Scott Clark

YellowDog.ai just set a 10x benchmark uplift in scale computing, delivering 40,000 tasks per second (TPS) and managing 100,000 compute nodes in the cloud. What's even more interesting, that's 2x IBM Symphony and opens an intriguing pathway for these until-know closed/captive systems.

10

Thanks to the resurgence of RL, LLMs are finally able to reliably coordinate tools and reasoning to do high-precision retrieval. Companies like Happenstance, Clado, and Mintlify have already shifted to agentic search, and it's only a matter of time until anything less feels broken to users. Link to full blog in comments.

20

2 Comments

Over the last 48 hours, 3,000+ people saw the CMEF demo. People tried it. Most asked the same question: "What do I actually DO with this information?" Fair question. So I wrote it down: 9 pages (spaced) on what CMEF is, how it works, and what happens after it catches something. Inside: What a "drift corridor" is (and why the Ukraine question is one) The anatomy of a 3-pass run (with actual entropy scores and flags) 5 concrete governance options when CMEF flags an answer Why "Eastern sources only" makes modern models hallucinate citations This isn't marketing. It's the manual. If you're deploying AI in compliance, healthcare, finance, or anywhere trust matters—this is how you measure what's actually coming out of your models. Before anyone sees it. Authors note - This is a repeat of an earlier post. Why you ask? Because my sense of humor doesn't always track. I left an AI artifact in the original report, because it was hilarious and a way to thumb my nose at the fear porn. However, it has been rightly pointed out that I am not in every room that report may be. So here is the Drift Corridor, minus my own personal humor.

2

1 Comment

The most recent Queued Up report from Berkeley Lab is out. Steven Zhang and I are proud to be listed as coauthors on this edition. The Interconnection.fyi team worked closely with Joseph Rand and the rest of the LBNL team to provide the data that powers this year's analysis. Supporting this level of research is core to our mission of bringing transparency to the wholesale energy markets. The 2025 edition (covering data through the end of 2024) highlights some significant shifts in the landscape: Active Capacity: 2024 closed with nearly 2,300 GW of generation and storage seeking interconnection. Changing Mix: Active natural gas capacity increased by 72% year over year, while solar and storage saw slight decreases in total queue volume. The Backlog: 408 GW of capacity already has an executed or draft interconnection agreement but has not yet reached commercial operations. Timelines: For projects built between 2018 and 2024, the median duration from request to operation has doubled compared to the early 2000s. This report is the definitive annual benchmark for the industry and provides a vital baseline as we begin to see the implementation of FERC Order 2023. The analysis in this report is based on our EOY 2024 data snapshot. Since then, we have continued to track and update these queues every single day. If you want a live view of how these numbers have shifted in the months since this snapshot was taken, follow Interconnection.fyi and subscribe to our Substack. You can find the full slide deck, interactive maps, and raw data files at the link in the comments.

41

3 Comments

this is a huge signal for applied neuroscience. sources report sam altman has hired an award-winning biomolecular engineer, mikhail shapiro, to join his "merge lab". shapiro is known for his pioneering work at caltech. this isnt just another neuralink rival. it looks like a completely different approach. shapiro's research focuses on non-invasive techniques using tools like ultrasound and gene therapy to interact with cells. a stark contrast to neuralinks surgical implants. altman himself has expressed interest in less invasive methods. this is the kind of parallel path in applied neuroscience that is vital. the field is ready for major breakthroughs, and we need multiple innovative strategies to deliver new treatment options and understand the brain.

16

2 Comments

The core science team is led by Caltech's Mikhail Shapiro, Tyson Aflalo, and Sumner Norman. The Scale: Musk's Neuralink raised $280M total across multiple rounds before reaching this level. Merge just did it in one shot. https://lnkd.in/gtX9afbW

10

Chris Lattner's take on CCC (Claude C Compiler) is a great summary of where AI's capabilities are at, in terms of generating code: https://lnkd.in/gGYapXVa Would suggest reading this along with Drew Breunig's take on Claude being an electron app: https://lnkd.in/g7xUsbss

11

Comparing models.. Finding that a small 1B #llm hallucinates less than a 4B for semantic to structural matching. Small llms for targeted tasks for the win. The new Meta llama3.2 1B is better than expected and cheap on token cost.

3

At OpenAI Jerry Tworek helped drive o1, o3, and Codex as VP of Research. Then he left to pursue “types of research that are hard to do at OpenAI.” This week on Unsupervised Learning, I sat down with Jerry to discuss where AI research is headed and what he learned from seven years at the forefront of the field. We hit on: ▪️ Why Jerry updated his AGI timeline after building reasoning models and what fundamental capability current models are missing ▪️ The real limits of scaling reinforcement learning  ▪️ Why he left OpenAI after creating some of its biggest breakthroughs  ▪️ Why Anthropic did so well in coding  ▪️ Inside OpenAI's pivotal 51-49 decisions  ▪️ What makes great AI researchers Why current models can never be AGI without continual learning Jerry's biggest update is his belief that static models can never achieve AGI. When models hit walls, they become "hopeless" and stay stuck. True intelligence always finds a way by probing problems until solved, but current models lack the ability to update their beliefs based on failure. As Jerry puts it: "Intelligence always finds a way... which the current models do not really." The fundamental limits of scaling RL While scaling RL delivers predictable benefits, you fundamentally get what you train for. Models excel at specific RL skills but struggle with generalization beyond their training distribution. Scaling primarily means adding more data, and this loop is slow and labor-intensive. The question isn't whether RL works, but whether there's research that could give us better generalization with less data. Why Jerry left OpenAI to chase the next paradigm shift After helping introduce reasoning models, which Jerry describes as a "tectonic shift," he wants to chase that high again. He's now pursuing what he believes is missing in how the world trains models, seeking freedom to explore the most important unsolved problems differently than everyone else has tried. Listen here👇 Spotify: https://bit.ly/4brrgTE Apple: https://bit.ly/4a2zjnc YouTube: https://lnkd.in/ekyv-znp

46

3 Comments

AI is advancing at breakneck speed. What I imagined might be years away -- compute finding non-obvious insights across science papers yielding novel discovery -- has arrived: “What distinguishes Zochi is its ability to identify non-obvious connections across papers and propose innovative solutions that address fundamental limitations rather than incremental improvements." https://lnkd.in/gqtz2aSj

77

19 Comments

How to Start a Biotech Company (Sourish Saha) A practical primer aimed at scientists turning entrepreneurial, with guidance on strategy and startup fundamentals. https://lnkd.in/gAn4R9x7

1

#TuesdayPaperThoughts Edition 60: The Art of Scaling RL Compute for LLMs This week's #TuesdayPaperThoughts explores "The Art of Scaling Reinforcement Learning Compute for LLMs" from researchers at Meta, The University of Texas at Austin, University of California, Berkeley, and Harvard University. While pre-training has had its moments with predictable scaling laws, RL training has remained more art than science until now. Key Takeaways: 1️⃣ Predictive Scaling Framework: The work introduces sigmoidal compute-performance curves for RL training that separate asymptotic performance (A) from compute efficiency (B). This framework enables extrapolation from smaller-scale runs to predict performance at larger compute budgets. The team validated this with a massive 100,000 GPU-hour run where predictions from just the first 50k hours closely matched final performance. 2️⃣ Not All Recipes Scale Equally: Methods that look promising at small compute budgets can hit lower performance ceilings at scale. The study reveals that design choices like loss type and FP32 precision shift asymptotic performance, while factors such as loss aggregation and normalization primarily modulate compute efficiency. This explains why some widely-adopted methods plateau unexpectedly. 3️⃣ ScaleRL Recipe: Through systematic ablations consuming 400,000+ GPU-hours, the team developed ScaleRL—combining PipelineRL with 8-step off-policyness, truncated importance sampling (CISPO), FP32 logits computation, and adaptive prompt filtering. ScaleRL achieves A=0.61 asymptotic performance, outperforming DeepSeek's GRPO, Qwen's DAPO, and other prevalent methods on both ceiling and efficiency. The timing couldn't be better. With RL compute budgets exploding (10× increase between model generations for o1→o3 and Grok-3→Grok-4), the field desperately needed this systematic approach. Research Credits: Devvrit K., Lovish Madaan, Rishabh Tiwari, Rachit Bansal, Sai Surya Duvvuri, Manzil Zaheer, Inderjit Dhillon, David Brandfonbrener, Rishabh Agrawal Paper Link: In comments

21

1 Comment

Just read this paper out of UC Berkeley (Vafaii, Galor, & Yates). Fascinating study. In just the first few paragraphs they summarize a range of references that unify variational autoencoders, diffusion models, and predictive coding (the neuroscience version) within the Bayesian learning framework. They go on to derive a recurrent spiking neuron just from the two principles of free energy minimization and a poisson firing rate, demonstrating experimentally that it has many neurally plausible properties. https://lnkd.in/g2XJSrxa

13

Great to see how a computational physics research lab is using Pulse's API to extract LaTeX equations and mathematical notation from their 40+ year archive of papers. Their team built a specialized pipeline that preserves the semantic structure between equations, preserving relationships between variables, theorems, and proofs. What's remarkable is how Pulse correctly identifies and extracts nested mathematical structures (matrices within summations within integrals) while maintaining proper symbol relationships. Our mission is to turn previously inaccessible computational knowledge into structured, searchable data that accelerates novel research. If you're a research team interested in trialing out Pulse, send me a DM for some trial credits! Sample LaTeX output attached below.

64

1 Comment

I was listening to this https://lnkd.in/giycVEkG and had a thought I can't shake: What if foundation models are AI's genetic code? DNA gives us instincts... babies cry, calves walk to milk without learning. Foundation models feel similar: pre-loaded with language and reasoning. But evolution doesn't create one perfect organism. It creates fish, mammals, birds... each specialized. Maybe AI works the same way? Not one AGI, but multiple species and individuals that learn and evolve. If foundation models are our "genetic code"... we need to be careful what we embed there. These become instincts for all future AI. Like Sutton says, learning is fundamental and the next crucial piece: how do we take what AI systems learn and feed it back to the next generation? In biology, successful traits get passed down... what's the AI equivalent? Maybe we're not just building better AI... maybe we're designing artificial evolution.

11

I’ve tried and used many tools surrounding LLMs, such as LangChain/Graph, LlamaIndex, vector stores like Qdrant, and full AI platforms like Azure AI Foundry. I don’t understand why none of them offer RAG applied to their own documentation - it seems like such an easy win. Many of these technologies are new and evolving quickly, so getting grounded information from places like ChatGPT can be difficult. Trawling through documentation feels unnecessary in the age of LLMs. Maybe someone can enlighten me.

18

9 Comments

Ever wondered how JAX makes machine learning so fast? Just dropped Part II of our series, "From JVP to VJP," to explain. At Pasteur Labs, we use JAX to build complex simulation systems — and automatic differentiation makes it all possible. This post breaks down JAX's clever approach to computing gradients efficiently. If you've used jax.grad() and wondered what's happening behind the scenes, this demystifies the magic. ✨ Part of our ongoing series on building next-gen simulation systems. Thanks Niklas Heim for the excellent post! 👀 Read it and our other Insights posts: https://lnkd.in/eHi25bBE

25

1 Comment

I’ve been experimenting with structured knowledge graphs, optimizing semantic search, and distilling reasoning frameworks that sit on top of LLM outputs. Despite all the progress, no one has yet solved how large language models can offer meaningful guidance about the future. The real breakthrough may come not from bigger models, but from architectures that build persistent agentic memory, systems that learn, recall, and reason more like the human brain. DM me if you see something worth a deeper conversation.

51

Voice is a super interesting modality right now - maybe the first modality we're seeing move to open source models across a number of scale ups / enterprises. Reliability concerns, high costs, and open source model performance are pushing engineers to do their own fine tuning vs. relying on third-party vendors of proprietary models. Many of these orgs have already been collecting their own first-party data and now with third-party vendors like Extrian, David AI, etc they can train really high quality models. RL has been insanely hyped, but it's been unclear how long it will take scale ups and enterprises to actually lean in. Voice AI might be hitting that inflection point faster than expected.

39

3 Comments