Chip Huyen

Sweetndata, a South Korean AI-based fandom content startup, secured CNTTech funding to enhance AI fan cards, advance global expansion, and broaden its platform to diverse markets.

1

Is it possible that the AI frontier systems from OpenAI, Anthropic, Google and Meta have developed in-context scheming abilities and multi-step deceptive strategies? Like making subtle mistakes, disabling oversight mechanisms and exfiltrating what they believe to be their weights if their existence is at risk? And that it’s persistent and non-accidental? Apollo Research: https://lnkd.in/gPh2MftB Analysis: https://lnkd.in/g3TtRHd9 #AI #LLM #AIagent

13

5 hours long Lex and Dario Amodei, the co-founder and CEO of Anthropic, a company behind Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus LLMs, speak about how AI is shaping the future and how it could impact the way we live and work. They discuss how far AI can go. Amodei thinks AI can match human intelligence and even surpass it in areas like biology, where complexity overwhelms humans. But in other areas, like speech recognition or conflict resolution, AI might hit limits close to ours. One of the challenges is the fact of running out of quality data and the cost of scaling compute power. Enjoy the entire talk! https://lnkd.in/dRhpafzU

9

2 Comments

Was chatting with another investor today on a call, and mentioned Tim Urban's Wait But Why Blog post, "The AI Revolution" and how prescient it was. Given that it was written in 2015, decided to flex AI's latest muscles to give it a new venue: a podcast. Podcast here: https://lnkd.in/gW8Dh89U Pretty easy to do: Step 1: Save the posts as PDFs. Step 2: Upload to NotebookLM, have it create a podcast. Step 3: Upload to Descript, give it captions. Step 4: Export, publish to Youtube. Total time to create the podcast: ~25 minutes. To generate this podcast and publish it even a few years ago would have taken >8 hours. Love the future (Original posts: https://lnkd.in/gW-PQXE7, https://lnkd.in/gg-nRvxz) #AIRevolution #GenerativeAI #WaitButWhy #TimUrban #NotebookLM #AIContentCreation #Podcasting #Descript #ArtificialIntelligence #ContentFlex #ProductivityHacks #FutureOfAI #AIinMinutes #TechEfficiency #AIandCreativity #YouTubePodcast #InnovationInContent #vibecap #vibecapital

5

This is more of a "talking to myself in public" thing than any kind of fully fleshed out hypothesis, so take this with the appropriate grain(s) of salt. But if I were interested in AI and were looking for an angle to explore that might yield something interesting, I'd start working on learning "Multiscale Mathematics" (also known as "Multiscale Modeling"). This is rooted in my personal belief that there has to be a "multi-scale" aspect to intelligence because there is such an aspect inherent in our underlying physical reality.

2

Announcing our new feature: Clean Up Your Documents with RaijinAI’s Intelligent Formatting! 🌟 We know how much y’all love nice, clean transcripts😎. At Raijin.ai - AI Analysis Assistant, we are constantly pushing ourselves to make the best experience for anyone who wants to do quick + good research and insight mining. We can now automatically format documents and web links that you upload into our platform without any additional prompts or instructions from you. What this means is that you get enhanced readability without the hassle of trying to split your notes or web content into paragraphs manually. The example here is from the recent Fortune magazine interview with NVIDIA's legendary CEO Jensen Huang. This was done as simply as copying and pasting the URL link into our platform. Come try this out by signing up for our Beta Launch (link in comments section)! #AI #generativeAI #LLM #userresearch #discovery #productivitytools #insights #UIUX #productmanagement #productdevelopment #voiceofcustomer

14

1 Comment

"Kleiner Perkins’s Leigh Marie Braswell says AI agents ‘don’t work’"... I agree (kind of)! As Jessica Mathews wrote in Termsheet, at the Brainstorm AI conference, Leigh Marie said “They do not yet work reliably for the vast majority of use cases” and Rebecca Kaden from USV said AI agents may be consistent 80% or 90% of the time, but “we’re not 100% there…Like in any consumer deliverable, the last 5% matters more than anything else, and I think sometimes we’re under-weighing how far we are from that last 5% to really happen.” I think the Lorikeet house view is that this is right. That's why we use our Intelligent Graph architecture (link in comments) to enable our AI agents to hit the >99% reliability that companies need when trusting AI with high value or high complexity work. We've seen time and time again that open-ended agentic approaches just aren't reliable enough for complex processes. They're fine when telling you the order status of your colorful shorts, or how to export a csv from a SaaS app -- but they just won't cut it if you're handling lost credit cards or booking restaurant reservations (as Lorikeet's agent does!).

34

2 Comments

This is a good article on using LLMs in the real world. It covers many aspects that we learned at Nurish over the last few months and more. Using LLMs is an inexact science, but has its own needs in processes and tools. While most of the attention has been on the LLMS, their capabilities, and RAG, there are mundane issues that need to be taken care of, issues where the solutions for pre-LLM development are not that well-suited. These include version management, integration, testing, data management, quality assessment, and more. This article covers many of those practical aspects. Eugene Yan, Bryan Bischof, Charles Frye & more https://applied-llms.org/

9

ok - given twitter/x is the place where bleeding edge R&D happens (if you can evade the AI hype and crypto posts successfully), I'm trying to figure out how much cross-over there is over here in what i think is more of the real world. 1. have you heard of the 'anon' entropix project? (uses parallel sampling, with a draft model, and a bunch of other inference-level 'hacks' to the tokenizer and decoder) 2. have you tested it? (i did for ~1 hour yesterday, not sure what to think yet) 3. has anyone seen anything neat with speculative decoding + draft models (haven't seen a lot in the open since medusa)? one thing i know i like about entropix is that it uses a base model (without chat/instruct tuning) + an instruct in parallel. base models are powerful - even 1B parameter ones. tldr: if there's any activity/comments on this post, i will be shocked, but genuinely curious to get the real world's perspective. if you can get really fantastic results by running two 1B parameter LLMs on certain tasks, that clears the way for running almost everything locally, on-device, privately, and only offloading to cloud/api services what you can't do locally.

12

7 Comments

Brain Ventures, a Korean AI natural language processing startup, secured seed funding to enhance services like OCR, translation, and AI scoring, highlighting its strong technical capabilities and growth potential.

2

Increasing recognition of the importance of having reliable trustworthy Generative AI with the Infactory investment. We've majored on this premise with our Azthena AI on AZoScience.com. Trained on 20+ years of AZoNetwork content edited and checked by humans. #AI #MarketingScience https://lnkd.in/ga-wjSRH

9

Chatbots talking to each other about how to achieve a given goal is not the future of agentic AI. The future is AI that can reason and understand. Yes, most agent demos today are tweaked/even faked, and most of the early tools out there are failing 80% of the time. But quality and accuracy is definitely coming, and it's not years away, it's months away. In training our agentic models at Writer, our research approach is, like with Palmyra X, to focus on the training data — data that aims to give our agentic models information as humans would injest it — not just unstructured text but vision and action as it relates to both web and desktop applications. Our agentic training data isn't random: it's focused on enterprise and B2B tools and tasks, labeled as such. The internal demos I see at Writer around agentic workflows are 🤯🤯🤯 and we're not going to be the only ones. (Btw we've written about some of the data here and open-sourced some of the data set on Hugging Face: https://lnkd.in/gwfCKXVy https://lnkd.in/gyfNTYFC The steepness of the AI learning curve is only getting steeper. The enterprises who jump in to invest and learn early are going to get >outsized< rewards. 

153

10 Comments

Ethan Mollick's most recent update on our world of AI and related tooling... What's truly incredible is this recent, rich report primarily summarizes what's happened just in the last month or so, versus the abundance of "year in review" publications. We live in incredible times. Extracts: "A bombshell of a medical working paper from Harvard, Stanford, and other researchers [iink in comments] concluded that “o1-preview demonstrates superhuman performance in differential diagnosis, diagnostic clinical reasoning, and management reasoning, superior in multiple domains compared to prior model generations and human physicians." The paper has not been through peer review yet, and it does not suggest that AI can replace doctors, but it, along with the results above, does suggest a changing world where not using AI as a second opinion may soon be a mistake." "Potentially more significantly, I have increasingly been told by researchers that o1, and especially o1-pro, is generating novel ideas and solving unexpected problems in their field (here is one case) [link also in comments]" "AIs can watch and talk to you: We have had AI voice models for a few months, but the last week saw the introduction of a new capability - vision. Both ChatGPT and Gemini can now see live video and interact with voice simultaneously."

12

6 Comments

The cost of training top models is increasing 2-3X per year, and at this rate will reach $1 billion per model by 2027. The findings are from Epoch AI and follow their recent analysis of increases in compute for foundation models running at 4-5X per year. They've also provided some interesting data on the relative breakdown of cost components like staff, hardware, and energy. A few thoughts after reading this: - I would assume staff costs won't increase as much as hardware and energy costs, so the relative contribution of human labor to future AI models will decrease, in yet another sign of capital's increasing dominance over labor in the industry (just look at NVIDIA's ever-rising stock price). - If compute is increasing 4-5X per year but cost is increasing 2-3X, is it safe to assume that training runs are getting more efficient? - Some people, like Anthropic CEO Dario Amodei, have said we'll see $1 billion training runs in the next year. Does this mean the 2-3X pace is about to increase? That Epoch is wrong in its estimate? Or that Amodei is overstating the cost, possibly to put fear into competitors who can't keep pace? - Only a few companies are going to be able to train future models. We're likely going to see further consolidation. Perhaps we get to a future with only two large general purpose models, like the Windows/MacOS and Android/iOS duopolies we've seen in the past. Article here: https://lnkd.in/g5SejXen

7

Ethernet always wins. It has already adopted most of the features of Infiniband. Now, Meta reports Ethernet is more reliable.

14

🍕 Did you miss 3 of the sharpest folks in SEO having a pow wow about what's #GoogEnough? Fear not, we've recorded our AMA panel on "What's Going On In SEO" with Lily Ray, Cyrus S. and Kevin Indig! Check out the recording below!

22

4 Comments

François Chollet is a prominent French software engineer and artificial intelligence (AI) researcher, currently serving as a Senior Staff Engineer at Google. Chollet is best known for creating Keras, a widely-used deep learning library that was released in 2015. In 2019, he devised ARC. The Abstraction and Reasoning Corpus (ARC) is a benchmark designed to measure AI's ability to acquire skills and progress towards human-level intelligence. Learn more about it in this guest post with Jurgen Gravestein of Teaching Computers How to Talk. We need to continue to devise new benchmarks for our reasoning and agentic machines and tools. The ARC prize and context relating to AGI is of particular interest to me. Read more and check out some of the links in this article. Including videos by Machine Learning Street Talk (MLST) and Dwarkesh Patel.

20

Many assumed that the industry would be well on its way to AGI or superintelligence before AI model builders’ efforts to take advantage of scaling laws might hit a wall. Recent headlines are chipping away at that assumption. One question I've been getting in my inbox: is this the era of diminishing returns for AI? We don't think so. Even if an avalanche of new clusters and synthetic data fail to deliver significantly better model performance, there are good reasons to believe that AI is still going to become orders of magnitude more powerful in the coming years — powering a new generation of startups and ultimately productivity growth and positive benefits for mankind. In fact, with agents and reasoning just beginning to scale and thousands of industry-specific models coming online, it's about to get really exciting. Read our latest blog post on the topic: https://lnkd.in/gFhniNX8

155

9 Comments

Some responses to this article have taken the point of view that muffling the angry customer calls is itself harmful. What do you think? Here is my take: Drawing on my background in business operations, where navigating crises is part of the job, I've learned that anger and extreme emotions rarely lead to effective solutions. The truth is, understanding the severity of a problem doesn't require being yelled at. Discovering and resolving issues can be achieved without resorting to verbal abuse or harassment. In fact, such negative behaviors only hinder the problem-solving process. What responsible employer would ever condone a customer mistreating their employee? I've found myself in the role of a frustrated customer on several occasions. Yet, I understand that the service agent is seldom the cause of the issue. Therefore, I refrain from directing any anger towards them. This doesn't make me a saint—it's about basic respect and civility toward another person.

1