Umut Eser

The new platform we have been working on, @weave_bio, is partially out of stealth. Stay tuned for exciting updates: weave.bio

@Yesterday_work_ Only the 14th author is somewhat affiliated with NVIDIA and the whole paper becomes “NVIDIA paper”.

I'm reading NVIDIA's new paper and its wild. Everyone keeps talking about scaling transformers with bigger clusters and smarter optimizers… meanwhile NVIDIA and Oxford just showed you can train billion-parameter models using evolution strategies a method most people wrote off as ancient. The trick is a new system called EGGROLL, and it flips the entire cost model of ES. Normally, ES dies at scale because you have to generate full-rank perturbation matrices for every population member. For billion-parameter models, that means insane memory movement and ridiculous compute. These guys solved it by generating low-rank perturbations using two skinny matrices A and B and letting ABᵀ act as the update. The population average then behaves like a full-rank update without paying the full-rank price. The result? They run evolution strategies with population sizes in the hundreds of thousands a number earlier work couldn’t touch because everything melted under memory pressure. Now, throughput is basically as fast as batched inference. That’s unheard of for any gradient-free method. The math checks out too. The low-rank approximation converges to the true ES gradient at a 1/r rate, so pushing the rank recreates full ES behavior without the computational explosion. But the experiments are where it gets crazy. → They pretrain recurrent LMs from scratch using only integer datatypes. No gradients. No backprop. Fully stable even at hyperscale. → They match GRPO-tier methods on LLM reasoning benchmarks. That means ES can compete with modern RL-for-reasoning approaches on real tasks. → ES suddenly becomes viable for massive, discrete, hybrid, and non-differentiable systems the exact places where backprop is painful or impossible. This paper quietly rewrites a boundary: we didn’t struggle to scale ES because the algorithm was bad we struggled because we were doing it in the most expensive possible way. NVIDIA and Oxford removed the bottleneck. And now evolution strategies aren’t an old idea… they’re a frontier-scale training method.

⏱️ When Health Authorities have questions, every hour counts. Meet HAQ Manager, Weave’s new AI-native solution that automates and accelerates Health Authority Question responses — built with Takeda. 🔗 Read more: businesswire.com/news/home/2025… 🎥 Register for the live demo (Nov 14): weave.bio/webinars/intro… #WeaveBio #HAQ #DrugDevelopment #RegulatoryInnovation #AIinPharma #Biotech #LifeSciences

⚡ Last week, the FDA took an important step toward transparency. By publishing the same filing checklists CDER reviewers use internally, the FDA is helping sponsors understand what “complete and reviewable” really means. It’s a practical move that could prevent avoidable Refuse to File (RTF) delays that can cost teams months or even years. You can read the full announcement here: 👉 fda.gov/news-events/pr… For anyone managing regulatory data, this release underscores what we see every day: clarity and organization aren’t just nice to have, they’re make-or-break. At Weave, we help teams stay ahead of these moments. Our platform connects data, context, and collaboration so you’re never scrambling to meet the standard—you’re already there. Learn how we can make your next submission go smoother at weave.bio. #WeaveBio #FDA #DrugDevelopment #RegulatoryInnovation #AIinPharma #Biotech #LifeSciences

We’re growing.

It’s now official! Excited to announce that we raised $20m Series A. I can already feel the phase transition from proving the value proposition to scaling and capturing the value creation. So proud of the team @weave_bio !

Exciting news! We're thrilled to announce our strategic partnership with @Parexel ! Parexel brings deep regulatory consulting expertise that enhances our AI solutions, enabling us to expand The Weave Platform across the entire regulatory lifecycle. Together, we’ve already delivered proven results with 50% faster IND authoring timelines, helping sponsors initiate clinical trials sooner while maintaining the highest quality and standards. Read the full announcement: newsroom.parexel.com/news-releases/…

LLMs aren’t dumb — they’re unstructured. Give GPT-4.1-mini a CoT instruction 10-digit multiplication: ❌ Give GPT-4.1-mini an algorithmic scaffold: ✅ 70–100% accuracy. 👉 Prompting = programming. 👉 Structure unlocks hidden capacity. Full essay here: umuteser1.substack.com/p/prompting-is…

LLMs can write poetry but can’t count stripes on a shoe. Intelligence needs structure, not scale. I ran an experiment in the most absurd lab: Brainfuck. Multi-scale solutions emerged! Full story 👇 umuteser1.substack.com/p/teaching-ai-… #AI #Emergence #NeuroSymbolic

Vision: Add a stripe to Adidas → AI still says “three.” Math: Add an irrelevant clause → accuracy drops 65%. Medicine: Replace answer with “None of the above” → collapse. Pattern ≠ reasoning. Post + examples → umuteser1.substack.com/p/the-reasonin…

💬 𝗪𝗲𝗮𝘃𝗲 𝗪𝗲𝗱𝗻𝗲𝘀𝗱𝗮𝘆 Your words keep us moving forward. Every time we hear how 𝗧𝗵𝗲 𝗪𝗲𝗮𝘃𝗲 𝗣𝗹𝗮𝘁𝗳𝗼𝗿𝗺 saves you hours, catches the details you missed,  or helps regulatory writing feel less like a bottleneck — we listen. And we build. This feedback doesn’t just make us smile (although it does) — it shapes our roadmap. It’s the reason we keep pushing to make The Weave Platform faster, smarter, and even more helpful for biotech and pharma teams. Your feedback isn’t just heard — it’s built into the future of The Weave Platform. 🚀 #WeaveWednesday #CustomerFeedback #Biotech #Pharma #RegulatoryWriting #ProductInnovation

After getting many positive feedback, I decided to write the webinar content as a Substack post: umuteser1.substack.com/p/mastering-th… Audience: non-developers I might write something more technical laying out the emerging best practices across domains that I have seen so far in the future.

🫶 🫶 Our customers are the reason Weave exists — and the reason we keep raising the bar. Whether they’re leading small but mighty teams in early-stage biotech or driving regulatory execution in later-phase clinical programs, they share one thing: They’re focused on the science — and they don’t have time for inefficient workflows. We built Weave to give them the space to focus on what matters most — not stitching together content by hand, and not managing formatting and version chaos. We’re proud to support the people who push science forward — and inspired by the work they do every day. These logos represent just a handful of the forward-thinking teams using Weave. We encourage you to check out their work — and imagine what’s possible. 🫶 #WeaveWednesday #SubmissionReady #BiotechOps #PoweredByOurPartners

RT @weave_bio: Sneak Peek Alert: AutoCT at SCT 2025 Headed to Vancouver for the @SCTorg Society for Clinical Trials meeting on May 19–20?…

Super Founders 📕 summary If a founder has prior experience of being a successful founder, their next startup has an 8x higher chance of becoming a unicorn. “This was the strongest difference I observed in my data between the billion-dollar and the random group.” — author Ali Tamaseb “Almost all the founders of billion-dollar companies are builders. Going back to high school, or even younger, many of them started side projects, student clubs, or even companies. Those projects weren’t about striking it rich so much as scratching an itch to create.” Study period: 2005 – 2018, 200 unicorns vs a random sample of 200 non-unicorns that also raised at least $3M of VC. It’s “one of the largest data sets ever compiled on startups and what leads to their success.” “Over 90% of billion-dollar startups were VC backed.” Co-investment with a top tier VC -> 5x the outcome. "Some of the largest companies in my study were founded during the recession. Perhaps innovation comes from constraints.” “Unicorns make up less than 0.1% of startups.” “Just over half of the unicorns were headquartered in the SF Bay Area, compared to 1/3 in the random group.” Not correlated with success: age of founders, number of founders, or domain expertise. "The mythology doesn't always match reality." “Fewer then 50% of founding CEOs had much, if any, work experience related to their startup. In science-related startups, the story is a little different. 75% of founders in biotech had direct relevant experience, enterprise tech 40%, and in consumer only 30% had direct relevant experience.” “The founders of billion-dollar companies were more likely to work for themselves or to have worked at a tier-one company.” “If there is one takeaway from this book, the best preparation for starting a wildly successful company is starting a company. If you have never started a company, the best preparation is to start something, maybe a club or a side hustle.” Unicorn founders “tend to have a compulsion to build stuff.”

If you are developing an SDK, create a button for exporting documentation for people to provide context to their favorite smart-IDEs or AI Assistant apps like ChatGPT or Claude. Maybe Cursor/Windsurf/Augment etc. already lets us to set the API doc sites to pull in?

i was surprised to learn how much of modern AI was developed by physicists: - scaling laws - alphafold - RLHF (dario got his phd in biophysics) - diffusion models - normalizing flows - VAEs - neural tangent kernel - hopfield networks - information bottleneck theory - neural style transfer - imagenet (apparently feifei has a bachelors in physics) - chinchilla optimality

✨ #𝗪𝗲𝗮𝘃𝗲𝗪𝗲𝗱𝗻𝗲𝘀𝗱𝗮𝘆 ✨ This week: 𝘁𝗵𝗲 𝗘𝗻𝗵𝗮𝗻𝗰𝗲𝗱 𝗗𝗮𝘁𝗮 𝗥𝗼𝗼𝗺 in #AutoIND. 🔍 The Data Room just got a whole lot smarter. Can’t find that one study? Need to answer Health Authority questions—fast? With the Enhanced Data Room, AutoIND now gives you a whole new way to interact with your data: 📁 Upload smarter   • Drag and drop full folders directly into the Data Room   • All PDFs and subfolders upload automatically—structure preserved   • Batch upload multiple folders and files with no extra manual work 🔎 Search better   • Keyword search across names, tags, and metadata   • Deep search uses AI to understand intent—not just words   • Instantly preview results, auto-scrolled to the relevant page 💬 Ask anything   • Use the Ask tab (formerly Explore) to query your files   • Save, rename, and revisit your questions   • Get draft responses with direct references and downloadable sources #WeaveBio #AutoIND #WeaveWednesday #DataRoom #RegulatoryTech #AIinBiotech

Blistering speed, real rigor. With a Top 20 Pharma partner, we used AI to draft 94 study summaries across 2 INDs. Avg time: 3.1 hrs vs. 100+ hrs manually (33x faster). Faithfulness: 96%. Quality: 74% on first draft. Evaluation? Still best done by expert humans. 👉 weave.bio #WeaveBio #AIinPharma #RegulatoryWriting #Biotech #PharmaTech

🚀 Master the Art of Prompting for Regulatory, Clinical, and Corporate Excellence Missed our live session? The recording is now available! Learn how to unlock the full potential of AI in your day-to-day workflows—from submission drafting to executive decision-making—with practical, real-world prompting strategies. 🔍 What You’ll Learn: ✨ Anatomy of an Effective Prompt ✨ Advanced Prompting Techniques ✨ Regulatory & Corporate Use Cases ✨ Shaping the Future of AI Interaction 📽️ Access the recording here: weave.bio/podcasts/maste… The future belongs to those who know how to talk to AI. Start mastering it today. #AIinRegulatory #ClinicalTrials #PromptEngineering #RegulatoryAffairs #AIProductivity #DrugDevelopment #MedicalWriting #ClinicalOperations #FutureofWork

Many people asked to re-watch the recording of art of promoting webinar. Some say there were unique insights hard to find online. So, I uploaded a bit shortened version of it to YouTube. It’s the very first YouTube video of my channel: youtu.be/vJVv8boZTFc?si…