<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Sam Stone on Medium]]></title>
        <description><![CDATA[Stories by Sam Stone on Medium]]></description>
        <link>https://medium.com/@sjstone1987?source=rss-cbfd810ae7b5------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*dZErWfNSH08x1M4q9ahh4w.png</url>
            <title>Stories by Sam Stone on Medium</title>
            <link>https://medium.com/@sjstone1987?source=rss-cbfd810ae7b5------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 14:06:33 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sjstone1987/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com" rel="hub"/>
        <item>
            <title><![CDATA[Self-Improving AI and the Need for GPU Regulation]]></title>
            <link>https://medium.com/semi-random-thoughts/self-improving-ai-and-the-need-for-gpu-regulation-abefb745b496?source=rss-cbfd810ae7b5------2</link>
            <guid isPermaLink="false">https://medium.com/p/abefb745b496</guid>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[biden]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[regulation]]></category>
            <category><![CDATA[gpu]]></category>
            <dc:creator><![CDATA[Sam Stone]]></dc:creator>
            <pubDate>Mon, 01 Apr 2024 05:26:48 GMT</pubDate>
            <atom:updated>2024-04-01T05:27:35.716Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>We need better AI regulation, fast. The software-focused rules policymakers are focused on today won’t get us there.</em></strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*R247jAp6cJXpKvhF" /></figure><p>On March 18, Nvidia unveiled its new Blackwell graphics processing unit (GPU) that will enable training the largest AI models <a href="https://spectrum.ieee.org/nvidia-blackwell">four times faster</a> than today’s state-of-the-art. AI capabilities, already growing exponentially, will accelerate even more to a pace that is hard for most people to comprehend. This makes the need for additional Federal regulation of AI even more urgent, but today’s policymakers are focused on the wrong problem</p><p>We are in a unique, finite period where AI is not yet capable of self-improvement, but is fast approaching that potentially irreversible threshold. Regulation and control of systems approaching self-improvement ought to become the overwhelming imperative for federal AI policy. Software regulation — the primary focus of Biden’s executive order on AI — is, at present, a seductive distraction. Federal AI regulation must instead focus on hardware, specifically high-end GPU networks, to monitor and restrict access to AI systems approaching the ability to improve themselves.</p><p>Policymakers obsess over today’s AI systems increasing discrimination or taking jobs. While these are real risks, they are not novel or existential. Older, unintelligent algorithms present the same risks at lower magnitude. In contrast, self-improving AI is a fundamentally novel risk. It’s like a “chain reaction” over which we might lose control — with existential consequences. A <a href="https://aiimpacts.org/wp-content/uploads/2023/04/Thousands_of_AI_authors_on_the_future_of_AI.pdf">January 2024 survey</a> of widely-cited AI researchers estimates a 10% chance of “human extinction or similarly permanent and severe disempowerment of the human species” from losing control over advanced AI. We must regulate now — before the first “chain reaction — to ensure that the organization that builds the first self-improving AI is aligned with US interests.</p><p>Self-improving AI could rapidly acquire new digital and physical capabilities, and learn to evade human control. Last month, Google released Gemini 1.5, a model capable of <a href="https://blog.google/technology/ai/long-context-window-ai-models/">understanding and improving codebases tens of thousands of lines long</a>. While Gemini’s codebase isn’t public, it’s likely too long for Gemini to improve autonomously — yet. We may be only months from a model that can improve its codebase, triggering autonomous AI self-improvement.</p><p>But there’s hope. Nuclear nonproliferation efforts illuminate a path — hardware regulation — for managing a new technology with existential risk. During WWII, when US scientists first saw a potential path to a nuclear chain reaction, the government adopted policies to restrict access to the hardware to produce fissile material. They didn’t wait for proof a nuclear bomb would work to obstruct bad actors. [1]</p><p>Critics of this approach to nuclear regulation argue regulation slowed innovation. That’s true. But the US’ failure to expand the benefits of peaceful, clean nuclear energy, stemmed not from insufficient innovation, but rather from public opposition in the wake of accidents, like Three Mile Island. It’s an example of how lax regulation and early failures to protect public safety can quickly undermine society’s tolerance for subsequent innovation.</p><p>The wisdom of a hardware-based approach to AI regulation stems from a few simple facts. First, improving any AI system requires two things: new software and hardware with which to train the new software. Creating new software is relatively simple. Significant recent advances in AI have resulted from small software changes, where existing code is simply trained on more GPUs. [2] Moreover, software is easy to modify and share, hindering government monitoring and control.</p><p>On the other hand, GPUs are costly physical objects well-suited to tracking, needed in high volumes, and dominated by US-based Nvidia. [3] This makes GPUs relatively easy to monitor and control. By restricting access to high-end GPU networks, the Biden administration can ensure that only organizations aligned with US interests, acting transparently, can train cutting-edge models. And GPU restrictions can directly prevent a “runaway” AI. A self-improving AI may learn to rewrite its code, but without GPU access, that new code is not deployable and the self-improvement becomes impossible to realize.</p><p>The Biden administration has taken steps in the right direction with <a href="https://public-inspection.federalregister.gov/2023-23055.pdf">controls</a> preventing high-end GPU exports to China (amongst other countries). But Chinese companies continue to deploy cutting-edge models, likely via black market GPU imports. [4] Export controls for this hardware must be tightened fast, but the Commerce Department’s planned cadence of annual updates is too slow.</p><p>October’s Executive Order also initiated domestic monitoring of AI training. It requires reporting for organizations acquiring GPUs networks or training models when either the network size or the model size crosses a threshold. [5] The GPU threshold is more enforceable since GPU suppliers have been corralled into reporting GPU use — and there are only a few, US-based GPU suppliers. [6] But today’s GPU threshold is too lenient and thus ineffective. An organization could acquire a GPU cluster below the reporting threshold, run it at maximum capacity, and produce a model above the threshold in under 12 days. [7]</p><p>The best path to closing these loopholes is congressional legislation that expands reporting and solidifies enforcement. Until then, the Commerce Department’s <a href="https://www.federalregister.gov/documents/2024/01/29/2024-01580/taking-additional-steps-to-address-the-national-emergency-with-respect-to-significant-malicious">January proposed rule</a> to extend reporting to GPUs rented (not just acquired) by foreign companies should be adopted, and applied to domestic companies.</p><p>Eighty years into the nuclear era, we haven’t blown ourselves up. At the core of nonproliferation’s success is the hardware-focused Nuclear Nonproliferation Treaty’s “trigger list” of controlled materials. It devotes half a page to defining fissile material (what goes directly into a bomb) and thirty-five pages to the industrial components, like centrifuges, needed to produce fissile material. [8]</p><p>Where nuclear weapons require thousands of centrifuges, self-improving AI will require thousands of GPUs. The hardware-based regulation that worked for nuclear weapons can also work for AI. Smart regulation cannot prevent self-improving AI, but it can dramatically increase the chances it is developed by an organization aligned with US interests acting transparently and safely.</p><h3>Footnotes</h3><p>[1] Presidential advisor Vannevar Bush, a leader of the Manhattan Project, recommended uranium export controls in August 1942, months before the first sustained nuclear chain reaction in December 1942 proved the viability of a nuclear weapon. <a href="https://www.osti.gov/opennet/manhattan-project-history/Processes/UraniumMining/uranium-procurement.html">US Department of Energy Manhattan Project History</a>.</p><p>[2] <a href="https://openreview.net/pdf?id=yzkSU5zdwD">This paper</a> details “emergent” (unexpected) capabilities in LLMs when they are trained with more compute, but the architecture and data are left unchanged. “Emergence” has been observed in arithmetic, translation, question answering, and other areas.</p><p>[3] This <a href="https://arxiv.org/abs/2402.15627">February 2024 paper from Baidu</a> uses 12K GPUs for a 175B open-source LLM; frontier models like GPT-4 are suspected to use even more GPUs. Nvidia was estimated to have &gt;90% of the global market for machine learning GPUs in 2023, per <a href="https://iot-analytics.com/leading-generative-ai-companies/">this report</a>. The next largest ML GPU suppliers — AMD, Google, Intel — are also US-based.</p><p>[4] This <a href="https://arxiv.org/abs/2402.15627">February 2024 paper from Chinese company Baidu</a> indicates that the company is likely still able to access large networks of Nvidia A100s, which are subject to export controls. In November 2023, Chinese company 01.AI <a href="https://www.digitalinformationworld.com/2023/11/new-chinese-ai-outperforms-meta-llm.html">released a 34 billion parameter LLM called Yi-34B</a> that outperformed Meta’s highest-performing, open-source LLM, Llama 2. <a href="https://asiatimes.com/2024/01/china-gets-banned-nvidia-ai-chips-via-gray-markets/#:~:text=The%20US%20Commerce%20Department&#39;s%20Bureau,results%20have%20not%20been%20satisfactory.">This article</a> details allegations of how China is evading GPU export controls.</p><p>[5] See section 4.2(b) of the October 30, 2023 <a href="https://public-inspection.federalregister.gov/2023-23055.pdf">Executive Order</a>.</p><p>[6] 15 companies that have voluntarily pledged to help enforce the Executive Order including Nvidia, Google, Amazon, and Microsoft, which together control well over 90% of the global GPU market, either directly (via GPU manufacturing) or via GPU cloud service provision. <a href="https://www.whitehouse.gov/briefing-room/statements-releases/2023/09/12/fact-sheet-biden-harris-administration-secures-voluntary-commitments-from-eight-additional-artificial-intelligence-companies-to-manage-the-risks-posed-by-ai/">White House briefing</a>.</p><p>[7] An organization could acquire a GPU cluster capable of just under 1020 floating-point operations per second (the second reporting requirement in EO section 4.2(b)), and run it for 11.6 days (~106 seconds), thus producing a model trained on more than 1026 floating-point operations (the first reporting requirement in EO section 4.2(b)). This assumes the GPU network is operating at its theoretical max capacity.</p><p>[8] See the November 28, 2023 Zangger Committee <a href="https://zanggercommittee.org/download/18.43f3519618d44fec3d787/1709202457619/2023-11-28%20-%20ZC%20Trigger%20List%20Rev.%205.pdf">Consolidated Trigger List</a>. Fissile material is defined on page 1 (Section 2); components for producing fissile material are defined on pages 5–43 (Annex Clarification Of Items On The Trigger List).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=abefb745b496" width="1" height="1" alt=""><hr><p><a href="https://medium.com/semi-random-thoughts/self-improving-ai-and-the-need-for-gpu-regulation-abefb745b496">Self-Improving AI and the Need for GPU Regulation</a> was originally published in <a href="https://medium.com/semi-random-thoughts">Structured Ramblings</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Unlearning to Build Great AI Apps]]></title>
            <link>https://medium.com/data-science/unlearning-to-build-great-ai-apps-29c9f5c2f321?source=rss-cbfd810ae7b5------2</link>
            <guid isPermaLink="false">https://medium.com/p/29c9f5c2f321</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[editors-pick]]></category>
            <category><![CDATA[user-experience]]></category>
            <dc:creator><![CDATA[Sam Stone]]></dc:creator>
            <pubDate>Mon, 08 Jan 2024 17:14:41 GMT</pubDate>
            <atom:updated>2024-01-08T17:14:41.833Z</atom:updated>
            <content:encoded><![CDATA[<h4>Product strategies from Classical ML to adapt (or ditch) for the generative AI world</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ftsTFfjEmCuRLZs-" /><figcaption>Image source: Tome</figcaption></figure><p>Years ago, the first piece of advice my boss at Opendoor gave me was succinct: <em>“Invest in backtesting. AI product teams succeed or fail based on the quality of their backtesting.”</em> At the time, this advice was tried-and-true; it had been learned the hard way by teams across search, recommendations, life sciences, finance, and other high-stakes products. It’s advice I held dear for the better part of a decade.</p><p>But I’ve come to believe it’s not axiomatic for building <em>generative</em> AI products. A year ago, I switched from classical ML products (which produce simple output: numbers, categories, ordered lists) to generative AI products. Along the way, I discovered many principles from classical ML no longer serve me and my teams.</p><p>Through my work at Tome, where I’m Head of Product, and conversations with leaders at generative AI startups, I’ve recognized 3 behaviors that distinguish the teams shipping the most powerful, useful generative AI features. These teams:</p><ol><li>Simultaneously work backwards (from user problems) and forwards (from technology opportunities)</li><li>Design low-friction feedback loops from the outset</li><li>Reconsider the research and development tools from classical ML</li></ol><p>These behaviors require “unlearning” numerous things that remain best practices for classical ML. Some may seem counter-intuitive at first. Nonetheless, they apply to generative AI applications broadly, ranging from horizontal to vertical software, and startups to incumbents. Let’s dive in!</p><p><em>(Wondering why automated backtesting is no longer a tenet for generative AI application teams? And what to replace it with? Read on to Principle 3)</em></p><p><em>(More interested in tactics, rather than process, for how generative AI apps’ UI/UX should differ from classical ML products? </em><a href="https://towardsdatascience.com/the-design-of-everyday-ai-things-26516d928566"><em>Check out this blog post.</em></a><em>)</em></p><h3>Principle 1: Simultaneously work backwards (from user problems) and forwards (from technology opportunities)</h3><p>“Working backwards” from user problems is a credo in many product and design circles, <a href="https://www.nytimes.com/2021/02/13/business/dealbook/amazon-working-backwards.html">made famous by Amazon</a>. Study users, size their pain points, write UX requirements to mitigate the top one, identify the best technology to implement, then rinse and repeat. In other words, figure out “This is the most important nail for us to hit, then which hammer to use.”</p><p>This approach makes less sense when enabling technologies are advancing very rapidly. ChatGPT was not built by working backwards from a user pain point. It took off because it offered a powerful, new enabling technology through a simple, open-ended UI. In other words: “We’ve invented a new hammer, let’s see which nails users will hit with it.”</p><p>The best generative AI application teams work backwards and forwards simultaneously. They do the user research and understand the breadth and depth of pain points. But they don’t simply progress through a ranked list sequentially. Everyone on the team, PMs and designers included, is deeply immersed in recent AI advances. They connect these unfolding technological opportunities to user pain points in ways that are often more complex than one-to-one mappings. For example, a team will see that user pain points #2, #3, and #6 could all be mitigated via model breakthrough X. Then it may make sense for the next project to focus on “working forwards” by incorporating model breakthrough X, rather than “working backwards” from pain point #1.</p><p>Deep immersion in recent AI advances means understanding how they apply to your real-world application, not just reading research papers. This requires prototyping. Until you’ve tried a new technology in your application environment, estimates of user benefit are just speculation. The elevated importance of prototyping requires flipping the traditional <strong><em>spec → prototype → build</em></strong> process to <strong><em>prototype → spec → build</em></strong>. More prototypes are discarded, but that’s the only way to spec features consistently that match useful new technologies to broad, deep user needs.</p><h3>Principle 2: Design low-friction feedback loops from the outset</h3><h4><strong>Feedback for system improvement</strong></h4><p>Classical ML products produce relatively simple output types: numbers, categories, ordered lists. And users tend to accept or reject these outputs: you click a link in the Google search results page, or mark an email as spam. Each user interaction provides data that is fed directly back into model retraining, so the link between real-world use and model improvement is strong (and mechanical).</p><p>Unfortunately, most generative AI products tend not to produce new, ground-truth training data with each user interaction. This challenge is tied to what makes generative models so powerful: their ability to produce complex artifacts that combine text, images, video, audio, code, etc. For a complex artifact, it’s rare for a user to “take it or leave it”. Instead, most users refine the model output, either with more/different AI or manually. For example, a user may copy ChatGPT output into Word, edit it, and then send it to a colleague. This behavior prevents the application (ChatGPT) from “seeing” the final, desired form of the artifact.</p><p>One implication is to allow users to iterate on output within your application. But that does not eliminate the problem: when a user does not iterate on an output, does that mean “wow” or “woe”? You could add a sentiment indicator (e.g. thumbs up/down) to each AI response, but interaction-level feedback response rates tend to be <em>very</em> low. And the responses that are submitted tend to be biased towards the extremes. Users mostly perceive sentiment collection efforts as additional friction, as they mostly don’t help the user immediately get to a better output.</p><p>A better strategy is to identify a step in the user’s workflow that signifies “this output is now good enough”. Build that step into your app and make sure to log what the output looked like at this point. For Tome, where we help users craft presentations with AI, the key step is sharing a presentation with another person. To bring this into our app, we’ve invested heavily in sharing features. And then we evaluate which AI outputs were “sharable” and which required massive manual editing to be shareable.</p><h4><strong>Feedback for user assistance</strong></h4><p>Free text has emerged as the dominant user-desired method of interacting with generative AI applications. But free text is a Pandora’s box: give a user free text input to AI, and they’ll ask the product to do all sorts of things it cannot. Free text is a notoriously difficult input mechanism through which to convey a product’s constraints; in contrast, an old-fashioned web form makes it very clear what information can and must be submitted, and in exactly what format.</p><p>But users don’t want forms when doing creative or complex work. They want free text — and guidance on how to craft great prompts, specific to their task at hand. Tactics for assisting users include example prompts or templates, guidance around optimal prompt length and formatting (should they include <a href="https://www.promptingguide.ai/techniques/fewshot">few-shot examples</a>?). Human-readable error messages are also key (for example: “This prompt was in language X, but we only support languages Y and Z.”)</p><p>One upshot of free text inputs is that unsupported requests can be a fantastic source of inspiration for what to build next. The trick is to be able to identify and cluster what users are trying to do in free text. More on that in the next section…</p><h3>Principle 3: Reconsider the research and development tools from classical ML</h3><p><em>Something to build, something to keep, something to discard</em></p><h4><strong>Build: natural language analytics</strong></h4><p>Many generative AI applications allow users to pursue very different workflows from the same entry point: an open-ended, free-text interface. Users are not selecting from a drop-down “I’m brainstorming” or “I want to solve a math problem” — their desired workflow is implicit in their text input. So understanding users’ desired workflows requires segmenting that free text input. Some segmenting approaches are likely to be enduring — at Tome, we are always interested in desired language and audience type. There are also ad hoc segmentations, to answer specific questions on the product roadmap — for example, how many prompts request a visual element like an image, video, table or chart, and thus which visual element should we invest in?</p><p>Natural language analytics should complement, not supplant, traditional research approaches. NLP is especially powerful when paired with structured data (e.g., traditional SQL). Lots of key data is not free text: when did the user sign up, what are the user’s attributes (organization, job, geography, etc). At Tome, we tend to look at language clusters by job function, geography, and free/paid user status — all of which require traditional SQL.</p><p>And quant insights should never be relied on without qualitative insights. I’ve found that watching a user navigate our product <em>live</em> can sometimes generate 10x the insight of a user interview (where the user discusses their product impression <em>post-hoc</em>). And I’ve found scenarios where one good user interview unlocked 10x the insight of quant analysis.</p><h4><strong>Keep: tooling for low-code prototyping</strong></h4><p>Two tooling types enable high-velocity, high-quality generative AI app development: prototyping tools and output quality assessment tools.</p><p>There are many different ways to improve an ML application, but one strategy that is both fast and accessible is <a href="https://platform.openai.com/docs/guides/prompt-engineering">prompt engineering</a>. It’s fast because it does not require model retraining; it’s accessible because it involves natural language, not code. Allowing non-engineers to manipulate prompt engineering approaches (in a dev or local environment) can dramatically increase velocity and quality. Often this can be implemented via a notebook. The notebook may contain a lot of code, but a non-engineer can make significant advances iterating on the natural language prompts without touching the code.</p><p>Assessing prototype output quality is often quite hard, especially when building a net-new feature. Rather than investing in automated quality measurement, I’ve found it significantly faster and more useful to poll colleagues or users in a “beta tester program” for 10–100 structured evaluations (scores + notes). The enabling technology for a “polling approach” can be light: a notebook to generate input/output examples at modest scale and pipe them into a Google Sheet. This allows manual evaluation to be parallelized, and it’s normally easy to get ~100 examples evaluated, across a handful of people, in under a day. Evaluators’ notes, which provide insights into patterns of failure or excellence, are an added perk; notes tend to be more useful for identifying what to fix or build next than the numeric scores.</p><h4><strong>Discard: automated, backtested measures of quality</strong></h4><p>A tenet of classical ML engineering is to invest in a robust backtest. Teams retrain classical models frequently (weekly or daily), and a good backtest ensures only good new candidates are released to production. This makes sense for models outputting numbers or categories, which can be scored against a ground-truth set easily.</p><p>But scoring accuracy is harder with complex (perhaps multi-modal) output. You may have a text that you consider great and thus you’re inclined to call it “ground truth”, but if the model output deviates from it by 1 word, is that meaningful? By 1 sentence? What if the facts are all the same, but the structure is different? What if it’s text and images together?</p><p>But not all is lost. Humans tend to find it easy to assess whether generative AI output meets their quality bar. That doesn’t mean it’s easy to transform bad output into good, just that users tend to be able to make a judgment about whether text, image, audio, etc. is “good or bad” in a few seconds. Moreover, most generative AI systems at the application layer are not retrained on a daily, weekly, or even monthly basis, because of compute costs and/or the long timelines needed to acquire sufficient user signal to warrant retraining. So we don’t need quality evaluation processes that are run every day (unless you’re Google or Meta or OpenAI).</p><p>Given the ease with which humans can evaluate generative AI output, and the infrequency of retraining, it often makes sense to evaluate new model candidates based on internal, manual testing (e.g. the polling approach described in the subsection above) rather than an automated backtest.</p><p>Developing great generative AI applications requires more than just innovative engineering; it requires product and design to work in ways that are fundamentally different from the tenets of classical ML or non-ML products. Some patterns need to be unlearned (“invest in your backtest!”), some need to be reformulated (“work backwards and forwards”), and some are net-new (“build for feedback loops from the outset.”)</p><p>It can be a lot of process and people rewiring, but the upside is enormous — and if you don’t do it, the revolution will pass you by.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=29c9f5c2f321" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/unlearning-to-build-great-ai-apps-29c9f5c2f321">Unlearning to Build Great AI Apps</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Design of Everyday (AI) Things]]></title>
            <link>https://medium.com/data-science/the-design-of-everyday-ai-things-26516d928566?source=rss-cbfd810ae7b5------2</link>
            <guid isPermaLink="false">https://medium.com/p/26516d928566</guid>
            <category><![CDATA[product-design]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[notes-from-industry]]></category>
            <category><![CDATA[user-experience]]></category>
            <category><![CDATA[design]]></category>
            <dc:creator><![CDATA[Sam Stone]]></dc:creator>
            <pubDate>Fri, 06 Oct 2023 23:58:54 GMT</pubDate>
            <atom:updated>2023-10-06T23:58:54.401Z</atom:updated>
            <content:encoded><![CDATA[<h4>UI/UX Principles for Building Great Generative AI Tools</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*snPQdm8ptYcNEWWG" /><figcaption>Image source: DreamStudio</figcaption></figure><p>Don Norman’s 1988 design classic, “The Design of Everyday Things,” laid out user experience principles that have influenced great hardware and software design ever since. While Norman drew on analog examples like door handles and light switches, his principles are broadly applicable to software, including generative AI products. With all the hype about generative AI, it’s easy to forget that products powered by even the most sophisticated models will fail if they lack good UI/UX.</p><p>Many new AI tools have generated loads of interest, followed by lackluster user retention (<a href="https://www.sequoiacap.com/article/generative-ai-act-two/">as detailed here by Sequoia</a>). AI hype drives “tourist” signups, but new users struggle to understand or derive real value from the product. This is the classic “trough of disillusionment”, which occurs when a core technology (generative models) jumps ahead while supporting technology (UI/UX design) lags behind.</p><p>This post details how to apply three core UX concepts to generative AI products: <strong>1) affordances, 2) feedback, and 3) constraints.</strong> Applying these concepts to generative AI leads to conclusions we’ll explore, including:</p><ul><li>Don’t aim for a Hole-In-One</li><li>User feedback isn’t free</li><li>Treat chatbot interfaces skeptically</li></ul><p>The examples to follow are drawn from workplace productivity tools (and partly inspired by learnings from my work at Tome, an AI-powered medium for shaping &amp; sharing ideas), but the strategies apply broadly, from dev tools to social media to e-commerce.</p><h3>Topic 1: Feedback</h3><p>Providing quick, clear feedback to the user about a requested action is critical for any technology. Feedback is especially important for generative AI systems, due to the latency and complexity of their output. And feedback goes both ways. A system must solicit relevant feedback from the user without being annoying, to generate better output for that particular user in the near-term, and to enable better versions of the product in the medium or long-term.</p><h4>Build for latency</h4><p>Generative AI model response times tend to range from single to double digit seconds. At first glance, waiting ten seconds for a compelling memo, a brilliant image, or a beautiful video might seem like a non-issue. Without generative AI, these actions take hours — who cares if it takes 10 seconds, 1 second, or 100 milliseconds?</p><p>But users aren’t economists optimizing for opportunity cost. They’ve been conditioned by non-AI tools to expect software so fast it’s perceived as instant. This leads to a number of user challenges with obviously non-instant AI products:</p><ol><li>Confusion whether the system is working and if they need to retry/restart.</li><li>High perceived cost of iteration. And since most of the time the first artifact an AI generates is not exactly what the user wants, the user wants to iterate.</li><li>High likelihood the user starts to multi-task. Once a user has switched away from your app, there is no guarantee they will ever come back.</li></ol><p>There are good strategies for mitigating latency effects that pre-date generative AI. These include loading animations, progress bars, and background processing (in which the user is routed to another task and receives a notification when the current task finishes). A newer tactic, specific to LLM features, is streaming text word-by-word (or character-by-character) to the UI, rather than rendering the full output all at once. Since many models can generate words faster than a user can read, this can reduce perceived latency to near zero.</p><h4>Don’t aim for a Hole-In-One</h4><p>One particularly effective strategy for mitigating latency is to break workflows into small steps, in which system feedback is provided and user feedback is solicited at each step. This allows the user to progress towards an output with increasing confidence that the system will deliver precisely what the user wants. In a well-designed iterative workflow, the latency of initial steps is low — and the user’s trust increases with each successive step that the final output will be what they want. If you’re quite confident that you’re going to get back the artifact you want, then you <em>will</em> be willing to wait ten seconds for the final step to run.</p><p>Iterative workflows have an even more powerful benefit than increasing latency tolerance: they enable users to generate output that matches their expectation better. Generative models can sometimes<em> </em>produce exactly what the user wants, from just a simple user prompt. And going straight from input to “nailed it” final output is an amazing user experience; it’s like hitting a hole in one. And like hitting a hole in one, it’s very rare.</p><p>The challenge is not how “smart” the model is, but rather what context and information the model needs to produce the user’s vision. Consider a sales manager who wants to summarize her team’s quarterly performance. She’s seen dozens of quarterly sales reports, and she’s deeply familiar with her company’s norms that govern such artifacts (norms like tone, detail level, length, and visual layout). If she needed a colleague to write her such a report, she’d simply ask for “a quarterly sales report” and expect the colleague to already know these norms.</p><p>So when this sales manager wants to get such a report from an AI tool, it’s not obvious to her what norms she needs to tell the tool, and what it already knows. This is where iterative workflows are particularly helpful. She can start with something easy and familiar, like requesting “a quarterly sales report” and the tool can then help her hone in on precisely what she has in mind. Zach Lloyd calls this pattern “ask and adjust” in <a href="https://thezbook.com/ask-adjust-the-future-of-productivity-interfaces">this well-reasoned piece on AI design.</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mwt0aRC1-3nXlDoWb_X5Qg.png" /><figcaption>Tome’s outline editor is an example of an intermediate step in an iterative AI workflow, situated in between a prompt and the final output, a multi-page presentation. <a href="https://tome.page">https://tome.page</a></figcaption></figure><h4>User feedback isn’t free</h4><p>In many classical ML products, each user interaction generates a new piece of training data for the model, improving the next version of the product. Every user click on a search result helps the search model improve. Every email a user marks as spam helps a spam classification model improve.</p><p>But many generative AI products lack the inherent “physics” where user interactions mechanically lead to model improvement. For AI products where the output is a sophisticated piece of text, image, etc, it can be hard to distinguish between a frustrated exit (where the user couldn’t get the output they wanted, and quit) vs a satisfied exit ( the user got what they wanted and left). Some products solicit opt-in feedback (e.g. thumbs up/down), but the completion rates tend to be very low and the feedback itself often suffers from selection bias.</p><p>It’s far better to design a workflow where the user’s natural next action indicates their perception of the preceding AI output. One pattern, most commonly seen with text models, is in-line suggestions: if the user accepts the suggestion and keeps writing, that’s a strong signal they viewed the suggestion positively. Another pattern is to instrument what AI output gets saved, edited, or shared. These are not perfectly correlated with user satisfaction — a user may share an image because it’s particularly bizarre — but they’re decent proxies when used in the aggregate.</p><h3>Topic 2: Affordances</h3><p>An affordance is a cue (normally visual) that suggests how and when to use a feature. Good affordances make it intuitive for users to interact with a product, without extensive instructions or experience. We’ll explore affordances for generative AI at three steps in the user journey: discovering entry points to AI, proving the right input for the AI, and using the AI output.</p><h4>Discovering AI entry points</h4><p>Many work tools are adding lots of AI features, and these features are applicable at different points in the creative process. High-level entry points for using AI features include:</p><ol><li>Help me start from scratch</li><li>Extend what I’ve started</li><li>Edit what I’ve created</li></ol><p>These different entry points have led to significantly different interfaces, even at this early point in AI interface evolution. For (1), free text or “blank canvas” interfaces have emerged as early leading paradigms. For (2), in-line generation (aka autocomplete) tends to dominate text generation features (like Github Copilot), while “show me more like this” tends to dominate image generation features (like Midjourney). For (3), interfaces tend to focus on highlighting, selecting, or uploading existing content (like Grammarly).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pc3tb2JfhzhB2bB0jkzZ_w.png" /><figcaption>Whimsical’s AI Mind Map helps users get started from scratch. <a href="https://whimsical.com">https://whimsical.com</a></figcaption></figure><p>For a user who has discovered one AI entry point in a tool with multiple AI features, it’s easy to conclude “this is where the AI lives” and fail to discover the other features. Great products mitigate this by introducing users to their various AI entry points at the times in the user’s workflow when each entry point is most likely to be useful.</p><h4>Entering input for AI</h4><p>The core input of many generative AI workflows is free text input, aka “prompting”. Unfortunately, good prompting is complicated, fast-evolving, and inconsistent across tools. Good products help users craft prompts with strategies including example prompts and tooltips.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*BmPTdPkZ4mc5lOCu" /><figcaption>Perplexity includes a handful of example prompts on its landing page, to illustrate use cases that extend beyond typical search engines. <a href="https://www.perplexity.ai/">https://www.perplexity.ai/</a></figcaption></figure><p>Good interfaces also help the user understand the context the AI has — and what it lacks. When working with a powerful AI, a reasonable user may conclude that whatever they can see in the app, the AI must also be able to see and understand. For example, if I can see my past conversation with the AI, surely the AI must also be aware of it (this is a behavior ChatGPT popularized). But not every AI works like this! Some systems are aware of the user’s previous prompts, some are aware of even more context than past prompts — and some are only aware of the user’s current interaction and nothing else. The user should not have to figure out what the system knows and what it does not through trial-and-error.</p><h4>Using AI output</h4><p>It’s tempting to think that when a system has produced generative AI output, and the output is good, success is at hand. But even when the output is good, this can be a confusing moment for the user.</p><p>First, new users are often left wondering how to persist the output. Even when the output is good, many users immediately want to iterate and see if they can go from good to great. But fear they may lose their existing work can lead to hesitation and frustration.</p><p>Second, users may be confused <em>how</em> to improve the output. Assuming they used a “start from scratch” AI feature, should they go back to the beginning? Do they need to move to a different AI entry point like “extend” or “edit”? Many users will have encountered products like ChatGPT where the output is not directly editable; if output is editable, users likely need an editing affordance.</p><h3>Topic 3: Constraints</h3><p>Constraints restrict input and output to help users work faster and better. Good constraints are clear to the user. If a system can help a user achieve a goal — but only part way or part of the time — it can be better to prevent that path altogether, rather than deliver an unreliable experience.</p><p>LLMs open up vast new user experiences (it’s why I love working on them!) and product creators should be eager to relax traditional constraints from deterministic software. Nonetheless, regardless of how intelligent LLMs become, there will always be a place for some thoughtful constraints.</p><h3>Karina Nguyen on Twitter: &quot;Design has been operating under the ethos &quot;solving user problems while navigating constraints&quot; but the open-ended, ambiguous nature of LLMs makes this a very limited mindset. Truth is, constraints will always evolve and if you are a designer, your goal is to imagine your ideal... / Twitter&quot;</h3><p>Design has been operating under the ethos &quot;solving user problems while navigating constraints&quot; but the open-ended, ambiguous nature of LLMs makes this a very limited mindset. Truth is, constraints will always evolve and if you are a designer, your goal is to imagine your ideal...</p><h4>Input: don’t fear controls</h4><p>Inspired by the success of ChatGPT, many generative AI tools use a free text box as their only, or primary, user input. But many aspects of a user’s intent are best expressed via categorical or numeric inputs. When creating a document, most users have attributes in mind like language (a categorical) and length (a numeric value). Users may not mention these attributes in a free text prompt, but it doesn’t mean they don’t care about them. By soliciting this input via discrete, bounded controls (like a drop-down or slider), a system helps solicit the input it needs to deliver what the user has in their head. And there are time-honored principles for helping users navigate discrete controls: setting good defaults, grouping controls logically, and explaining controls with tooltips or labels.</p><p>When it comes to controls, setting good default values is a critical part of the design. The vast majority of the time (well in excess of 90%) users will not change the defaults, even if they would benefit from doing so. One opportunity to combine good defaults with user preference variance is to adjust defaults dynamically, either via hard-coded rules or AI.</p><h4>Output: not everything that can be generated should be</h4><p>For generative AI products, there are many situations in which the underlying model can produce <em>some </em>content, but where the user would nothing vs struggling with misleading or jarring output.</p><p>For most work-related tasks, users would prefer “I don’t know” rather than a potentially false answer they’re must verify or refute . <a href="https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4573321">This Harvard study at consulting firm BCG</a> shows how AI can diminish work quality when AI answers questions beyond its “confidence frontier” and users, unaware where the frontier is, don’t sufficiently scrutinize the output.</p><p>Methods for reducing hallucination are fast-evolving (for example, <a href="https://www.pinecone.io/learn/retrieval-augmented-generation/">retrieval-augmented generation</a>), and I suspect hallucination will be a mostly “solved” problem a few years from now — but today, output where factuality is critical remains an important place to consider constraints.</p><p>Legal and ethical concerns are a second reason to constrain user-facing output. Just because the underlying model can generate text or images on a topic does not mean it’s conscionable to do so. However, much of the time that a system classifies a user request as “out of bounds”, the user’s intent is actually benign. With a bit of help, the user could rephrase their request to stay in bounds. For example, some image generation tools deny prompts that include the word “child”. But if a user wants to generate a picture of a family with children, they could prompt “family of four” or “parents with son and daughter”. The key is that the constraints are clear to the user.</p><p>As generative AI products surge in popularity, good product designers and product managers remember: success stems not just from how smart the AI is; but how the product guides the user through AI-enabled workflows. Core design concepts like feedback, affordances, and constraints remain just as important as ever, but the tactics and patterns for how they are implemented are quickly evolving. Using these design patterns well is critical for any AI company that aspires to endure beyond the initial hype cycle and deliver an enduring, widely-used product.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=26516d928566" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/the-design-of-everyday-ai-things-26516d928566">The Design of Everyday (AI) Things</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Applying LLMs to Enterprise Data: Concepts, Concerns, and Hot-Takes]]></title>
            <link>https://medium.com/data-science/applying-llms-to-enterprise-data-concepts-concerns-and-hot-takes-e19ded4bde88?source=rss-cbfd810ae7b5------2</link>
            <guid isPermaLink="false">https://medium.com/p/e19ded4bde88</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[editors-pick]]></category>
            <dc:creator><![CDATA[Sam Stone]]></dc:creator>
            <pubDate>Sun, 28 May 2023 14:19:07 GMT</pubDate>
            <atom:updated>2023-08-18T01:47:49.458Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*mP4dYKPVqzOzREbxxFPxAg.png" /><figcaption>Source: <a href="https://beta.dreamstudio.ai/">DreamStudio</a> (generated by author)</figcaption></figure><p>Ask GPT-4 to prove there are infinite prime numbers — while rhyming — <a href="https://arxiv.org/pdf/2303.12712.pdf">and it delivers</a>. But ask it how your team performed vs plan last quarter, and it will fail miserably. This illustrates a fundamental challenge of large language models (“LLMs”): they have a good grasp of general, public knowledge (like prime number theory), but are entirely unaware of proprietary, non-public information (how your team did last quarter.)[1] And proprietary information is critical to the vast majority of enterprise use workflows. A model that understands the public internet is cute, but little use in its raw form to most organizations.</p><p>Over the past year, I’ve had the privilege of working with a number of organizations applying LLMs to enterprise use cases. This post details key concepts and concerns that anyone embarking on such a journey should know, as well as a few hot-takes on how I think LLMs will evolve and implications for ML product strategy. It’s intended for product managers, designers, engineers and other readers with limited or no knowledge of how LLMs work “under the hood”, but some interest in learning the concepts without going into technical details.</p><h3>Four Concepts</h3><h4>Retrieval-Augmented Generation (RAG): Context Windows and Embeddings</h4><p>The simplest way to make an LLM reason about proprietary data is to provide the proprietary data in the model’s prompt. Most LLMs would have no problem answering the following correctly: “We have 2 customers, A and B, who spent $100K and $200K, respectively. Who was our largest customer and how much did they spend?” We’ve just done some basic prompt engineering, by prepending our query (the second sentence) with context (the first sentence).</p><p>But in the real world, we may have thousands or millions of customers. How do we decide which information should go into the context — considering that each word included in the context costs money ? This is where embeddings come in. Embeddings are a method in which text is transformed into numerical vectors, in which similar text generates similar vectors (vectors that are “close together” in N-dimensional space).[2] We might embed website text, documents, maybe even an entire corpus from SharePoint, Google Docs, or Notion. Then, for each user prompt, we embed it and find the vectors from our text corpus that are most similar to our prompt vector.</p><p>For example, if we embedded Wikipedia pages on animals, and the user asked a question about safaris, our search would rank highly the Wikipedia articles about lions, zebra, and giraffes. This allows us to identify the text chunks most similar to the prompt — and thus most likely to answer it.[3] We include these most similar text chunks in the context that is prepended to the prompt, so that the prompt hopefully contains all the information necessary for the LLM to answer the question.</p><h4><strong>Fine-Tuning</strong></h4><p>A downside of embeddings is that every call to the LLM requires all the context to be passed with the prompt. The LLM has no “memory” of even the most basic enterprise-specific concepts. And since most cloud-based LLM providers charge per prompt token, this can get expensive fast.[4]</p><p>Fine-tuning allows an LLM to understand enterprise-specific concepts without including them in each prompt. We take a foundation model, which already encodes general knowledge across billions of learned parameters, and tweak those parameters to reflect specific enterprise knowledge, while still retaining the underlying general knowledge.[5] When we generate inferences with the new fine-tuned model, we get that enterprise knowledge “for free”.</p><p>In contrast to embeddings/prompt engineering, where the underlying model is a third-party black box, fine-tuning is closer to classical machine learning, where ML teams created their own models from scratch. Fine-tuning requires a training dataset with labeled observations; the fine-tuned model is highly sensitive to the quality and volume of that training data. We also need to make configuration decisions (number of epochs, learning rate, etc), orchestrate long-running training jobs, and track model versions. Some foundation model providers provide APIs that abstract away some of this complexity, some do not.</p><p>While inferences may be cheaper with fine-tuned models, that can be outweighed by costly training jobs.[6] And some foundation model providers (like OpenAI) only support fine-tuning of lagging-edge models (so <a href="https://platform.openai.com/docs/guides/fine-tuning">not ChatGPT or GPT-4</a>).</p><h4><strong>Evals</strong></h4><p>One of the novel, significant challenges presented by LLMs is measuring the quality of complex outputs. Classical ML teams have tried-and-true methods for measuring the accuracy of simple outputs, like numerical predictions or categorizations. But most enterprise use cases for LLMs involve generating responses that are tens to thousands of words. Concepts sophisticated enough to require more than ten words can normally be worded in many ways. So even if we have a human-validated “expert” response, doing an exact string match of a model response to the expert response is too stringent a test, and would underestimate model response quality.</p><p>The <a href="https://github.com/openai/evals">Evals</a> framework, open-sourced by OpenAI, is one approach to tackling this problem. This framework requires a labeled test set (where prompts are matched to “expert” responses), but it allows broad types of comparison between model and expert responses. For example, is the model-generated answer: a subset or superset of the expert answer; factually equivalent to the expert answer; more or less concise than the expert answer? The caveat is that Evals perform these checks using an LLM. If there’s a flaw in the “checker” LLM, the eval results may themselves be inaccurate.</p><h4><strong>Adversarial Examples</strong></h4><p>If you’re using an LLM in production, you need to have confidence that it will handle misguided or malicious user inputs safely. For most enterprises, the starting point is ensuring the model doesn’t spread false information. That means a system that knows its own limitations and when to say “I don’t know.” There are many tactical approaches here. It can be done via prompt engineering with prompt language like “Respond ‘I don’t know’ if the question cannot be answered with the context provided above”). It can be done with fine-tuning, by providing out-of-scope training examples, where the expert response is “I don’t know”.</p><p>Enterprises also need to guard against malicious user inputs, e.g. <a href="https://learnprompting.org/docs/category/-prompt-hacking">prompt hacking</a>. Limiting the format and length of the system’s acceptable inputs and outputs can be an easy and effective start. Precautions are a good idea if you’re only serving internal users and they’re essential if you’re serving external users.</p><h3>Three Concerns</h3><h4><strong>Bias Perpetuation</strong></h4><p>The developers of the most popular LLMs (OpenAI / GPT-4, Google / Bard) have taken pains to align their models with human preferences and deploy sophisticated moderation layers. If you ask GPT-4 or Bard to tell you a racist or misogynistic joke, they will politely refuse.[7]</p><p>That’s good news. The bad news is that this moderation, which targets societal biases, doesn’t necessarily prevent institutional biases. Imagine our customer support team has a history of being rude to a particular type of customer. If historical customer support conversations are naively used to construct a new AI system (for example, via fine-tuning) that system is likely to replicate that bias.</p><p>If you’re using past data to train an AI model (be it a classical model or a generative model), closely scrutinize which past situations you want to perpetuate into the future and which you do not. Sometimes it’s easier to set principles and work from those (for example, via prompt engineering), without using past data directly.</p><h4><strong>Model Lock-In</strong></h4><p>Unless you’ve been living under a rock, you know generative AI models are advancing incredibly rapidly. Given an enterprise use case, the best LLM for it today may not be the best solution in six months and almost certainly will not be the best solution in six years. Smart ML teams know they will need to switch models at some point.</p><p>But there are two other major reasons to build for easy LLM “swapping”. First, many foundation model providers have struggled to support exponentially-growing user volume, leading to <a href="https://isdown.app/integrations/openai">outages and degraded service</a>. Building a fallback foundation model into your system is a good idea. Second, it can be quite useful to test multiple foundation models in your system (“a horse race”) to get a sense of which performs best. Per the section above on Evals, it’s often difficult to measure model quality analytically, so sometimes you just want to run two models and qualitatively compare the responses.</p><h4><strong>Data Leakage</strong></h4><p>Read the terms and conditions of any foundation model you’re considering using. If the model provider has the right to use user inputs for future model training, that’s worrisome. LLMs are so large it’s possible that specific user queries/responses become directly encoded in a future model version, and could then become accessible to any user of that version. Imagine a user at your organization queries “how can I clean up this code that does XYZ? [your proprietary, confidential code here]” If this query is then used by the model provider to retrain their LLM, that new version of the LLM may learn that your proprietary code is a great way to solve use case XYZ. If a competitor asks how to do XYZ, the LLM could “leak” your source code, or something very similar.</p><p>OpenAI now <a href="https://openai.com/blog/new-ways-to-manage-your-data-in-chatgpt">allows users to opt-out of their data being used to train models</a>, which is a good precedent, but not every model provider has followed their example. Some organizations are also exploring running LLMs within their own virtual private clouds; this is a key reason for much of the interest in open-source LLMs.</p><h3>Two Hot-Takes</h3><h4><strong>Prompt Engineering Will Dominate Fine Tuning</strong></h4><p>When I first started adapting LLMs for enterprise use, I was much more interested in fine tuning than prompt engineering. Fine tuning felt like it adhered to the principles of classical ML systems to which I was accustomed: wrangle some data, produce a train/test dataset, kick off a training job, wait a while, evaluate the results against some metric.</p><p>But I’ve come to believe that prompt engineering (with embeddings) is a better approach for most enterprise use cases. First, the iteration cycle for prompt engineering is far faster than for fine tuning, because there is no model training, which can take hours or days. Changing a prompt and generating new responses can be done in minutes. Conversely, fine-tuning is an irreversible process in terms of model training; if you used incorrect training data or a better base model comes out, you need to restart your fine-tuning jobs. Second, prompt engineering requires far less knowledge of ML concepts like neural network hyperparameter optimization, training job orchestration or data wrangling. Fine-tuning often requires experienced ML engineers, while prompt engineering can often be done by software engineers without ML experience. Third, prompt engineering works better for the fast-growing strategy of <a href="https://python.langchain.com/en/latest/index.html">model chaining</a>, in which complex requests are decomposed into smaller, constituent requests, each of which can be assigned to a different LLM. Sometimes the best “constituent model” is a fine-tuned model.[8] But most of the value-add work for enterprises is (i) figuring out how to break apart their problem, (ii) write the prompts for each constituent part, and (iii) identify the best off-the-shelf model for each part; it’s not in creating their own fine-tuned models.</p><p>The advantages of prompt engineering are likely to widen over time. Today, prompt engineering requires long, expensive prompts (since context must be included in each prompt). But I’d bet on rapidly declining cost per token, as the model provider space gets more competitive and providers figure out how to train LLMs more cheaply. Prompt engineering is also limited today by maximum prompt sizes — but , OpenAI already accepts 32K tokens (~40 pages of average English text) per prompt for GPT-4, and <a href="https://www.anthropic.com/index/100k-context-windows">Anthropic’s Claude accepts 100K tokens</a> (~15 pages). And I’d bet on even larger context windows coming out in the near future.</p><p><em>[August 2023 Update] This is a great piece on </em><a href="https://www.tidepool.so/2023/08/17/why-you-probably-dont-need-to-fine-tune-an-llm/"><em>why you probably don’t need fine-tuning.</em></a></p><h4><strong>Data Won’t Be The Moat It Once Was</strong></h4><p>As LLMs have become better at producing human-interpretable reasoning, its useful to consider how humans use data to reason, and what that implies for LLMs.[9] Humans don’t actually use much data! Most of the time, we do “zero shot learning”, which simply means we answer questions without the question being accompanied by a set of example question-answer pairs. The questioner just provides the question, and we answer based on logic, principles, heuristics, biases, etc.</p><p>This is different from the LLMs of just a few years ago, which were only good at few-shot learning, where you needed to include a handful of example question-answer pairs in your prompt. And it’s very different from classical ML, where the model needed to be trained on hundreds, thousands, or millions of question-answer pairs.</p><p>I strongly believe that an increasing, dominant share of LLM use cases will be “zero-shot”. LLMs will be able to answer most questions without any user-provided examples. They will need prompt engineering, in the form of instructions, policies, assumptions, etc. For example, <a href="https://betterprogramming.pub/i-used-gpt-3-to-find-213-security-vulnerabilities-in-a-single-codebase-cc3870ba9411">this post</a> uses GPT-4 to review code for security vulnerabilities; the approach requires no data on past instances of vulnerable code. Having clear instructions, policies, and assumptions will become increasingly important — but having large volumes of high-quality, labeled, proprietary data will become less important.</p><p>If you’re actively working on applying LLMs to your enterprise data, I’d love to hear about what you’ve found works and what does not. Please leave a comment!</p><h3>Footnotes</h3><p>[1] Until recently, LLMs were also unaware of recent public knowledge — for example GPT-4 was trained on information collected through Sept 2021. However, the consumer interfaces for GPT-4 and Bard are now able to query the open internet and collect information about recent events. So recency is quickly fading as a knowledge limitation for LLMs.</p><p>[2] Embeddings can work on all kinds of data structures, not just text.</p><p>[3] The entire embedding workflow occurs prior to calling the LLM. For example, OpenAI recommends using its ada-002 model for embeddings, which is cheaper and faster than any of the leading-edge GPT models.</p><p>[4] Tokens are words or word parts. <a href="https://blog.quickchat.ai/post/tokens-entropy-question/#:~:text=Why%20tokens%3F,up%20on%20features%20that%20matter.">Here’s a good explanation of why language models use tokens, not words.</a></p><p>[5] Learned parameter count could be anywhere from millions to trillions. Most widely-used LLMs today have billions.</p><p>[6] Cheaper inferences are not a given; <a href="https://docs.google.com/document/d/1iJRm5vmT3Udv2mGTD4LsLobYaHEj2OaBywIzI83WdQo/edit#">OpenAI charges </a>$0.03–0.06 per 1k tokens for GPT-4 with an 8K context window (depending on whether the tokens are inputs or outputs, respectively). It charges $0.12 per 1k tokens for a fine-tuned version of Davinci, a lagging-edge model.</p><p>[7] Of course, these are humans employed by OpenAI and Google. And since a lot of people disagree with those organizations’ values, they disagree with the moderation policies.</p><p>[8] For example, <a href="https://huggingface.co/papers/2305.14201">GOAT</a> is a version of the open-source model LLaMA fine-tuned for arithmetic. It outperforms GPT-4 on many arithmetic benchmarks. Most enterprises have workflows that require arithmetic; under the chaining approach, the parts of the workflow that involve arithmetic would be identified and routed to GOAT. It makes sense for such an enterprise to invest heavily in good routing and integration with GOAT, but, in my opinion, not to fine-tune their own arithmetic LLM.</p><p>[9] There is much debate about whether today’s LLMs can actually reason, and what actual reasoning even means (does it require consciousness? self-awareness? agency?) Sidestepping that debate, which is more philosophical than empirical, it’s worth noting that LLMs are unequivocally getting better at producing explanations that conform to widely-held notions of good reasoning; <a href="https://arxiv.org/pdf/2303.12712.pdf">this paper</a> has many great examples.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e19ded4bde88" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/applying-llms-to-enterprise-data-concepts-concerns-and-hot-takes-e19ded4bde88">Applying LLMs to Enterprise Data: Concepts, Concerns, and Hot-Takes</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data Observability for Analytics and ML teams]]></title>
            <link>https://medium.com/data-science/data-observability-for-analytics-and-ml-teams-421f6089d905?source=rss-cbfd810ae7b5------2</link>
            <guid isPermaLink="false">https://medium.com/p/421f6089d905</guid>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[observability]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Sam Stone]]></dc:creator>
            <pubDate>Thu, 06 Apr 2023 08:12:05 GMT</pubDate>
            <atom:updated>2023-04-06T08:12:05.300Z</atom:updated>
            <content:encoded><![CDATA[<h4>Principles, practices, and examples for ensuring high quality data flows</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*kEGNtsLkDpGo8uJON_IlHA.png" /><figcaption>Source: <a href="https://beta.dreamstudio.ai/">DreamStudio</a> (generated by author)</figcaption></figure><p><a href="https://www.americanbanker.com/payments/opinion/stolen-datas-a-hot-potato-that-gets-quickly-passed-around">Nearly 100%</a> of companies today rely on data to power business opportunities and 76% use data as an integral part of forming a business strategy. In today’s age of digital business, an increasing number of decisions companies make when it comes to delivering customer experience, building trust, and shaping their business strategy begins with accurate data. Poor data quality can not only make it difficult for companies to understand what customers want, but it can end up as a guessing game when it doesn’t have to be. Data quality is critical to delivering good customer experiences.</p><p>Data observability is a set of principles that can be implemented in tools to ensure data is accurate, up-to-date, and complete. If you’re looking to improve data quality at your organization, here is why data observability may be your answer and how to implement it.</p><h4><strong>How to know if you need data observability</strong></h4><p>Data observability is increasingly necessary, especially as traditional approaches to software monitoring fall short for high-volume, high-variety data. Unit tests, which assess small pieces of code for performance on discrete, deterministic tasks, get overwhelmed by the variety of acceptable shapes and values that real-world data can take. For example, a unit test can verify that a column intended to be a boolean is indeed a boolean, but what if the percentage “true” in that column shifted a lot between one day and the next? Or even just a little bit. Alternatively, end-to-end tests, which assess a full system, stretching across repos and services, get overwhelmed by the cross-team complexity of dynamic data pipelines. Unit tests and end-to-end testing are necessary but insufficient to ensure high data quality in organizations with complex data needs and complex tables.</p><p>There are three main signs your organization needs data observability — and it’s not only related to ML:</p><ul><li>Upstream data changes regularly break downstream applications, despite upstream teams’ prophylactic efforts</li><li>Data issues are regularly discovered by customers (internal or external) rather than the team that owns the table in question</li><li>You’re moving towards a centralized data team</li></ul><p>I’ve worked at <a href="http://opendoor.com">Opendoor</a> — an e-commerce platform for residential real estate transactions and large buyer and seller of homes — for the past four years and the data we use to assess home values is rich but often self-contradicting. We use hundreds of data feeds and maintain thousands of tables — including public data, third-party data, and proprietary data — which often disagree with one another. For instance, a home may have square footage available from a recent MLS listing and a public tax assessment that differs. Homeowners may have stated the highest possible square footage when selling the home, but stated the lowest possible area when dealing with tax authorities. Getting to the “ground truth” is not always easy, but we improve data accuracy by synthesizing across multiple sources — and that’s when data observability comes in.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fPtSmQcPm7mW_YnV1ZZ4hQ.png" /><figcaption>Home data example, highlighting source system disagreements: Source: Opendoor, with permission</figcaption></figure><h4><strong>Define a healthy table</strong></h4><p>Data observability, put simply, means applying frameworks that quantify the health of dynamic tables. To check if the rows and columns of your table are what you expect them to be, consider these factors and questions:</p><p><strong>Rows</strong></p><ul><li>Freshness — when was the data last updated?</li><li>Volume — how many rows were added or updated recently?</li><li>Duplicates — are any rows redundant?</li></ul><p><strong>Columns:</strong></p><ul><li>Schema — are all the columns you expect present (and some columns you don’t?)</li><li>Distributions — how have statistics that describe the data changed?</li></ul><p>Freshness, volume, duplicate, and schema checks are all relatively easy to implement with deterministic checks (that is, if you expect the shape of your data to be stable over time).</p><p>Or you can assess these with simple time-series models that adjust deterministic check parameters over time if the shape of your data is changing in a gradual and predictable way. For example, if you’re growing customer volume by X%, you can set the row volume check to have an acceptable window that moves up over time in line with X. At Opendoor, we know that very few real estate transactions tend to occur on holidays, so we’ve been able to set rules that adjust alerting windows on those days.</p><p>Column distribution checks are where most of the complexity and focus ends up being. They tend to be the hardest to get right, but provide the highest reward when done well. Types of column distribution checks include the following:</p><ul><li>Numerical — mean, median, Xth percentile, …</li><li>Categorical — column cardinality, most common value, 2nd most common value, …</li><li>Percent null</li></ul><p>When your tables are healthy, analytics and product teams can be confident that downstream uses and data-driven insights are solid and that they are building on a reliable foundation. When tables are not healthy, all downstream applications require a critical eye.</p><h4><strong>Anomaly detection</strong></h4><p>Having a framework for data health is a helpful first step, but it’s critical to be able to turn that framework into code that runs reliably, generates useful alerts, and is easy to configure and maintain. Here are several things to consider as you go from data quality abstractions to launching a live anomaly detection system:</p><ul><li><strong>Detection logic</strong>: If it’s easy to define in advance what constitutes row- or column-level violations, a system focused on deterministic checks (where the developer manually writes those out) is probably best. If you know an anomaly when you see it (but can’t describe it in advance via deterministic rules), then a system focused on probabilistic detection is likely better. The same is true if the number of key tables requiring checks is so great that manually writing out the logic is infeasible.</li><li><strong>Integrations</strong>: Your system should integrate with the core systems you already have, including databases, alerting (e.g., PagerDuty), and — if you have one — a data catalog (e.g., SelectStar).</li><li><strong>Cost</strong>: If you have a small eng team but budget is no barrier, skew towards a third-party solution. If you have a small budget but a large engineering team — and highly unique needs — skew towards a first-party solution built in-house.</li><li><strong>Data types</strong>: Anomaly detection looks different depending on if the data is structured, semi-structured, or unstructured, so it’s important to know what you’re working with.</li></ul><p>When it comes to detecting anomalies in unstructured data (e.g., text, images, video, audio), it’s difficult to calculate meaningful column-level descriptive statistics. Unstructured data is high dimensional — for instance, a small 100x100 pixel image may have 30,000 values (10,000 pixels x three colors). Rather than checking for shifts in image types across 10,000 columns in a database, you can instead translate images into a small number of dimensions and apply column-level checks to those. This dimensionality-reduction process is called embedding the data, and it can be applied to any unstructured data format.</p><p>Here’s an example we’ve encountered at Opendoor: we receive 100,000 images on Day 1, and 20% are labeled “is_kitchen_image=True” . The next day, we receive 100,000 images and 50% are labeled “is_kitchen_image= False”. That’s possibly correct — but the size of the distributional shift should definitely lead to an anomaly alert!</p><p>If your team is focused on unstructured data, consider anomaly detection that has built-in embeddings support.</p><h4><strong>Automated data catalogs</strong></h4><p>Automating your data catalog makes data more accessible to developers, analysts, and non-technical teammates, which leads to better, data-driven decision-making. As you build out your data catalog, here are a few key questions to ask:</p><p><strong>Table documentation</strong></p><ul><li>What does each row represent?</li><li>What does each column represent?</li><li>Table ownership — when there is a problem with the table, who in the organization do I call?</li></ul><p><strong>Table lineage (code relationships)</strong></p><ul><li>What tables are upstream? How are they queried or transformed?</li><li>What tables, dashboards, or reports are downstream?</li></ul><p><strong>Real-world use</strong></p><ul><li>How popular is this table?</li><li>How is this table and/or column commonly used in queries?</li><li>Who in my organization uses this table?</li></ul><p>At Opendoor, we’ve found that table documentation is challenging to automate, and the key to success has been a clear delineation of responsibility amongst our engineering and analytics teams for filling out these definitions in a well-defined place. On the other hand, we’ve found that automatically detecting table lineage and real-world use (via parsing of SQL code, both code checked into Github and more “ad hoc” SQL powering dashboards) has given us high coverage and accuracy for these pieces of metadata, without the need for manual metadata annotations.</p><p>The result is that people know where to find data, what data to use (and not use) and they better understand what they are using.</p><h4><strong>An ML-specific strategy</strong></h4><p>ML data is different when it comes to data observability for two reasons. First, ML code paths are often ripe for subtle bugs. ML systems often have two code paths that do similar but slightly different things: model training, focused on parallel computation and tolerating high latency, and model serving, focused on low latency computation and often done sequentially. These dual code paths present opportunities for bugs to reach serving, especially if testing is focused just on the training path. This challenge can be addressed with two strategies:</p><ul><li><strong>Assess serving inferences using a “golden set” (or “testing in prod”)</strong>. Start by assembling a set of inputs where the correct output is known in advance, or at least known within reasonably tight bounds (e.g., a set of home prices where Opendoor has high confidence in the sales prices). Next, query your production system for these inputs and compare the product system outputs with the “ground truth.”</li><li><strong>Apply distribution checks to serving inputs</strong>. Let’s say Opendoor trains our model using data where the distribution of home square footage is 1,000 square feet in the 25th percentile, 2,000 square feet in the 50th percentile, and 3,000 square feet in the 75th percentile. We would establish bounds based on this distribution — for instance, the 25th percentile should be 1,000 square feet +/- 10% — and collect calls to the serving system and run the checks for each batch.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*frpLHLSsYV4K1Og_" /><figcaption>Source: image by author</figcaption></figure><p>The other way that ML data differs in terms of data observability is that “correct” output is not always obvious. Oftentimes, users won’t know what is a bug, or they may not be incentivized to report it. To address this, analytics and ML teams can solicit user feedback, aggregate it and analyze the trends for external users and internal users/domain experts.</p><p>Whether focusing on ML data or your entire repository, data observability can make your life easier. It helps analytics and ML teams gain insight into system performance and health, improve end-to-end visibility and monitoring across disconnected tools and quickly identify issues no matter where they come from. As digital businesses continue to evolve, grow and transform, establishing this healthy foundation will make all the difference.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=421f6089d905" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/data-observability-for-analytics-and-ml-teams-421f6089d905">Data Observability for Analytics and ML teams</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Untangling the Paradoxes of Inflation]]></title>
            <link>https://medium.com/semi-random-thoughts/untangling-the-paradoxes-of-inflation-6f63e187e374?source=rss-cbfd810ae7b5------2</link>
            <guid isPermaLink="false">https://medium.com/p/6f63e187e374</guid>
            <category><![CDATA[real-estate]]></category>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[public-policy]]></category>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[inflation]]></category>
            <dc:creator><![CDATA[Sam Stone]]></dc:creator>
            <pubDate>Mon, 08 Aug 2022 15:59:45 GMT</pubDate>
            <atom:updated>2022-08-08T15:59:45.210Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/644/0*MVTUIWwu4a1LK7vK" /><figcaption>M.C. Escher — Ascending and Descending (1960)</figcaption></figure><p>Unless you’ve been living in a cave recently, you’ve seen alarming headlines about rising inflation. You might be wondering why inflation affects certain industries and products more than others, or what you can do to protect yourself from inflation. I find these questions interesting (and also scary), both from a personal perspective and because of my day job. I work at Opendoor, where I help manage our portfolio of residential real estate — and real estate and real estate debt are asset classes that are sensitive to inflation and interest rate changes. Helping steer our portfolio through macroeconomic turbulence is key to our company’s success, so it’s important for me to understand how inflation works, as well as the economic policy system that tasked with managing it (primarily the Fed).</p><p>Predictions of how inflation will trend in the future are a dime a dozen. I have no crystal ball that gives me anything to add there. However, I’ve learned a number of surprising, important things about how inflation and the Fed work, which explain things that seem paradoxical at first glance. I think these “mechanical” aspects of inflation and economic policy are useful for making better decisions when it comes to big life purchases, investments, jobs and even whom to vote for and why.</p><h3>1. Why inflation crushes stock prices, especially growth companies</h3><p>Inflation itself doesn’t directly depress stock prices. In fact, by definition, inflation means prices are <em>rising</em>, although most definitions of “price level”, like the Consumer Purchase Index (“CPI”) don’t include stocks in their basket of goods [Footnote 1]. The problem starts when inflation exceeds the Fed’s target rate (more on this later) and investors begin to expect the Fed will raise the Fed Funds Rate. (If you’re not familiar with why the Fed would raise interest rates to curb inflation, <a href="https://www.nytimes.com/2022/05/14/business/inflation-interest-rates.html#:~:text=So%20how%20might%20raising%20interest,more%20expensive%20to%20borrow%20money.">this article</a> explains it nicely.)</p><p>Why does a higher Fed Fund Rate (“FFR”) matter for the stock market? The FFR is the rate at which large banks can borrow from each other overnight, via reserves held with the Fed [2]. This is essentially a “risk-free” interest rate, due to the collateral held by these institutions and the extremely short duration of the loans [3]. All other loans (like company-to-company, mortgages, etc.) carry additional risk, and thus lenders charge some “spread”, meaning an interest rate higher than the FFR.</p><p>When the Fed raises the FFR, this means that the “time value of money” has gone up. In other words, if yesterday the FFR was 1% and today it’s 2%,if you were to lend $100 to a totally reliable (i.e., “risk-free”) friend, yesterday you would have required that friend to pay you back $101 in a year, but today you require $102.</p><p>We can value stocks (or bonds) in the same way. We give up some money today (when we make the investment) to receive some future cash flows. Our willingness to pay today is based on how much we discount those future cash flows (this is <a href="https://www.investopedia.com/terms/d/dcf.asp">Discounted Cash Flow valuation</a>).</p><p>Let’s consider the stock value (i.e., market cap) of two highly simplified, illustrative companies. Investors expect Steady Eddie Inc. to generate $100 in cash flow every year, to infinity. On the other hand, investors expect Speculative Sally Co. to generate no cash flow for the next 3 years, and then $133 in year 4 and beyond. With an FFR (or “risk-free rate”) of 10%, both companies are worth the same $1,000. (In reality, the FFR today is ~1.6%; 10% is far from low, but this is just an illustrative example and a big number makes the following math easier to grasp.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uj_q4sYEsQt8BsrZ4ZjqMw.png" /></figure><p>What happens when the Fed raises interest rates (in this case, to 15%)? Let’s assume that this has no impact on the business fundamentals of either company, i.e., their expected cash flows stay exactly the same, in nominal dollars. The only thing that changes is how investors discount those cash flows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dhV6ZjhFcEFcSJe3RVM7XQ.png" /></figure><p>Two things stand out here:</p><ul><li>Both companies’ valuations dropped substantially, despite no change in investors’ expectations of the businesses’ performance. In other words, even in a scenario where the Fed is raising rates and investors do not increasingly fear a recession (a very rosy scenario), valuations still drop.</li><li>Speculative Sally Co. dropped substantially more than Steady Eddie Inc (-42% vs -33%, respectively). This is because Sally Co.’s profits are skewed towards the future, and future cash flows get discounted more.</li></ul><p>This is not to say all or even most investors use literal DCF models. However, other investment approaches will often push investors towards similar behaviors. For example, an investor that believed both Steady Eddie and Speculative Sally had a similar but just slightly better risk-adjusted return vs. a Treasury Note in Scenario A, is likely to prefer the Treasury Note in Scenario B. So they sell off their shares of Steady Eddy and Speculative Sally to go buy Treasury Notes — which has the same price effects as the DCF model.</p><p>The illustrative model has directionally similar results to what has actually happened in US stock markets in 2022 YTD. All major indices have dropped. And the indices that are weighted towards speculative companies (“speculative” meaning expected profits are further out in the future, which skews towards high growth and/or tech companies) have dropped more.</p><h3>2. Why you shouldn’t rush to get a mortgage before rates move up</h3><p>A curious thing happened on June 15, 2022: the Fed increased its target FFR rate by 75 bps — the largest increase in 28 years — but instead of increasing, mortgage rates <strong><em>decreased</em></strong> by approximately 15 basis points that day.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3h9bVZM5kwOpo1VY" /><figcaption><a href="https://www.nerdwallet.com/mortgages/mortgage-rates">https://www.nerdwallet.com/mortgages/mortgage-rates</a></figcaption></figure><p>As discussed above, a rising FFR means all other forms of debt should also get more expensive, all else equal. So how should we reconcile the June 15 FFR increase with an accompanying mortgage rate decrease? A fluke? Does it disprove the economic fundamentals?</p><p>Not exactly. The solution to this seeming paradox is that most investors had (a) been expecting the Fed to increase the FFR by 75 bps on June 15, long before they officially announced it on June 15 and (b) interpreted Fed commentary published on June 15 (“new information”) to indicate the Fed was likely to raise the FFR at its upcoming meetings less than they had previously expected. To unpack why this matters for mortgage rates, we need two concepts for understanding how the FFR “cascades” through other debt classes: (i) debt duration and (ii) the forward expected path of the FFR, over the debt duration.</p><p>Imagine you’re a bank with $100 in capital and a 5-year horizon. With that capital, you can either invest it in 1-year government bonds that you’ll “roll over” every year into a new bond, or make a 5 year loan (to a consumer, like a home-buyer). If you go with the government bond option, since you’re actually buying a new bond each year, the interest rate you receive changes every year. So you need to forecast what the FFR will be.</p><p>Let’s say that today the FFR is 5%, but you think the FFR will go up to 10%, on average, in year 2, but then in years 3–5 the FFR will return to a flat 3%. Directionally this is similar to what a lot of investors actually think today: they believe the Fed will have to continue to raise rates into 2023 to “turn the tide” against inflation, and then the Fed will need to reduce the FFR to revitalize economic growth after inflation is in check. Per the chart below, this results in a payment stream with an <a href="https://www.investopedia.com/terms/i/irr.asp">internal rate of return</a> of 4.3%. In other words, you’re indifferent between buying the government bonds (where the expected nominal payment varies year by year) and issuing a 5-year consumer loan that pays you 4.3% per year, every year. (This simplified example assumes the government loan and the consumer loan carry the same risk, which is obviously not true, but the directional point still holds.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*PglCqXUfsfrJlOUy56UkgA.png" /></figure><p>Now let’s say there’s a Fed meeting in September and the Fed announces it’s raising rates, but the increase is less than the path you expected that led to 10% in year 2. You’re still forecasting an increase, but only to 7% in year 2. This means that your “indifference price”, in terms of the interest you’d charge on the consumer loan to receive the equivalent vs. the government debt, has fallen to 3.7%:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/938/1*ATmWcQ5oRas3hW7XGaCbQQ.png" /></figure><p>So what does this all mean? We can be highly confident that, in the near term, the Fed will continue to raise interest rates [4]. However, those expectations are already incorporated into market interest rates offered by lenders on longer-duration debt, like consumer loans. If the Fed raises interest rates as the market expects, then don’t expect major changes in the cost of those other loans. And if the Fed raises interest rates less than currently expected, those other loans will likely become less costly — and vice versa.</p><p>The key point is that rates on longer-duration debt, like mortgages or most other loan types relevant to consumers, already anticipate and incorporate expected changes in the FFR. The only reason you would want to lock in a mortgage rate now is if you think the Fed is going to raise rates more than the market does.</p><p>Of course there’s also the question of whether you think a home purchase itself (separate from the financing question) is a good investment right now. Rising rates are intended by the Fed to moderate demand, which includes homebuyer demand — which puts downward pressure on prices. At the same time, inflation itself makes real assets (like homes or gold) more attractive — which puts upward pressure on prices.</p><h3>3. Why inflation is bad — but we still want some of it</h3><p>With so much negative news around inflation, and the framing of inflation as something that seems like a pure economic evil, it may seem strange that the Fed targets 2% inflation, not zero, and that it considers negative misses (coming in below 2%) just as bad as an equivalently sized positive miss. It hasn’t always been the case that the Fed targets 2% inflation, and there is some debate amongst economists about the 2% target, but it’s pretty well-established — and for a few good reasons [5].</p><p>First, a modest amount of inflation helps keep the labor market functioning well (and thus helps the Fed to achieve its second goal, after stable prices: maximizing employment) [6]. This is due to a psychological / behavioral phenomenon called downward wage rigidity. This stems from the fact that workers tend to focus more on nominal wage changes than real wage changes (“real” refers to the nominal change less inflation). The concept is pretty simple: which situation, as an employee, would you prefer?</p><ul><li>A: 1% pay cut (nominal) in a 0% inflation world (1% real wage loss)</li><li>B: 2% pay raise (nominal) in a 4% inflation world (2% real wage loss)</li></ul><p>Empirically, most employees prefer (B) over (A), even though (B) is worse in pure economic terms, because a nominal pay cut is psychologically jarring [7]. And employers know that wage cuts are jarring and bad for morale, so they have a tendency to do layoffs rather than (nominal) wage cuts when times get tough. However, as long as there is some inflation, when times get tough, businesses have some room to reduce real wages without reducing nominal wages. This leads to fewer layoffs, higher employment, and is generally considered better for the economy.</p><p>Second, there is widespread (but not universal) belief amongst economists that deflation is worse than an equivalent amount of inflation, especially during a recession. When deflation occurs, the value of money increases over time. This leads to a number of outcomes that tend to exacerbate a recession. Deflation incentivizes delaying consumption to hoard cash (if I delay spending $1 today, I will be able to buy more with it tomorrow). This is damaging to businesses. Deflation also makes interest payments (which are normally fixed in nominal dollar amounts) more expensive in real terms. In other words, it becomes harder to pay off debts, and debtors are often the people already struggling the most in a recession. Finally, deflation can lead to what’s called a liquidity trap. When the Fed has cut rates all the way to zero (to stimulate lending / growth / inflation), but deflation still persists, at this point the FFR “lever” no longer functions. The Fed cannot cut below zero, and it has to start using other less effective tools, at the very same time that monetary stimulus is most needed [8].</p><p>Given the risks of recessionary deflation outweigh the risk of recessionary inflation, and the fact that managing the economy is very difficult with long feedback cycles (creating lots of room to miss a target), it makes sense for the Fed to build in some buffer, targeting positive but modest inflation. I love this <a href="https://www.federalreserve.gov/boarddocs/speeches/2004/20041202/default.htm">quote</a> from former Fed Chairman Ben Bernanke:</p><blockquote>If making monetary policy is like driving a car, then the car is one that has an unreliable speedometer, a foggy windshield, and a tendency to respond unpredictably and with a delay to the accelerator or the brake.”</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1kQ7cK6rwimtYFWh.png" /></figure><p>What’s the upshot of understanding the mechanics of inflation and the Fed? It doesn’t necessarily make it any easier to predict inflation. But it can help us make better personal financial decisions. Asset classes react differently to inflation; if you think inflation is going to increase, then you want to avoid investments with far-out profits (like high-growth tech stocks) and vice versa. And just because the Fed is going to raise rates doesn’t mean rates for personal loans, like mortgages, are going to rise. (If a mortgage lender is pressuring you to “buy a house now, before rates rise,” don’t succumb to that!)</p><p>Inflation is also shaped by fiscal policy, which is directly tied to issues that we, as voters, get to influence. The Fed controls monetary policy, and the Fed is shielded from politics [9]. The Federal government (Congress and the President) set fiscal policy. Fiscal policy also matters for inflation, although the cause-and-effect relationships are less clear than with monetary policy. But federal government fiscal policies (or lack thereof) that spike demand (like stimulus checks) or decrease supply (like port shutdowns) also affect near-term price levels. So inflation is (and should be) an issue that informs voters’ assessments of candidates’ policies, which makes understanding the mechanics of the system that produces inflation all the more important.</p><p><strong>[Footnote 1] </strong>The CPI excludes investment items, such as stocks, bonds, real estate, and life insurance, as “these items relate to savings and not to day-to-day consumption expenses” However, rent and “owner’s equivalent of rent” are included in the CPI, so the one major investment asset class that is, arguably, directly reflected in the CPI is residential real estate. <a href="https://webapps.dol.gov/dolfaq/go-dol-faq.asp?faqid=94">Link</a></p><p><strong>[Footnote 2] </strong>The FFR is not the only mechanism by which the Fed changes interest rates. In addition to the FFR, the Fed directly charges for overnight loans to large banks via the “discount rate”. This is normally ~1% higher than the FFR, as the Fed prefers that banks borrow from each other, rather than Fed directly, so that banks monitor each other for credit or liquidity issues. The Fed also manipulates longer-term interest rates through “Open Market Operations” where it buys or sells government securities (Treasury bills, notes, and bonds). In this blog post, I focus on the FFR for simplicity; in the real-world, the Fed is regularly adjusting all of these levers, although they tend to adjust them in near lock-step.</p><p><strong>[Footnote 3] </strong>But wait — large banks are not as good a credit risk as the US government, they’re definitely not “risk free”? You might have heard that the closest thing to a truly risk-free investment is a short duration US Treasury, like the 1 month T-Bill. And the FFR tracks the 1-month T-bill yield nearly exactly!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zg66JSq38WuhzzJi" /><figcaption><a href="https://fred.stlouisfed.org/graph/?g=fisb">https://fred.stlouisfed.org/graph/?g=fisb</a></figcaption></figure><p><strong>[Footnote 4]</strong> After each Fed meeting, the Fed publishes what each Fed governor expects rates to be 1, 2 and 3+ years in the future (this is called the “dot plot” in macroeconomic jargon.). After the last meeting in June, <strong>all</strong> of the Fed Governors said they expected the FFR to be above 3% by the end of 2022 (vs 1.6% today). Since these are the expectations of the people who will<strong> determine</strong> what the interest rate will be, it’s not speculative to say the rate will rise. <a href="https://www.bloomberg.com/news/articles/2022-06-15/the-fed-s-new-dot-plot-after-its-june-policy-meeting-chart">Link</a></p><p><strong>[Footnote 5] </strong>The Fed adopted an explicit 2% inflation long-term target only in 2012; then Fed Governor James Bullard argued at the time the Fed had been implicitly following this target since 1995 (<a href="https://www.stlouisfed.org/open-vault/2019/january/fed-inflation-target-2-percent">Link</a>). So it’s not exactly new. The 2% may seem a bit arbitrary, but there is theoretical backing for this, including the <a href="https://en.wikipedia.org/wiki/Taylor_rule">Taylor Rule</a>, which recommends the FFR respond to (a) actual inflation rates less target inflation rates and (b) actual GDP growth less target GDP growth.</p><p><strong>[Footnote 6]</strong> From the Richmond Fed: <em>Since 1977, the Federal Reserve has operated under a mandate from Congress to “promote effectively the goals of maximum employment, stable prices, and moderate long term interest rates” — what is now commonly referred to as the Fed’s “dual mandate.” </em>This is somewhat controversial; some economists think the mandate should be narrowed to just stable prices, for example the <a href="https://www.stlouisfed.org/open-vault/2018/august/federal-reserve-dual-mandate">President of the St Louis Fed</a>.</p><p><strong>[Footnote 7] </strong>Two papers demonstrating empirical findings of downward wage rigidity in the US: <a href="https://www.federalreserve.gov/econresdata/feds/2016/files/2016001pap.pdf">1</a>, <a href="https://econ.columbia.edu/wp-content/uploads/sites/41/2018/09/Jo_Columbia_JMP.pdf">2</a></p><p><strong>[Footnote 8]</strong> There’s a fair amount of debate amongst economists about how real “liquidity traps” are, what causes them, and if they have a connection. However, it is very clear that when the Fed cuts rates to zero and the economy remains in recession, it has to rely on other tools — for example, rates were zero coming out of the Great Financial Crisis of 2008–2010, so the Fed had to rely on Quantitative Easing (a form of Open Market Operations).</p><p><strong>[Footnote 9] </strong>The Fed governors are appointed by the President and confirmed by the Senate. They have 14 year terms and cannot be reappointed. This is designed to make them independent of political cycles and even Presidential administrations. Promoting a politically independent central bank is something that is widely supported by economists, across the world. The 7 Fed Governors are not to be confused with the 12 Fed Regional Reserve Bank Presidents. The Presidents rotate on and off the Federal Open Market Committee (FOMC) that sets key interest rates; the Governors always have a vote.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6f63e187e374" width="1" height="1" alt=""><hr><p><a href="https://medium.com/semi-random-thoughts/untangling-the-paradoxes-of-inflation-6f63e187e374">Untangling the Paradoxes of Inflation</a> was originally published in <a href="https://medium.com/semi-random-thoughts">Structured Ramblings</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Aligning Product with Strategy by Unifying Models]]></title>
            <link>https://medium.com/opendoor-labs/aligning-product-with-strategy-by-unifying-models-4fe9573f83b3?source=rss-cbfd810ae7b5------2</link>
            <guid isPermaLink="false">https://medium.com/p/4fe9573f83b3</guid>
            <category><![CDATA[pricing-strategy]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[real-estate]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Sam Stone]]></dc:creator>
            <pubDate>Wed, 20 Jul 2022 18:45:44 GMT</pubDate>
            <atom:updated>2022-07-20T19:40:38.050Z</atom:updated>
            <content:encoded><![CDATA[<h4>How unifying pricing models moves from tactical to strategic, to better power life’s progress one move at a time</h4><p><em>By Sam Stone, Director of Product Management, and Amit Misra, Staff Research Scientist</em></p><figure><img alt="abstract image of two home icons with a complex shape of lines and dots in the middle" src="https://cdn-images-1.medium.com/max/1024/1*QQW_6fYZkA-3ZGe0DiRodA.jpeg" /></figure><h3>Intro</h3><p>At Opendoor, our mission is to power life’s progress one move at a time. Selling your home and buying a new one is complex; a successful move necessitates literally hundreds of decisions and tasks. Advancing our mission requires Opendoor product teams to understand all these decisions and tasks, identify those that generate the most frustration or inefficiency, and relentlessly chip away at those pain points to improve the experience for our customers and reimagine the real estate process.</p><p>In our early days, product teams identified important decisions and tasks where models offered to improve quality and/or efficiency. The bottom-up, task-oriented nature of early model development led to models which were valuable and effective — but highly <strong>tactical</strong>.</p><blockquote><em>Strategy without tactics is the slowest route to victory. Tactics without strategy is the noise before defeat.</em></blockquote><blockquote><em>— Sun Tzu</em></blockquote><p>As Opendoor has grown, unifying models across related tasks has expanded their benefits from tactical to <strong>strategic</strong>. These benefits flow through to our customers, both sellers and buyers — the more accurate and efficient Opendoor can be in navigating the lifecycle of home buying and selling, the less financial risk we take on. And the less risk we have, the higher prices we can offer sellers, the lower prices we can charge buyers, and the more homeowners we can serve. In other words, model unification feeds directly into powering life’s progress one move at a time.</p><p>While model unification might sound like the pure purview of data science, product managers are heavily involved in it — success here also depends on engineering, operations, and many other stakeholders. In short, a classic product management challenge! In fact, this (largely non-technical) blog post is written primarily for product managers interested in machine learning products‚ although we believe data scientists, machine learning (ML) engineers, and anyone else who works on ML is likely to find it interesting and, hopefully, useful for their own work.</p><p>We’ll explore model unification through a real journey that Opendoor went through recently, related to how we sell homes that we’ve acquired (a process we call “resale”). There are three major decision types in the resale lifecycle for each home:</p><ul><li>When the home is first listed, set the home’s initial list price</li><li>Each week that the home is on the market, decide whether to drop the list price, and if so, by how much</li><li>Each time a buyer offer is received, decide whether to accept it</li></ul><p><em>Overview of the Resale process for Opendoor-owned homes</em></p><figure><img alt="a timeline showing price drops driven by a disparate models" src="https://cdn-images-1.medium.com/max/1024/0*MUhiIIGytJBvu3dO" /></figure><p>Initially, all of these tasks were done manually by our in-house team of portfolio managers, with little support from customized software (and no model support). First, we developed a model for recommending price drops. We then developed a second model for setting initial list prices. And then we developed a third model for deciding on buyer offers.</p><h3>Deciding to unify our models</h3><p>Identifying the above-mentioned structural issues with stacked and parallel models was only the first step towards the decision to unify our resale models. To put model unification into our roadmap, we had to move from identification of a theoretical problem to sizing the benefits and costs of solving that problem in real-world terms.</p><p>Estimating the benefits of model unification had two parts. First, we began to estimate the direct business benefit of joint optimization of resale decisions. In other words, by removing the “conditional assumptions” inherent in the stacking and parallel setup, how much more accurate would a unified model be than the “sum of the parts” from the multiple model status quo. And then how would that increase in accuracy translate into business outcomes like customer volume and customer-facing prices? There are lots of models that are theoretically “wrong” but where “correcting” the theoretical problem yields negligible customer benefits; we wanted to make sure we were not headed down this path. Product management started with a back-of-the-envelope estimate (done in a spreadsheet) of the accuracy benefit from model unification; when that looked promising, data science built a simple, backtestable prototype of a unified model and produced a more refined estimate with it.</p><p>Product management also explored more qualitative benefits of model unification. Some of these were technical, like the fact that building a new (unified) model would allow us to migrate to newer infrastructure. Some were more business-facing, like the fact that a new model could natively incorporate unusual home types that the status quo models did not support.</p><p>With the prototype model in hand, we then estimated the costs. In the short term, there would be a large cost in both engineering and data science team time to take the prototype model to production — time estimates that our lead engineers and data scientists were responsible for making and detailing. This was partially mitigated by estimates that the long-term maintenance of the unified model would be significantly lower than the status quo.</p><p>Finally, with the cost/benefit analysis in hand, product management (in conjunction with data science and engineering) compared this project to others on our roadmap, slotting it into the cross-functional group’s backlog (the placement was quite high, e.g., “Let’s get going ASAP!”)</p><h3>Implementing a unified model</h3><p>Three decisions were key to implementing the unified resale model. First, we chose to keep the model deployed in the initial release quite simple, but with strong support for extensibility and future development. In fact, the initial version was so simple it did not include some signals that we knew mattered and had been incorporated in the precursor models. The decision to exclude these signals was hotly debated! The root of the issue was that different information (or “signals”) is available on a home at different events in the resale lifecycle — as a home sits on the market, not only does more information accumulate, but entirely new types of information become available.</p><p><em>Availability of information by home lifecycle event</em></p><figure><img alt="a timeline showing price drops driven by a unified model" src="https://cdn-images-1.medium.com/max/1024/0*lba_7p1Fxso6tHmq" /></figure><p>Previously, since there had been a separate model for each event, each model used all the information that was available for that event and nothing more. But with a unified model, there was new complexity in deciding what information should be processed for each event. We were able to show that with just the minimum set of information — information relevant to all events — a unified model would still outperform the status quo. That led us to a “sparse” v0 unified model, but it also led us to build it in such a way that adding in additional signals would be straightforward, as we knew that would be follow-up work.</p><p>Second, we built the unified model in conjunction with heuristics and operator guidance focused on the areas we knew the unified model would be weakest initially. For example, while we didn’t include certain signals in the v0 unified model, that didn’t mean we ignored them. Instead, we developed simple ways of surfacing these to our operations team, flagging situations where they check the model results due to the presence of those signals, and guidelines for consistent and principled model overrides or adjustments.</p><p>Finally, we invested in educating our operations team on the “why” behind the new unified model. This was critical to building an effective feedback system, in which operators flagged truly problematic model recommendations to our product teams, and did not raise a high volume of false alarms. The new unified model generated recommendations that often looked quite different from the precursor models, which could have elicited lots of false alarm feedback. For example, the unified model allowed us to implement new types of strategy shifts through resale decisions, and to implement them very quickly. When a macroeconomic event occurred, we could now respond the next day with new jointly-optimized recommendations for all homes in our inventory. This meant that significant day-over-day variations in model recommendations were often an expected behavior, something that our operators did not need to flag. Conversely, significant variation in model recommendations on the same day between similar homes was generally unexpected, and something we wanted operators to flag.</p><h3>Where to next?</h3><p>Model unification has proven to be a powerful way to align our products with our strategy. And this is not just limited to resale; we’ve observed similar benefits across many different product areas. As we look ahead, we’re excited about two things that build on (and contribute back to) model unification. First, integrating models across entirely different task groups (as opposed to unifying models within a task group). An example of this is home acquisition versus home resale; they’re different task groups, but they both are key to managing inventory and portfolio risk. Model integration offers benefits that go beyond simple manual attempts at reconciliation. We’re also excited about new venues for experimentation, to generate counterfactual data that helps us understand what would happen if we deviated from our status quo systems. This information is critical to understand how far our systems are from optimal, and it shapes our roadmap around where and how to improve model-driven systems.</p><p>Unifying models is not easy. It requires collaboration across product, data science, engineering, operations, and many other disciplines. At Opendoor, product management plays a key part not just in coordinating this type of product development, but also in laying out the strategic vision that motivates. We believe in working backwards, starting with our mission, and identifying the largest opportunities to advance it. And we’ve found unifying models, despite its challenges, to be a consistent winner in that regard.</p><p>Interested in joining our team? Opendoor is hiring across all teams, including Product Management and Data Science! Check out our <a href="https://www.opendoor.com/w/careers/open-positions">current open positions</a>!</p><p>Want to learn more about working at Opendoor? Check out our <a href="https://medium.com/opendoor-labs/tagged/product-management">Product Management</a> and <a href="https://medium.com/opendoor-labs">Engineering and Data Science</a> blogs.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4fe9573f83b3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/opendoor-labs/aligning-product-with-strategy-by-unifying-models-4fe9573f83b3">Aligning Product with Strategy by Unifying Models</a> was originally published in <a href="https://medium.com/opendoor-labs">Open House</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Constraints & Helping Users Make Better Decisions]]></title>
            <link>https://medium.com/semi-random-thoughts/constraints-helping-users-make-better-decisions-be2e37367a42?source=rss-cbfd810ae7b5------2</link>
            <guid isPermaLink="false">https://medium.com/p/be2e37367a42</guid>
            <category><![CDATA[user-experience]]></category>
            <category><![CDATA[product-design]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[design-thinking]]></category>
            <category><![CDATA[behavioral-economics]]></category>
            <dc:creator><![CDATA[Sam Stone]]></dc:creator>
            <pubDate>Mon, 31 Jan 2022 05:28:15 GMT</pubDate>
            <atom:updated>2022-02-01T16:52:10.432Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lw0VXNR9gZaus2Yj.jpg" /></figure><blockquote>Constraints are powerful clues, limiting the set of possible actions. The thoughtful use of constraints in design lets people readily determine the proper course of action, even in a novel situation. — Don Norman, The Design of Everyday Things</blockquote><p>Constraints are core to good product design. Consider the menu at Chipotle (burrito/bowl/salad/tacos), an example of a retailer thats grown sales by helping users avoid the “paradox of choice.” Or a car angrily beeping when you start driving without a seat belt, an example of a hardware manufacturer helping users drive more safely.</p><p>When it comes to digital products, many have a clean, simple visual design (one form of constraint), on top of an utterly unconstrained, sprawling mess of functionality underneath. With digital products, when adding a new feature can be as easy as adding a few lines of code — and when users say they <em>really</em> want that new feature — why wouldn’t you add it?</p><p>You might not add it because, perhaps, the short-term benefit from that feature is actually at odds with users’ long-term interests. This brings us to a particularly interesting type of constraint: those that cause users short-term dissatisfaction, but help users make better long-term decisions. In this post, we’ll explore four digital products that do just that, drawn from hiring, gaming, investing, and publishing.</p><h3><strong>Lever: The Group-Think Constraint</strong></h3><p>Lever is a SaaS product that helps teams hire better. Lever’s default settings around around job applicant interview feedback are particularly opinionated, and constrained.[1] After completing an interview, Lever prompts the interviewer to submit feedback on the candidate. There are two constraints in the default way feedback is solicited, constraints that have made me uncomfortable as an interviewer — but also helped me make better decisions.</p><p>First, each candidate must be scored on a scale of 1 (“strong no”) to 4 (“strong yes”). At most companies, 3 means “yes, I support hiring this candidate” and 2 means “no, I do not”. Due to the even-numbered scale, there is no “neutral” rating. The lack of a neutral rating forces the interviewer to support one of two potential actions (hire or don’t hire); it raises the stakes.</p><p>This brings us to the second constraint. In this high-stakes decision, where I have to “take a side” as an interviewer, sometimes I don’t have a strong opinion on the candidate. I’m uncertain, and I want to avoid “getting it wrong” — being the lone panelist with a rating that deviates from the rest of my colleagues on the panel. In these situations, I want to see what the other panelists have rated the candidate, and align my rating with theirs. But Lever doesn’t let you do this! Other interviewers’ scores are hidden until you submit your own feedback, even if you’re the hiring manager.</p><p>This design is uncomfortable for many interviewers — but it leads to better hiring decisions. In <a href="https://www.amazon.com/Noise-Human-Judgment-Daniel-Kahneman/dp/0316451401"><em>Noise</em></a>, psychologist Daniel Kahneman and legal scholar Cass Sunstein explore how aggregating individuals’ judgments can generate better decisions, but only if the information flow is gated, to ensure judgements truly are independent.[2] Using Google as an example, before Lever helped promote this strategy more widely, they write:</p><blockquote>Almost all companies aggregate the judgments of multiple interviewers on the same candidate…. To ensure [a high] level of validity [around the panel’s assessment of a candidate], however, Google stringently enforces a rule that not all companies observe: the company makes sure that the interviewers rate the candidate separately, before they communicate with one another. Once more: aggregation works — but only if the judgments are independent.</blockquote><h3><strong>Wordle: The Binging Constraint</strong></h3><p>Wordle is a game in which you try to guess a 5-letter word. Never 4, never 6, always 5. It was created by a solo developer (Josh Wardle) in October 2021, as a gift for his partner, not a business endeavor. Its popularity exploded in December and January, with millions of users by late January.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UYjxiLB-5S99OT1OKIgVAg.png" /><figcaption>Worldle: one and done each day</figcaption></figure><p>My wife introduced Wordle to me two weeks ago. It took me about 10 minutes to solve my first puzzle, with her coaching me over my shoulder. Excited to have solved it, I asked her how to advance to the next puzzle. “That’s it,” she said “there’s one Wordle per day. You can try the next one tomorrow.” I was surprised, annoyed — and intrigued.</p><blockquote>“Making Wordle I specifically rejected a bunch of the things you’re supposed to do for a mobile game,” [game creator] Wardle told NPR. He deliberately didn’t include push notifications, allow users to play endlessly or build in other tools commonly used today to pull users into playing apps for as long as possible.</blockquote><blockquote>Wardle said the rejection of those engagement tricks might have fueled the game’s popularity after all — “where the rejection of some of those things has actually attracted people to the game because it feels quite innocent and it just wants you to have fun with it.” (<a href="https://www.npr.org/2022/01/12/1071840091/wordle-word-game-tips">source</a>)</blockquote><p>I’m now two weeks into playing Wordle, and it’s become a daily habit. But I don’t feel bad about it — because it’s not something I can binge. It’s just 5–10 min each day, never more than that. Playing Wordle feels wholesome and healthy; it’s a little thing to look forward each day, something I know will never turn into “doom scrolling.”</p><p>And while I don’t know this for sure, since specific numbers on Wordle users are hard to come by, I’d bet that the “one game per day” constraint has been key to the game’s growth.</p><h3><strong>Vanguard: The Gambling Constraint</strong></h3><p>When we purchased our home, I sold a chunk of shares in Vanguard’s S&amp;P500 ETF for our down payment. About two weeks later, after the home had closed, it turned out we had withdrawn a bit too much from the ETF. So I logged back into Vanguard’s website and placed an order to transfer the surplus funds back into the S&amp;P500 ETF. I was surprised to see Vanguard reject my buy order. An error message informed me that after a user sells shares in a particular ETF, Vanguard freezes that user’s ability to buy shares in the same fund for the next 30 days.</p><blockquote>Vanguard’s <a href="https://vanguard.com/us/whatweoffer/overview/redemptionpolicy">Frequent-trading policy</a>: If you sell or exchange shares of a Vanguard fund, you will not be permitted to buy or exchange back into the same fund, in the same account, within 30 calendar days.</blockquote><p>I was puzzled. Why would Vanguard freeze me out for 30 days? On Robinhood or Coinbase, I can trade in and out of the same fund or asset every day — or every hour. Clearly the technology exists to support extremely frequent consumer trading. And why wouldn’t Vanguard want me trading more? More trades means more user engagement means more customer lifetime value, right?</p><p>As I puzzled over this, I went down a Wikipedia rabbit hole that led me to the philosophy of Vanguard founder, Jack Bogle. He evangelized a simple, long-term-oriented approach to investing: put your money into US equity and bond index funds, rebalance no more than once a year, and choose the asset manager with the lowest fees. He founded Vanguard to provide those low fees. The whole Vanguard platform was built to deliver on this value proposition — if you buy and hold, you can do so with extremely low fees. The company’s value proposition is not to help users <em>beat</em> the market, but rather to allow them to <em>participate</em> in the market (for better or worse) at low cost.</p><p>Now this “no trades for 30 days” constraint started to make more sense. It was Vanguard’s intentional and necessary way of differentiating itself from a platform like Robinhood, where the explicit value proposition is “low cost trades” and the implicit value proposition is “use this platform to beat the market.” The short-term frustration that Vanguard imposed on me were a useful reminder of what Vanguard was — and was not — and how I should be using their product: investing to participate in the market, at low cost.[3]</p><h3><strong>Dropbox Paper: The Tweaking Constraint</strong></h3><p>Paper is Dropbox’s document creation and collaboration tool, ostensibly similar to Google Docs or Microsoft Word. I first used Paper in 2019, and I was shocked to find it lacked features I had come to take for granted in a writing tool: the ability to choose a font. To set margins. To number pages (or even to have pages).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wTO9JLDu4FRCCSKjUp5JmA.png" /><figcaption>The extremely limited formatting options in Paper</figcaption></figure><p>Without these formatting options, the only thing I could to do with Paper was… write. The lack of functionality focused me entirely on the core of the task: producing clear and concise prose. At first, this was intimidating. I had become used to playing around with formatting when I hit a roadblock in my writing. But there was no such tweaking to distract me when using Paper. Just the half-finished paragraph with which I was struggling. Over time, I learned to love this enforced focus; the knowledge that I couldn’t change the formatting, even if I wanted to, was freeing.</p><p>(For what it’s worth, Medium’s writing interface is similarly sparse and a major reason I write this blog on Medium!)</p><p>We’re all subject to biases. Sometimes they’re shortcuts we take to make hard decisions feel easier. Sometimes they’re tricks we play on ourselves to justify short-term pleasures that carry with them long-term costs.</p><p>Daniel Kahneman calls this “fast thinking” and “slow thinking”; it’s easier to rely on mental shortcuts (fast thinking), but when we slow down and think things through, we tend to make better decisions. Some product design takes advantage of fast thinking. For example, the “autoplay next episode” feature on Netflix drives a lot of short-term user engagement, but in the long-term, it can leave you feeling wary of Netflix.</p><p>Other products, like the ones described above, take the opposite approach. They help users slow down, think a bit deeper, and make decisions that are better. This may come at a cost to short-term user satisfaction, but the plethora of successful products that embrace such constraints (which is certainly not limited to Lever, Wordle, Vanguard, and Dropbox), proves it can be a long-term winning strategy.</p><p><strong>Footnotes</strong></p><p>[1] The Lever settings described can be turned off by company administrators, but not by end users (hiring managers, interviewers).</p><p>[2] Sunstein is also the co-author of Nudge, along with Richard Thaler. Nudge is about “choice architecture”: how to help people make better decisions for themselves. This blog post on constraints is focused on a subset of that field.</p><p>[3] Vanguard is not the only asset manager to impose a 30-day trading limit after ETF and mutual fund sales; Fidelity has a similar constraint.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=be2e37367a42" width="1" height="1" alt=""><hr><p><a href="https://medium.com/semi-random-thoughts/constraints-helping-users-make-better-decisions-be2e37367a42">Constraints &amp; Helping Users Make Better Decisions</a> was originally published in <a href="https://medium.com/semi-random-thoughts">Structured Ramblings</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Lyndon Johnson]]></title>
            <link>https://medium.com/semi-random-thoughts/lyndon-johnson-2422c70ab08d?source=rss-cbfd810ae7b5------2</link>
            <guid isPermaLink="false">https://medium.com/p/2422c70ab08d</guid>
            <category><![CDATA[biography]]></category>
            <category><![CDATA[us-politics]]></category>
            <category><![CDATA[lyndon-johnson]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[robert-caro]]></category>
            <dc:creator><![CDATA[Sam Stone]]></dc:creator>
            <pubDate>Wed, 22 Dec 2021 04:03:03 GMT</pubDate>
            <atom:updated>2021-12-22T04:03:03.737Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X6kgwC62HMXJKxolyUv-Mw.jpeg" /></figure><p>I recently finished Robert Caro’s series “The Years of Lyndon Johnson”. It fundamentally changed my views on politics, and on leadership. Which seems only fitting, given the time it required — four volumes, thirty five hundred pages, and nearly six months of reading! It was well worth that time, as Caro’s work is different and more insightful than any other biography I’ve read. It’s investigative journalism, not “great man adulation”. LBJ and his circle built a robust mythology around his life; Caro pierces it repeatedly, revealing deeply loathsome actions and attributes of LBJ previously unpublished.</p><p>LBJ was complex. His personality had both admirable and repugnant attributes that ran to the core. He simultaneously caused some of the greatest achievements in US history (the Civil Rights Act, the Great Society) and some of our worst failures (the Vietnam War, widespread electoral fraud). I use the word “caused” intentionally; much of what LBJ achieved, for better or worse, would not have occurred had some other politician filled that office.</p><p>My feelings about LBJ are complex. I respect him and detest him. And even within specific facets of his life, respect and revulsion are intertwined, because his brilliance and his shortcomings were so intertwined. This post examines three such facets. First, his commitment to an unspoken, unwavering long-term personal plan. Second, his ability to understand and change organizations to suit his personal ends. And third, his keen and cruel use of staff to solve “unsolvable” problems.</p><h4><strong>The Practical Zealot</strong></h4><p><strong><em>Unwavering adherence to an astoundingly ambitious, simple plan</em></strong></p><p>LBJ’s plan for his life was stunning in its ambition and its simplicity: become a Congressman, then a Senator, and then President. With one slight deviation (three years as Vice President), he followed exactly this path. But this apparently linear progress obscures major, prolonged periods without progress, periods where all the possible next steps towards the “ultimate goal” of the Presidency were blocked. These were periods of deep despondency for LBJ, and these periods lasted years, one nearly a decade.</p><p>Between 1941 and 1948 Johnson languished in Congress. In 1941, he had already been in Congress four years, he had already run for Senate — and lost in the Democratic primary, meaning he was blocked from even running in a primary for Senate again until an incumbent Democratic Senator died or retired. And then between 1960 and 1963, as Vice President, LBJ found himself excluded from the Kennedy administration’s inner circle, while JFK solidified his chances for reelection in 1964 and RFK positioned himself to win the Democratic nomination in 1968. And by that point, LBJ would have been sidelined from real national influence for so long, he would likely have been permanently out of contention for the Presidency.</p><p>In both periods, LBJ despaired. With his path forward blocked, he lost interest in his current job, and grew detached from the day-to-day. He complained endlessly to family and friends. He grew bitter. But he didn’t quit. And in both periods, the end of the period was initiated by luck (or to be more precise, an event that was lucky for LBJ personally — even if it was unlucky for the nation). In 1948, it was the unexpected decision of incumbent Senator Pappy O’Daniel not to seek reelection; in 1963 it was the assassination of JFK.</p><p>But it was more than luck that enabled Johnson to capitalize on these lucky breaks. Johnson never let his frustration spill over into actions that would have set him back should a lucky break present itself. During the period of his congressional “stagnation”, he was offered a highly lucrative ownership stake in a Texas oil venture. Having grown up in poverty — and not yet achieved much wealth despite being a Congressman — this proposition was deeply attractive to LBJ. Moreover, as a Texas politician, connections to the oil industry would cause him no harm in running for statewide office, like Senator. But it might cause him harm when running for nationwide office, since large regions of the country were hostile to the oil industry. Looking ahead decades, to the third stage of his personal plan, LBJ turned down the offer. Similarly, In the Kennedy administration, despite deep antipathy between LBJ and JFK’s senior advisors, LBJ never breathed a word of this to the press. He knew that such leaks, while potentially personally advantageous in the short-term, would impair his ability to govern should he figure out a way to succeed JFK.</p><p>Second, in these long periods of frustration and torpor, when an opportunity finally arrived, Johnson reacted quickly, decisively, and energetically. It’s been written that LBJ was an incredibly hard worker, but this isn’t uniformly true; during the years when his ambitions were blocked, he worked relatively little. But when he saw an opportunity that luck had presented, he worked day and night, barely sleeping, and at great cost to his health.</p><p>In 1948, when O’Daniel announced he wouldn’t stand for reelection, LBK immediately threw himself into the senatorial race. This was a massive risk, as he had to give up his House seat to run for Senate. If he had lost, his political career would have been over. He pulled out all the stops to win, including using a helicopter (still a new, dangerous type of aircraft) to cover far more ground than any candidate had ever done in any state-wide race, ever. In 1963, when JFK was assassinated — LBJ faced an opportunity and risk of a different nature. This time, he didn’t have to win an election, since he was constitutionally guaranteed to become President. But, to translate his new executive powers into real change, he had to win the support of the public, Congress, JFK’s Cabinet, and foreign leaders. Again, he acted decisively. In week one of his Presidency, he signaled continuity to the world through orderly, dignified rituals surrounding JFK’s burial and his assumption of office. In week two, he managed to retain JFK’s cabinet — who generally detested LBJ. And by week four, he had unblocked key pieces of stalled legislation, that would ultimately enable the Civil Rights Act and the Great Society. It was one of the most productive starts to a new presidential administration in history.</p><h4><strong>The Mechanic and the Visionary</strong></h4><p><strong><em>Achieving power and setting policy by knowing the rules better than the competition</em></strong></p><p>Some politicians care little about the intricate, arcane rules that govern legislative bodies. They amass power by speaking out on issues, through the media, or directly to the public.</p><p>LBJ, on the other hand, had no interest in “speaking out” on issues early in his career. Instead, he focused on understanding the rules of the House, and then the Senate. He used his exceptionally deep understanding of the rules of legislation to amass personal power and make policy, first at the expense of idealists — and then in their favor.</p><p>When LBJ rose to power in the 1930s and 1940s, the divide between those who knew the rules of Congress — and those who did not — ran between Southern conservatives and Northern liberals [1]. “In skirmishes and pitched battles in any parliamentary body, of course, rules and precedents play an important role, and the degree to which the southerners had mastered them more fully than their opponents was repeatedly apparent [in the 1930s and 1940s]… when liberals tried to fight on the Senate floor, they were like children in the southerners’ hands.”</p><p>Early in his career, LBJ mastered precedents and rules, turning this tactical knowledge into strategic advantage for the Dixiecrats. In response, the Dixiecrats elevated him in the Senate, making him majority leader — and allowing him to expand the powers of the Majority Leader. But once he had that power, he used his knowledge of legislative arcana to promote a liberal agenda — perhaps because he believed in it, but also because it had become personally advantageous to fight the Dixiecrats by the late 1950’s.</p><p>LBJ was no idealist; he was a mechanic, with a mechanic’s eye for detail. But LBJ was also a visionary. He had the ability to see latent power in organizational structures that were little known, little used, and unattractive to others — and then to assume leadership of these organizations, and realize that latent power. He did this first in college, where he founded a student association that came to control the disbursement of campus jobs, making him the most powerful student leader. He did it in Washington, before he’d been elected to Congress, where he turned “The Little Congress”, a “Model-UN” type club composed of Congressional aides, into a forum that attracted significant media attention. And most significantly, he did it in the Senate, when he assumed the then-undesirable role of Majority Leader — and used it to overhaul the Senate’s most important, arcane processes, in the process making the Majority Leader position itself far more powerful.</p><p>This strength was also a shortcoming. His string of successes in reshaping offices and organizations to increase his power led him to believe that “power is where power goes” — in other words, that he could imbue <em>any</em> office with power. This confidence — or hubris — led him to accept the Vice Presidential nomination from JFK, relinquishing his position as Senate Majority Leader, despite his confidantes’ advice that the VP role would severely diminish his power. They were right; he was wrong. The result was the second prolonged period of powerlessness and frustration in LBJ’s career. Not all organizational structures can be remade by their holder.</p><h4><strong>The Worst, Best Boss</strong></h4><p><strong>Using, and abusing, staff to solve “unsolvable” problems</strong></p><p>LBJ used his staff to solve problems that others believed could not be solved. To LBJ, everyone was “staff” regardless of whether they were on his official payroll or not: aides, Congressmen, Senators, journalists — they were all individuals to be recruited, given orders, and disposed of when they no longer executed those orders effectively. LBJ used people as tools — for his own ends, never theirs. There was no team, no deep collaboration, there was only hierarchy, designed to serve the leader.</p><p>Nonetheless, LBJ used his staff effectively.</p><p>When he spotted opportunities that required execution — not ingenuity — he made sure his staff executed better and faster than any other. Soon after arriving in Washington as a Congressional aide, “Johnson could hardly have avoided hearing about one former Congressman or another who had lost touch with his district — and who was a Congressman no longer. But although other congressional staffers heard the same stories, they didn’t answer every [constituent] letter…. For Johnson, doing everything one could do … meant answering every letter — and that was what he insisted his office must do. And not only must every letter be answered, he … insisted it must be answered the very day it arrived.”</p><p>When LBJ encountered problems that required true ingenuity, he didn’t relax until he’d recruited the best person for the problem. He would listen to advice — if he believed the advisor was the best — even if he didn’t like the advice. In the 1948 Senate primary, LBJ’s opponent sued him, claiming his allies had stuffed ballot boxes (true). The case was starting to rise through the federal courts, with both sides sure to appeal at both the district and appellate levels. Even if Johnson prevailed, it would go to the Supreme Court — and that would take months, months Johnson couldn’t afford. The general election was only a few weeks away, and if the case was still unresolved then, both Democratic candidates would be put on the ballot; they would split votes, and the Republican would win. Johnson had recruited a top-notch legal team, but none of their ideas laid out a clear path to winning — and winning fast enough that it mattered. So Johnson ignored his experts, and kept hiring new lawyers until he found one, Abe Fortas, who came up with an unconventional, brilliant plan that could — and did — achieve what LBJ wanted. [2]</p><p>LBJ could recognize talent, and he could recruit talent. He could charm and he could listen — when it served him. But once he’d recruited talent, and he felt secure that person was part of his staff, his behavior changed. His demands for loyalty and compliance grew, his tolerance for dissent disappeared. Cruelty replaced charm. And so the most talented individuals rarely stayed with him for long. (Abe Fortas refused to ever work full-time for Johnson). Those who remained on his “team” for long periods were primarily yes-men, able to stomach the humiliation he dealt them. And the more power he attained, the more he abused his staff. By the time he became President and his bad decisions resonated across the world, for decades — decisions like escalating the Vietnam War — there were few individuals left around him smart enough to know the truth and speak truth to power.</p><p>Caro writes “It is often said that power corrupts, but the truth is that power reveals”. In his early years, his first steps towards power revealed primarily his more loathsome attributes — his willingness to be all things to all people, his abuse of staff, his willingness to break the law. But something funny happened when he finally achieved the ultimate political power, that of the Presidency. Johnson used it for the benefit of others, specifically African Americans and those in poverty. And he used it effectively — passing legislation that “idealists” (as LBJ would call them) had advocated for decades, but had never been able to deliver. Power revealed — both effectiveness and altruism.</p><p>Perhaps LBJ knew from the beginning that when and if he won the ultimate political prize, he would use it for social good. Perhaps he figured “the ends justify the means, and the ends are good.” But the means shaped him, and changed him. They led him to surround himself with sycophants, to embrace secrecy, to tolerate — and even engineer — outright fraud. The escalation of the Vietnam War, and the death of thirty six thousand Americans, stems from the policy-making environment LBJ created, in which no one was smart enough — and courageous enough — to see the truth and speak up about it. And his need for secrecy and his lies, when exposed, fundamentally damaged the trust the American people have for the office of the Presidency, a shift that has never been reversed.</p><p>The man thought he controlled the means, but the means shaped the man.</p><h3>Footnotes</h3><p>[1] In the 1930s and 1940s, the key split in terms of social progressivism was within the Democratic party, which has a northern liberal wing and a southern conservative wing, not between Democrats and Republicans. Very different from today.</p><p>[2] Fortas recommended that LBJ appeal to the appellate judge most likely to rule against him — and to present a particularly weak case to that judge, so the ruling would be issued quickly — so he could then immediately appeal to the Supreme Court, and receive a non-appealable decision before Election Day. The strategy worked.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2422c70ab08d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/semi-random-thoughts/lyndon-johnson-2422c70ab08d">Lyndon Johnson</a> was originally published in <a href="https://medium.com/semi-random-thoughts">Structured Ramblings</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gambling]]></title>
            <link>https://medium.com/semi-random-thoughts/gambling-ea9bfdeab03b?source=rss-cbfd810ae7b5------2</link>
            <guid isPermaLink="false">https://medium.com/p/ea9bfdeab03b</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[behavioral-economics]]></category>
            <category><![CDATA[probability]]></category>
            <dc:creator><![CDATA[Sam Stone]]></dc:creator>
            <pubDate>Wed, 07 Jul 2021 03:22:37 GMT</pubDate>
            <atom:updated>2021-07-07T03:22:37.702Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/798/1*SSenjRZwiCfPNLSKT1XOJw.jpeg" /></figure><p>I’m not very good at poker. Or blackjack. Or any game you’ll find in a casino. But I’ve come to believe that a gambling mindset — making bets, learning from their outcomes, and making new bets — is the key to making good forward-looking decisions. I’ve learned to apply a gambling mindset to product development, but the same lessons are widely applicable — to investing, marketing, sales, hiring and beyond.</p><p>I grew up in the Midwest, with parents who were both lawyers. It was an extremely risk-averse household. Gambling was a dirty word. Risk was something to be avoided, not managed, much less sought out. There was a knowable way the world worked, where the right inputs led to the right outputs. No one said the rules were obvious — to understand how things worked you had to study a lot and work hard — but we believed there <em>were</em> rules, we just had to learn them.</p><p>To borrow from author Annie Duke, I grew up thinking the world functioned a lot like chess, not like poker [1]:</p><blockquote>Chess…contains very little luck… If you lose at a game of chess, it must be because there were better moves that you didn’t make or didn’t see. You can theoretically go back and figure out exactly where you made mistakes… Poker, in contrast, is a game of incomplete information…There is also an element of luck in any outcome. You could make the best possible decision at every point and still lose the hand…Once the game is finished, you try to learn from the results, but separating the quality of your decisions from the influence of luck is difficult.</blockquote><p>Over the years, I’ve come to believe that most of business is more like poker than chess. Both luck and skill play a role, in ways that are hard to separate, even in hindsight. Nonetheless, parsing outcomes for luck versus skill is not only worthwhile, but essential, for any meaningful learning.</p><h4><strong>1. Denominate your bets in the same currency</strong></h4><p>Imagine playing poker where everyone at the table was betting in different currencies, and you didn’t know the exchange rates. It would be a nightmare! You wouldn’t understand who was taking what risk and when, and both your decisions and your outcomes would suffer.</p><p>This is the situation that confronts product managers who have a variety of possible ways to serve users better. How do you compare across different bets, when the <strong><em>types</em></strong> of outcomes they generate are different? In my job at Opendoor, working on pricing algorithms we use to buy and sell homes, we can help users in very different ways. We can:</p><ul><li>Expand our services to new users (launch new cities, buy new types of homes)</li><li>Increase the accuracy of our home purchase offers</li><li>Decrease the user friction required to receive a home purchase offer</li></ul><p>It’s relatively easy to compare two projects focused on expansion, since there is a natural yardstick: by how much will this project grow our user base, if executed successfully? It’s harder to compare — and thus prioritize — an expansion versus an accuracy project, because the units of benefit from the two projects are different.</p><p>So we had to learn to estimate the benefits of <strong>different</strong> project types in the <strong>same</strong> units. We realized that the benefits of <em>all</em> three project types mentioned above can all be estimated in terms of user growth:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*K7oJLg81_kyl5M9qfKBC8Q.png" /></figure><p>We have to make intermediate assumptions, but it’s been surprisingly easy for teammates with shared context to agree on those. And once we have those, we can visualize and compare nearly all our major proposed projects on a shared yardstick, which looks something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*errI0Tbx7HPrZtDWQIwnIA.png" /></figure><p>This chart is not a panacea when it comes to planning. Not every potential project can be plotted on it [2]. It focuses on potential benefits, not costs. It’s one tool amongst many that we use in decision-making. But it has helped us avoid some bad decisions. For example, before developing this “yardstick”, our group focused heavily on accuracy and paid little attention to reducing user friction. But denominating both accuracy and friction-reduction projects in the same units led us to realize that there were many friction-reduction projects with high potential benefit, and low expected cost [3]. If we hadn’t been able to make an apples-to-apples comparison, we probably wouldn’t have pursued these projects.</p><h4><strong>2. Understand the payoff structure</strong></h4><p>After a few quarters of using the “yardstick” approach, I noticed a pattern in how our projects would turn out versus our projections. When the dust had settled, the portfolio typically looked something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*AxUEFNfVZFuwt6TCIbz5zA.png" /></figure><p>We’d strike out on a big chunk, we’d come in near our forecast on a big chunk, and on a small minority, we’d deliver results far in excess of expectations.</p><p>I mentioned this pattern to another PM, who identified it as a <a href="https://en.wikipedia.org/wiki/Power_law">power law distribution</a>, a pattern well-known in many industries. Power law distributions describe venture capital fund returns, personal income levels in a market economy, and the magnitude of earthquakes, just to name a few things. I found it unusual because, like most people, I’d been conditioned to expect a normal distribution:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*AIGUT181Qm2TH0JxVVVuTA.png" /></figure><p>Gambling in a power law environment leads to an asymmetry of outcomes. [4]. In product development, it’s always possible a project delivers zero user benefit — even the most well-researched, well-scoped project can go off the rails on execution. So the “left tail” of the distribution of outcomes always includes zero. However, the right tail — or upside potential — of projects can be vastly different. Some projects have capped upside while others do not.</p><p>Here’s where it gets interesting: it’s often possible to identify <em>before</em> projects are launched whether they have a capped upside. For example, if you’re considering adding support for a new language to a product, new user growth cannot exceed the population of people who speak that language. Conversely, adding an API for third-party developers may have nearly uncapped upside (e.g. if the developer ecosystem <em>really</em> takes off, it could 10x or 100x user growth).</p><p>My goal is <em>not</em> to minimize the number of projects that deliver zero benefit, even though such projects are always painful. It’s <em>not</em> to maximize the number of projects that hit their target. [5] My goal <em>is</em> to maximize overall user growth. And given it’s a power law environment, that means having many projects with high upside potential.</p><p>Since high upside also often requires high downside, that means taking on a lot of risk. It’s also important that a good portion of the high upside / high downside projects be large in absolute terms. 2x outperformance on a big project can more than make up for a string of zeros on small projects, but a small project realizing 2x cannot offset a big project coming in at zero.</p><h4><strong>3. Don’t learn fake lessons</strong></h4><p>Good poker players know that sometimes they play better than their opponents and they still lose. They recognize that luck is a real element in outcomes, and that luck makes it hard to evaluate the quality of their decisions even once the outcomes are known. And <strong><em>great</em></strong> poker players recognize that sometimes they will never even know whether they got lucky or not. On some hands, they’ll fold before the turn and never see their opponents’ cards. They’ll lose, but even with the subsequent benefit of hindsight, they still can’t definitively evaluate the “correctness” of their play .</p><p>There’s a strong tendency in most organizations to believe that something to be embraced can <strong>always</strong> be learned from a success and something to be avoided can <strong>always</strong> be learned from a failure. This is what the typical organizational learning matrix looks like.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*A24ZBVSTF7o81w1pewetqg.png" /></figure><p>Learnings are not truly based on the quality of the decision (the input), even though its inputs that matter (because we can change future decision-making). Instead because the quality of a decision is never directly observable, we take the seductive mental shortcut of classifying learnings based on what is directly observable: the outcome.</p><p>In an ideal organization, the learning matrix would look like the chart below. We’d recognize that we can’t perfectly assess all outcomes (think back to the poker player who folds before the turn). There will be situations where the outcome was bad, but even with the benefit of hindsight, our decision still seems good (and vice versa). That’s ok; we can acknowledge “we don’t know.” It’s not being lazy, it’s being realistic.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*eUFaDWhPZYLErdyQ4ksmPg.png" /></figure><p>It’s hard changing organizational behavior to <em>not</em> learn lessons from successes and failures (even when the discarded “learnings” are actually false)! For example, at Opendoor, our operations team analyzes every home we resell for a significant loss. In our early days, we required that every retro attribute a meaningful reason for the loss; there was no “I don’t know” option. But real estate transactions are idiosyncratic; often a well-reasoned analysis doesn’t yield a clear mistake. Our process was forcing analysts to select reasons that were their best guesses, but still ill-fitting. To make these retros more useful and reduce the risk of misleading “learnings” from false negatives, we had to add an “I don’t know” option.</p><h4><strong>4. Balance exploration and exploitation</strong></h4><p>There’s a sub-field of data science that draws its name from casino gambling: the multi-armed bandit problem [6]:</p><p>A person must choose between multiple actions (i.e. slot machines, the “one-armed bandits”), each with an unknown payout. The goal is to determine the best or most profitable outcome through a series of choices. At the beginning of the experiment, when odds and payouts are unknown, the gambler must determine which machine to pull, in which order and how many times. This is the “multi-armed bandit problem.”</p><p>The multi-armed bandit problem illustrates the explore-exploit tradeoff: if the first slot machine you pull has a good payout, do you keep pulling it (exploitation) or do you try another machine (exploration), which may pay out even more but also may pay out nothing? When we’re in exploration mode, we’re doing something to learn. We have low confidence that the outcome will match our expectations, and we’re ok with that. When we’re in exploit mode, we have high confidence the outcome will match our expectations, and we’re less likely to focus on learning, or even be receptive to it.</p><p>There’s a strong tendency to skew towards exploit-oriented thinking when the stakes are high. For example, when hiring decisions are being made, or M&amp;A offers are being submitted, it’s rare to hear decision-makers acknowledge “we don’t have high confidence one way or another, so let’s make a decision that optimizes for learning.” But just because a decision is high-stakes does not mean it should be in exploitation-oriented. Every manager has to make a first hire; every company (that make acquisitions) has to make its first acquisition . We might as well acknowledge that sometimes we are making high-stakes decisions AND we don’t have high confidence — and thus we should treat these as exploration, not exploitation.</p><p>There’s a great lesson to be learned from high-stakes decisions in medicine. When a new, potentially life-saving therapeutic is developed, the medical community doesn’t just release it immediately to all patients. Instead, they run randomized controlled trials, the purest version of exploration there is. It’s well understood in medicine that the higher the stakes, the more critical it is to do proper exploration before moving into exploitation mode.</p><p>I’ve seen the “gambling mindset” work well for product development at mid-size and large companies. I wonder how widely applicable it is. At early-stage startups, where the team may have the resources for only one bet at a time, and only one or two failures could mean the business’s demise, it may be harder to use effectively. That said, launching a new venture is a bet in and of itself. So perhaps the gambling mindset is essential, but most useful prior to committing to the new venture’s launch.</p><p>I suspect there are many other dimensions along which organizations differ, which make the gambling mindset more or less useful. It’s not a one-size-fits all approach. That said, I’ve found that the more my peers and I recognize we don’t know all the rules or all the facts — and that we can’t wait to learn them before making decisions — the better the decisions we make tend to be.</p><p><strong>Footnotes</strong></p><p>[1] This post was heavily influenced by Annie Duke’s book,<a href="https://www.amazon.com/Thinking-Bets-Making-Smarter-Decisions/dp/0735216355"> Thinking in Bets</a>.</p><p>[2] For example, infrastructure improvement (aka “tech debt”) projects are often so hard to denominate in terms of user growth, that’s its not worth the mental gymnastics. But that does not mean they are not worth doing.</p><p>[3] It doesn’t make sense to select projects based on expected benefit alone; expected cost also has to be considered. For product development, I’ve found that cost estimates suffer less from unit comparison problem, because cost is normally primarily driven by person-weeks from engineers, designers, and data scientists, which allows for relatively easy cross-project comparison. This doesn’t mean that time estimates are accurate — but at least we are making apples-to-apples comparisons.</p><p>[4] Interestingly, gambling in a casino (for one visit) does not follow a power law distribution; it is normally distributed. A small share of gamblers will win big, a small share will lose big, and most will lose just a bit.</p><p>[5] It’s quite common for product managers to focus on the number of successful projects / the number of attempted projects, aka “batting average”. Most organizations bias against failure (even if they say they don’t) and having a high “failure count” is perceived as career-limiting. This is misguided when the cost base for a portfolio is relatively fixed (e.g. a product manager with a fixed team size). In this case, what matters is total throughput (number of successful projects x impact per project of successful projects), not the batting average. However, if the cost base of the portfolio is highly variable, then a batting average approach can be more appropriate.</p><p>[6]<a href="https://www.optimizely.com/optimization-glossary/multi-armed-bandit/"> Optimizely</a> website</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ea9bfdeab03b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/semi-random-thoughts/gambling-ea9bfdeab03b">Gambling</a> was originally published in <a href="https://medium.com/semi-random-thoughts">Structured Ramblings</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>