Bahman Rabii

Can LLMs strengthen LP relaxations and speed up solver convergence on MILPs? This is a question that we (Milad Yazdani, Samin Aref, Zirui Zhou, and I) explored in our recent work, and the results are promising. In this year’s IMO, leading AI systems (from OpenAI, Google/DeepMind) achieved a gold-medal–level score by solving 5 out of 6 problems, yet P6 (a notoriously difficult problem) remained unsolved. We modeled P6 as an integer program (LLMs, with some guidance, also produced correct formulations). Still, top MILP solvers did not converge in a reasonable time, suggesting a weak LP relaxation. Our method: EvoCut. EvoCut combines evolutionary search + LLMs to automatically discover problem-specific cuts (valid inequalities) that strengthen the relaxation. Results: - ~55% average gap reduction at the same runtime - Order-of-magnitude faster convergence in many instances We also observed similar improvements on other combinatorial problems across different domains. Read more & try EvoCut: - Paper: EvoCut: Strengthening Integer Programs via Evolution-Guided Language Models (https://lnkd.in/g6Kydtp6) - Code: Test EvoCut on your own MILP (https://lnkd.in/g42wMy4s) #MILP #OperationsResearch #Optimization #AI #LLMs #CuttingPlanes #Combinatorics #IMO #LinearProgramming

123

4 Comments

Today we’re launching our Physical AI Lab, led by UCLA professor and renowned researcher Bolei Zhou, who’s joining Coco as Chief AI Scientist. Bolei has spent his career advancing physical AI with 60,000+ citations and pioneering work. Now, he’s bringing that expertise to Coco. What makes this lab special is Coco’s fleet itself. Our robots run urban deliveries across the U.S. and Europe, navigating unpredictable environments (construction, pedestrians, weather, etc.). These are the kinds of real world challenges you can’t capture in simulation and that data is becoming the raw material for advancing physical AI. As we scale to 10,000+ robots in 2026, this fusion of research and deployment will define the next generation of physical AI and reimagine how goods move through cities everywhere. TechCrunch has the scoop - link in comments.

213

12 Comments

I was honored to host a fireside chat with Sherwin Wu, head of engineering for OpenAI API at last week's Fellows Forum 2025 organized by Fellows Fund. Sherwin shared his incredible journey at OpenAI in scaling OpenAI's developer platform and building its vibrant ecosystem, and his deep insights into various topics we covered.   A few quick notes:   Model Evolution & Platform Development: Outlined 3 major model capability inflection points 📍GPT-3: Limited to copywriting applications 📍GPT-3.5/4: Enabled the post-ChatGPT explosion (Perplexity, Snapchat integrations) 📍Reasoning paradigm (O1): Unlocked long-term planning for coding and autonomous agents   Lessons from GPT-5 launch 📍Latency concerns: Model's extensive thinking time frustrated users 📍Product showcase mismatch: ChatGPT interface didn't effectively demonstrate the model's capabilities 📍Learning: Future major launches will pair API releases with first-party products that better showcase model strengths   Unexpected use cases 📍Consumer apps like enabling toys for kids to chat with   Common patterns in startups that successfully leverage OpenAI models and API 📍Most successful startups "swim with the current" - building products that push models to their limits and position for future capabilities rather than current limitations   Common patterns in enterprises that successfully leverage OpenAI models and API 📍Common use cases include internal productivity enhancement (e.g., using Codex) and customer interaction transformation (support, sales automation)   Barriers in enterprise Adoption 📍Lack of both top-down and bottom-up organizational buy-in (top-down alone is not sufficient) 📍Insufficient AI-ready infrastructure and data platform investment   Responsible use: OpenAI is investing in 📍Pre-launch Red-teaming in Model Development 📍Post-launch monitoring 📍Tools to help developers manage safety risks, build guiderails, and setup evals   Exciting future breakthroughs 📍Excitement centers on advancing the reasoning paradigm (longer thinking periods) and multimodal breakthroughs, particularly in voice interactions that feel more naturally conversational Full fireside chat video: https://lnkd.in/gJzjzMBm

120

2 Comments

How do LLMs really compare to the human brain? In a sequence of studies, our team at Google Research and partners at Princeton, HUJI and NYU have been exploring this question, aligning brain activity with the internal contextual embeddings of speech and language within LLMs as they process everyday conversations. What have we discovered? AI language models and the human brain share underlying organizational principles for representing words. Specifically, the internal representations used by these AI systems correspond to the neural activity patterns observed in brain regions crucial for both understanding and generating speech, hinting at potential shared computational mechanisms, all in all demonstrating the power of deep language model’s embeddings to act as a framework for understanding how the human brain processes language. Thank you to our research partners: Hasson Lab at Princeton University, the DeepCognitionLab at the Hebrew University, and researchers from the NYU Langone Comprehensive Epilepsy Center. Google Research Blog Post, authored by Mariano Schain and Ariel Goldstein: https://lnkd.in/d9cbP4Ee Nature Human Behaviour Paper: https://lnkd.in/dCbvdC-A Nature Communications Paper: https://lnkd.in/dbPCGY_m Nature Neuroscience Paper: https://lnkd.in/dvWjEg3C Hasson Lab, Princeton: https://lnkd.in/diMh8RA4 DeepCognitionLab at the Hebrew University: https://lnkd.in/dFPDuM7x  NYU Langone Comprehensive Epilepsy Center: https://lnkd.in/dU8rWe5e

680

26 Comments

The story of the rise of Nvidia, the world’s first 4 trillion dollar tech company, and it’s relentless CEO, Jensen Huang, is the most extraordinary business story I have ever heard. No one knows it better than New Yorker writer Stephen Witt, author of the acclaimed book, The Thinking Machine, who I recently had as a guest on the Next Big Idea podcast. I asked Stephen after our podcast conversation — What can we learn from Jensen about how to build successful companies? His answer — for starters, “You have to be completely paranoid about the competition at all times… and you have to work as hard as you can every day of your life.” Okay ... But the real differentiator is where you choose to compete. Witt explains that Jensen deliberately chased what he called a “zero billion dollar market.” Instead of sprinting toward the obvious TAM, he built low-budget supercomputers—turning graphics cards into tools for scientists—“a very small market, so nobody followed him.” Then he grew the customers around him until the “small market” wasn’t small anymore. Translation: build where others won’t go, not just where they can’t go. My favorite operating principle from the conversation: “Work at the speed of light.” That doesn’t mean “go fast.” It means identify the physically fastest a thing could be done, then work backward to your schedule. Once you anchor to the theoretical ceiling, “you’re not constantly looking over your shoulder,” because your competitors can’t beat physics either. You’re racing the limit, not the noisy neighbor. And then there’s Jensen’s ruthless time hygiene: “Do the hardest—actually the most important—thing first.” Even with “60 meetings a day,” he carves out protected time for the work that moves Nvidia’s stock over the very long term. As Witt put it, the team is comfortable making five-year and even ten-year bets—and Jensen touches the long-term project every day. Here’s a summary of Jensen’s playbook: - Chase the zero-billion market and let customers scale it with you. - Set timelines by the fastest physically possible, not the median meeting estimate. - Front-load the highest-leverage task before the day gets noisy. - Make decisions on a 5–10 year horizon, then show up for them daily. Do you start the day by doing the task that is most critical for your long term success? Would love to hear your answers in the comments. 🚀

106

14 Comments