Federico Ulfo

💭 Meet Raza Habib : The London-based entrepreneur who’s raised over $7 million to build LLM apps with his AI startup: Humanloop 🧑‍💻 HumanLoop helps companies that want to build products with Large Language Models (LLMs) to develop and evaluate these products. 🤏 Unlike software engineering, when you’re dealing with deterministic code and you can predict how it will behave; using AI is non-deterministic. It’ll produce different outputs depending on very minor changes. Product teams use HumanLoop as their development environment where in that UI, they can collaborate with engineers to develop prompts and iterate. 🦜 Here’s a quick example of how DuoLingo uses HumanLoop to build their LLM applications: 👷 A lot of prompt engineering is required when building AI apps like AI assistants or chatbots that users can practice languages with. 🗣️ The best people to do the prompt engineering for applications like this are the domain experts: Linguists themselves. 🚀 HumanLoop allows linguists to collaborate with engineers to build prompts and measure their performance. 🗞️ In 2023, HumanLoop was also named one of the 21 generative AI startups to watch, according to Sifted. 🛃 Today, HumanLoop works with customers like DuoLingo, Gusto, Vanta, Apollo.io and Fundrise to build LLM applications for them. 📈 Since 2022, they’ve grown over 60x in usage and have around 14 employees. 💰 They’ve now raised about $7M in total from investors like Index Ventures, Y Combinator and the UCL Technology Fund. Check out my latest post for the full story here!! 👉 https://lnkd.in/dzNs9GEu

32

2 Comments

$2.8B of disclosed funding for 104 AI Agent startups in 2024 (Through to Dec 10, 2024) … and that excludes the big LLM investments like OpenAI ($6.6B), Anthropic ($4B), Cohere ($0.5B), etc… and also companies that aren’t primarily AI Agents (eg. Intercom, Replit). 🤯 Will it be higher or lower in 2025? For reference total VC funding in 2024 was ~$100B Credit to CB Insights for the analysis / graphic #AIAgents #VC

15

2 Comments

Model Context Protocol or short MCP is likely the new hottest thing in LLMs almost nobody talks about. The Problem: Usually LLMs / AI only is as smart as its training data and can't interact with the world outside its confined borders. Think about ChatGPT, you can ask it to generate an image or to give you feedback on text, but it can't buy you something on Amazon. How's Model Context Protocol solving this? MCP, developed by Anthropic, allows to write and enable powerful extensions for the LLM. Basically whatever is accessible through your computer can be made accessible to the LLM. Examples: - Your local Filesystem for read / write permission - Google Maps - Slack - Spotify - Github - Other LLMs so they can talk and discuss about a solution Although maybe a bit more technical than just interacting with a nice polished User Interface by the big companies, I believe MCP is the most significant new development when it comes to AI tools. Has anyone tested it yet? GitHub: https://lnkd.in/eyA-XTgp

1

Groq Partners with Aramco to Propel Saudi AI Ambitions In a groundbreaking move, Groq, the innovative chip startup renowned for its cutting-edge technology, has inked a significant deal with aramco, the world’s leading oil company. This strategic partnership is set to be a game-changer for Saudi Arabia's burgeoning AI landscape. What This Means: Groq’s advanced chip technology will play a crucial role in supporting Saudi Arabia's ambitious plans to position itself as a global leader in artificial intelligence. The collaboration is expected to significantly enhance the computational power available for AI applications, paving the way for breakthroughs in various sectors including energy, healthcare, and smart cities. Key Highlights: Boosting AI Capabilities: Groq’s state-of-the-art chips are designed to handle complex AI tasks with unprecedented speed and efficiency, aligning perfectly with Saudi Arabia's AI objectives. Strategic Alignment: This partnership reflects a broader trend of technological alliances aimed at fostering innovation and economic diversification in the region. Future Impact: By leveraging Groq’s technology, aramco and Saudi Arabia are positioning themselves at the forefront of the AI revolution, potentially setting new standards in the industry. As the tech world watches closely, this deal underscores the growing intersection of AI and semiconductor technology, and how strategic collaborations are shaping the future of these industries. For more information: Bloomberg https://lnkd.in/dr89qeCV #AI #ChipTechnology #Innovation #TechPartnerships #Groq #Aramco #SaudiArabia #FutureOfAI #TechNews

6

Today we’re releasing Chai-1r, an updated version of Chai-1 with full support for restraints. Prompting Chai-1r with prior structural knowledge can boost accuracy by >2x. https://lnkd.in/euP7frVQ Custom restraints can refine predictions with knowledge obtained from scientific literature or lab assays. For example, a “contact” restraint specifies that two tokens are within a certain distance. A “pocket restraint” indicates a token should be in contact with another chain. Chai-1r excels in antibody-antigen interaction prediction with mean DockQ % of - Chai-1r (+4 epitope residues): 43.7 - Chai-1: 35.6 - Boltz (an AlphaFold3 reproduction from MIT & Genesis): 26.4 - AlphaFold2: 20.6 Restraints are available immediately both in our public codebase (https://lnkd.in/e3SGx_dj) and on our web server (https://lnkd.in/e9YfR6GV).

575

13 Comments

What is the state of the AI for Good ecosystem? This was a question proposed by a panel at UNGA with Payal Dalal, Dana Ima Hamzah,Uyi (O.T.) Stewart , Fred Swaniker, Komal Sahu, and I have been thinking a lot about it ever since. ⬆️ Hope: The potential of AI to drive meaningful change is immense. We’ve seen real, tangible solutions emerge—AI systems that support climate resilience, improve healthcare outcomes, and streamline social services. These aren’t just concepts; they’re proven examples of what’s possible when AI is applied thoughtfully. ➡️ Challenge: Yet, in many critical areas, the technology still falls short. While we’ve made progress, the reality is that most AI tools are not yet robust enough to deliver high-quality results in complex, domain-specific applications. Isolated success stories are promising, but they’re not enough. The challenge now is taking these breakthroughs and replicating them across the broader ecosystem to ensure that AI’s impact is felt where it’s needed most. ⬆️ Hope: The cost of building AI solutions has dropped dramatically—by around 1,000x in the last decade. And it will continue to drop. ➡️ Challenge: Access to high-quality, well-documented datasets. Currently, 85% of the organizations that can afford all data they need are defense & intelligence agencies. We need to invest in data interoperability and make these datasets available to those working on the frontlines of social impact. One good example is ODE Partners, which is working on making Earth observation models more accessible. AI should not be a hammer in search of nails. Too many organizations treat it that way, jumping on the bandwagon without a clear understanding of what problem they’re trying to solve. The focus should always start with the problem, and only then should we decide if AI is the right tool. 🔜 Winning strategy? Ecosystems! Success in AI for Good will not happen in a vacuum. We need to build ecosystems around every major impact vertical. Every accelerator or AI for Good program should consider having an “Ecosystem Coordinator-in-Chief.” We need people dedicated to connecting the dots, both within our organizations and across the sector, if we want to turn isolated wins into systemic change. If we want to move from theory to impact, we need the right tools, the right data, and most importantly—the right strategy. #AIEcosystem #SustainableDesign #AISolutions

18

4 Comments

My wide ranging conversation on enterprise AI automation with Alp Arhan U. on the Masters of Automation is now available on YouTube, Spotify and Apple Podcasts (links in comments). Some of the things we go over are: 🧠 My journey from Turkey to Harvard to YC 🧠 How Brainbase (YC W24) decides what to automate with AI 🧠 Our unique insight at Brainbase 🧠 Why immigrant founders make some of the best entrepreneurs Enjoy our conversation, and please sound out with comments and questions below!

90

6 Comments

Every time we sit down with enterprise CIOs and AI teams, the top of mind concern we hear is "Reliability" - how do we go from our use case to reliable AI as fast as we can? In our experience, reliable AI is extremely hard to engineer. There is no silver bullet - it is about the many long tail problems that you need to grind at for weeks and months. But, this is what makes enterprise AI extremely resource intensive today. We are building Awen to change this. We are building the core agentic infrastructure that helps AI teams ship reliable AI 100x faster - from data curation to running evals at scale. Experience more in our playground: https://lnkd.in/gFGnpcdR Abhishek Borar Himanshu Muchhal

9

Llama3 + Groq can generate tokens at a 208.9 per second, which is nearly 10 times faster than GPT-4 Turbo and Claude 3 Opus.

3

The spot of best available LLM has been kinda volatile. In February Claude 3 Opus dethroned GPT 4 for like 3 weeks till openai released another version of GPT 4 to reclaim the top spot. Openai further cememted the win with the release of GPT4o which was not that much better than GPT4 but much more capable. I believe its gonna play out the same way. The highlight for this release is the price point. Gpt4o and Claude 3.5 sonnet are both very good at agentic workflows Claude 3.5 sonnet is priced so well as to become the first choice for this. it is a whooping $2 per million tokens cheaper than GPT4o which is so significant by the scale at which these models are used.

2

🧠 AI Showdown: Grok 2.0 vs. Claude 3.5 Sonnet vs. ChatGPT 4 I recently put Grok 2.0, Anthropic's Claude 3.5 Sonnet, and OpenAI's ChatGPT 4 through their paces with a simple yet revealing engineering challenge. The task? Implement an efficient JavaScript function to find the longest palindromic substring in a string up to 1000 characters. (Original Prompt: https://lnkd.in/gnrT9NwE) The results were eye-opening: - Grok 2.0 (xAI): • Implemented Manacher's algorithm • Achieved O(n) time complexity • Used boundary markers for uniform palindrome handling • Demonstrated higher-level thinking in algorithm selection - Claude 3.5 Sonnet & ChatGPT 4: • Both used the "Expand Around Center" method • O(n^2) time complexity • Potentially less efficient for longer strings Why Grok's approach stands out: - Unified palindrome detection (treats all as odd-length) - Reduced redundancy in checks - Superior scalability for larger inputs Interestingly, when asked to evaluate the solutions: - Claude 3.5 and Grok 2 both chose Grok's solution - ChatGPT picked its own (citing simplicity, despite efficiency not being the primary goal) Key Takeaway: In less than 24 hours since its release, Grok 2.0 has shown it's not just about solving problems, but solving them smartly. Its approach to palindrome detection is a masterclass in efficiency, showcasing the rapid progress in AI development. Kudos to the teams at X on a job well done! BTW, for anyone curious: Claudes answer: https://lnkd.in/gbwSkzRc ChatGPT's Answer: https://lnkd.in/gMxRcwKP Grok's Answer: https://lnkd.in/g6dSW2du #AI #MachineLearning #TechInnovation #Grok #Claude #ChatGPT

7

1 Comment

At Parcl Labs, our objective is to provide transparency and accessibility to global housing data. We do this through several mechanisms: ✅ Accessible data to anyone via our API ✅ Transparent research ✅ Improving open source Today, we advance this work into Dubai. When we first started, we quickly recognized a global problem in housing data: 🐌 It was lagged 🛡️ It was inaccessible 😵‍💫 The way it was interpreted and presented often had spin There wasn't consensus on the numbers. TL;DR - no one knew what was happening. We began rebuilding it 3 years ago, one 🧱 at a time. We started in the US. We consumed data from thousands of sources and triangulated across them to create clean signals. Our first product was a daily pricing mechanism for the housing market, speeding up data by months. We called this product the Parcl Labs Price Feeds: ✅ Now available in 100 markets across the country ✅ Updating every day for years ✅ Improving housing data velocity by months ✅ Improving transparency ✅ Accessible to everyone The first place it was available? The Parcl real estate exchange. Now with over $2.1 billion in total volume. And 18 major US markets being traded by people around the world. Others noticed. We've partnered on cutting-edge housing data stories with The Wall Street Journal, CNBC, Moody's, and industry-leading housing data journalists like Lance Lambert. Other venues to trade the price feeds started opening, like with John Dolan and Home Price Futures. All the while, we kept building and expanding coverage: 🇬🇧 Into the UK with our London Price Feed 🇫🇷 And then France with our Paris Price Feed And today, we are pleased to announce another market: our Parcl Labs Price Feed for Dubai. ✅ Updating daily ✅ Transparent methodology (whitepaper in comments) ✅ Accessible to anyone via our API Dubai marks a significant step forward for us on our mission of creating transparency and accessibility in global housing data. Join us and people from around the world in advancing the shared understanding of housing. 🧱 by 🧱

11

3 Comments

rumor I’m hearing: good chance there will be no more step function improvements in base ai models source is a founder building agents on top of base models; he also has connections at base model co's a few takeaways from our convo: 1/ base model intelligence isn't mature enough for pure autonomous agents to work well 2/ a significant step function improvement in base model intelligence would be required for them to work well ... but could be as far out as 12-18 mo away (if they come at all) 3/ autonomous agents were working fine for smaller and easier workflows, but if tasks are complex agents don’t solve reliably. this is leading to lots of customer churn. personally, i think even if the models don't improve much, still think there's tons of innovation and great companies to build what do you think? -- follow me Mike Hinckley

16

NVIDIA's Jensen said "ChatGPT democratized computing, Llama2 democratized generative AI" ...and now Llama3 is the next level 🔥 It's fascinating to watch Meta's strategic moves. With Llama3, they prioritized training on a massive 15T token dataset to pack all into a lean 70B param model, instead of building a massive model. This allows Llama 3 to match trillion+ parameter models like GPT-4, but at 1/10th the compute, storage and inference costs! 💰 This technique was published by Google DeepMind a few years ago https://lnkd.in/gx7VU8aA Meta is not a dark horse in AI anymore. They might be the top dog. https://lnkd.in/gRRJYuxt #llama2 #llama3 #llm #chatgpt #gpt4 #ai #ml

16

5 Comments

It was a pleasure to hear Dr. Ayesha Khanna, the resident guru of #ai, talk about the future of the space. There is not one industry that won’t be disrupted by real artificial intelligence. I don’t mean the stuff that has AI on the surface and is still powered by humans below the hood. Real AI will transform the nature of life and work whether we like it or not. It’s times like these that it is refreshing to hear a presentation at a business conference end on a note revolving around responsibility. Building Large Language Models powered by data from a biased world, will produce biased results. The question is whose vision of the future should we use to guide us? What does a fair world look like? Who gets to decide what “responsible” means? I don’t have the answers, but it’s great to know smart people are trying to tackle the problem. #echelon #e27 #publicpolicy

48

1 Comment

It’s encouraging to see Llama 3 has such good performance. I wonder how other large open source model do (if they participate) on the leaderboard. It will be great to see Grok 1.5 also evaluated here.

6

Insights on Venture Capital for Generative AI from a Scientist-Founder's Perspective As the founder of two genAI startups, I continue to benefit from VC’s perspectives. Here’s a helpful presentation from Omri Amirav-Drory at NFX: What I Wish I Knew as a Scientist-Founder: https://lnkd.in/ergQ6nus A relevant article is here: https://lnkd.in/eyg-pPmi Learning from Omri's talk, my objective is to provide structure and actionable questions. 1. Show potential for significant scale. How do you calculate the market potential? Specifically, what is the number of potential customers (A), the price per unit (B), and the gross margin (C)? Based on these values, what is the projected market cap or revenue that indicates your business can achieve substantial scale? How can you project future market growth? What expansions or new product lines could tap into the existing infrastructure or customer base to drive this growth? 2. Develop a unique technological edge. What intellectual property or technological features ensure that your technology is difficult for competitors to imitate? How do you maintain the exclusivity of your technology? What makes your technology unique and superior to existing solutions? How does it stand out in terms of novelty and effectiveness? How have you solidified the technological edge with strategic intellectual property rights? How do these rights serve as both a shield and a weapon in the market? 3. Assemble a capable, dynamic team. What specific expertise do team members bring that directly correlates with the business needs? How do their backgrounds and previous successes contribute to the potential success of the venture? How does the team’s dynamic facilitate overcoming business challenges, adapting to market changes, and driving innovation? 4. Align the business model with investor expectations. How do your projections resonate with investor incentives, especially considering the typical "2 and 20" structure in venture capital? How and when does the business expect to generate returns? How do these align with the typical 10-year fund lifecycle of venture capital investments? 5. Effectively pitch. How do you ensure your pitch is direct and clear, avoiding unnecessary jargon, and directly addresses how the investment will generate returns? How does your pitch start by showing that the business addresses a big enough problem or taps into a large enough market? #VentureCapital #GenerativeAI #StartupFunding #AIStartups #TechInnovation #Entrepreneurship #ScientistFounders #BusinessGrowth #InvestmentStrategy

3

1 Comment

Reinforcement fine-tuning from OpenAI is here. In a post last week, I highlighted how providing the ability to customize models locally can be a major revenue source for providers but, more importantly, a game changer for the LLM commercialization world. The reinforcement training announcement from OpenAI today is just that, going by the wording of the press release: "This new model customization technique enables developers to customize our models using dozens to thousands of high-quality tasks and grade the model’s response with provided reference answers. This technique reinforces how the model reasons through similar problems and improves its accuracy on specific tasks in that domain." This is the secret sauce if you want to commercialize AI applications in enterprises massively. But as you can imagine, it largely depends on how well and how fast organizations can master the art of "customization". And this in itself is a major business opportunity, helping organizations with this customization. Other providers will replicate this soon. But in the end, the one that can help enterprises "customize" faster, better, and with marketing-worthy results will prevail. In the next five years, you will not see the same landscape of AI model providers. The great news is, enterprises can now paint their own vision, on their own canvas, leveraging such offerings. Exciting times and let the journey begin! #ai #ml #datascience #tech #technology

1

13 years ago @marc andreessen of A16Z wrote a piece titled "Software Is Eating The World" (link in bio!) - arguing that the next 20 years would see the proliferation of software throughout every enterprise/workflow across the world 🥣 🌎 30 days ago, Emergence Capital wrote a piece titled "Service-As-A-Software" (link also in bio!), arguing that with new advances in AI, you could flip the economics of traditional service driven business models and now can offer that service in the form of a low cost, scalable, software offerings - which was impossible before the recent LLM evolution 🤖🚀 We Factored Quality are biased, but I deeply believe the biggest distribution opportunity of Service-As-A-Software is via fundamental platforms/vendors in vertical applications like us, that sit in key enterprise workflows & have strong trust with end users. Over the next few months, Factored Quality is going to release a number of FREE tools for VP's of Supply Chain, Operations, and Dir's of Quality and Regulatory that should automate 95% of what you've historically had to pay consultants to do. We're starting today with our free AI Powered GAP Analysis - a tool that let's consumer product brands / importers / retailers enter in a few questions - and in <30 seconds receive a detailed product regulatory analysis outlining what product safety and testing certifications you need to hit. Check it out and give it a spin below, and then reach out to Andrew Hoagland John Reynolds and Jenson Tuomi to help you run an expert certified analysis + get onboarded to the FQ platform to run audits, inspections, or testing across 28 countries 🔥 Link Here: https://lnkd.in/dEMVk9R7 + More details in bio!

86

6 Comments