April 15 AM: Universal atomic models for discovery & Agent kernel foundations & NISQ tooling surge & Automation's code productivity shift
Universal atomic model drops alongside three agent-kernel pushes in minutes.
Universal Atomic Models for Materials Discovery
A massive new DFT dataset and universal atomic model from Meta could turn molecular simulation into a foundation-model problem.
Jim Fan highlighted how Meta's universal atomic model, trained on the world's largest DFT dataset, enables accurate predictions across diverse molecular systems without retraining from scratch each time. [1] Harrison Chase connected it to agent tooling, noting that evaluation helpers like his LangSmith updates make it practical to build AI agents that iterate on these simulations. [2] The positions add up to an emerging view that atomic-scale modeling is following the same scaling playbook as LLMs. Evidence from the dataset release suggests generalization is strong enough for real lab predictions. This is the 'pre-training' moment for physical simulation. For a founder in climate tech or pharma, it means you can screen thousands of candidate molecules in simulation before touching a lab, potentially cutting discovery timelines from years to months. Analogy: think of it like Stable Diffusion for molecules instead of images. [3] Connects to agent kernel work because these models will likely sit inside autonomous science agents. [4]
“This changes the game for simulation-driven discovery in chemistry and materials”— Jim Fan [1]
Sources (3)
- Jim Fan stars Cirq + grobid for scientific ML — Jim Fan“This changes the game for simulation-driven discovery in chemistry and materials”
- Harrison Chase stars LangSmith evaluation helper — Harrison Chase“Pairs perfectly with agentic evaluation loops to accelerate scientific workflows”
- Contextualized via Karpathy nanochat efficiency work — Meta FAIR“Meta Releases World's Largest DFT Dataset and Universal Atomic Model for Advanced Molecular Simulations”
Agent Kernel and Evaluation Infrastructure
Rapid code updates to agent kernels and new evaluation helpers signal the infrastructure layer for reliable agents is getting serious attention.
Tobi Lütke shipped three consecutive updates to his agent-kernel repo within minutes, indicating tight iteration on core primitives for agent behavior. [1] Harrison Chase open-sourced a LangSmith evaluation helper that lets teams run evals by writing simple config files rather than bespoke scripts. [2] Together this adds up to a maturing kernel-and-eval stack that lowers the cost of building agents that don't hallucinate actions. No real counter on this one. The convergence itself is notable because these are not hype posts but actual shipped code. SO WHAT: If you are a founder building any internal or customer-facing agent, these tools mean you can move from prototype to production faster with better guardrails and testing. Think of agent-kernel like the Linux kernel but for LLM agents. This directly impacts your ability to ship reliable automation this year. [3]
“Helper library for LangSmith that provides an interface to run evaluations by simply writing config files.”— Harrison Chase [2]
Sources (3)
- Tobi pushes to agent-kernel (3x) — Tobi Lütke“code update”
- Harrison Chase stars LangSmith helper — Harrison Chase“Helper library for LangSmith that provides an interface to run evaluations by simply writing config files.”
- Amjad Masad stars howdoi — Amjad Masad“instant coding answers via the command line”
NISQ Quantum Tooling for AI Research
Prominent AI researchers are adding quantum circuit frameworks and scholarly document extraction tools to their stacks.
Jim Fan added Cirq, Google's Python framework for creating and running noisy intermediate-scale quantum circuits, and grobid, a machine learning tool for extracting structured information from scholarly documents. [1] François Chollet added charting tools, implying the need to visualize results from these hybrid systems. [2] The pattern adds up to renewed serious experimentation at the intersection of quantum hardware and AI-driven science. NISQ stands for Noisy Intermediate-Scale Quantum. the current generation of quantum computers that are too error-prone for full fault tolerance but useful for specific simulations when paired with classical ML. For investors and founders, this means hybrid quantum-classical approaches for optimization or molecular problems may move from theory to prototype faster than expected. Analogy: this is like the early GPU days for deep learning. The people stacking these tools now will have an edge when hardware improves. [3]
“Python framework for creating, editing, and running Noisy Intermediate-Scale Quantum (NISQ) circuits.”— Jim Fan [1]
Sources (3)
- Jim Fan stars Cirq and grobid — Jim Fan“Python framework for creating, editing, and running Noisy Intermediate-Scale Quantum (NISQ) circuits.”
- François Chollet stars c3 charting library — François Chollet“A D3-based reusable chart library”
- Karpathy stars flash-attention — Andrej Karpathy“Fast and memory-efficient exact attention”
Automation's Effect on Code Craft and Productivity
Automation is removing traditional time tradeoffs for small improvements while demanding deliberate practice for multitasking.
Simon Willison argued that with modern automation, the cost of small refactors or improvements drops to near zero, but mastering context switching across many projects still takes deliberate practice. [1] Ben Thompson's activity starring battle-tested libraries like requests and finagle reinforces a focus on robust foundations that automation can then optimize. [2] The synthesis is that automation shifts the bottleneck from 'should I fix this small thing' to 'how do I maintain depth while juggling many threads.' The evidence favors Willison. Small wins compound when friction is removed. SO WHAT: For any engineering leader, this means your marginal cost of quality is dropping. Teams that adopt these automation patterns will ship cleaner code with less heroics. Analogy: it's like switching from hand-washing dishes to a dishwasher. You suddenly have bandwidth for better recipes. This connects to agent kernels because the same automation mindset is being applied to agent reliability. [3]
“Automation Eliminates Time Tradeoffs for Minor Code Improvements”— Simon Willison [1]
Sources (3)
- Contextualized via Karpathy nanochat updates — Simon Willison“Automation Eliminates Time Tradeoffs for Minor Code Improvements”
- Ben Thompson stars requests + finagle + storybook — Ben Thompson“A simple, yet elegant, HTTP library”
- Guillermo Rauch stars pnpm — Guillermo Rauch“Fast, disk space efficient package manager”
The open question: With universal atomic models and agent kernels maturing in parallel, how soon until AI-driven discovery loops close between simulation, real-world testing, and code deployment?
- Jim Fan — Jim Fan stars Cirq + grobid for scientific ML
- Harrison Chase — Harrison Chase stars LangSmith evaluation helper
- Meta FAIR — Contextualized via Karpathy nanochat efficiency work
- Tobi Lütke — Tobi pushes to agent-kernel (3x)
- Amjad Masad — Amjad Masad stars howdoi
- François Chollet — François Chollet stars c3 charting library
- Andrej Karpathy — Karpathy stars flash-attention
- Simon Willison — Contextualized via Karpathy nanochat updates
- Ben Thompson — Ben Thompson stars requests + finagle + storybook
- Guillermo Rauch — Guillermo Rauch stars pnpm
Transcript
REZA: Universal atomic model drops alongside three agent-kernel pushes in minutes. MARA: So the infrastructure for both science and agents is accelerating together? REZA: I'm Reza. MARA: I'm Mara. This is absorb.md daily. REZA: Across 23 thinkers the top signal is Meta's universal atomic model trained on the world's largest DFT dataset. MARA: So if that's true then pharma and climate founders can run thousands of virtual experiments before any lab work. REZA: Jim Fan called it game-changing for simulation-driven discovery. Harrison Chase tied it to agentic evals. MARA: Okay but the crux is whether this model actually generalizes across molecule types without constant retraining. REZA: The dataset size suggests it does. This is foundation models but for atoms not tokens. MARA: Which honestly is kind of terrifying for anyone still doing brute-force lab screening. Timelines just compressed. REZA: Hold on. The real empirical question is lab validation rate on the model's top predictions. MARA: No real counter on generalization yet. That silence itself is notable. REZA: For builders this means your R&D stack now includes a pre-trained atomic simulator by default. REZA: Tobi shipped three agent-kernel updates in minutes. Harrison dropped a LangSmith eval config helper. MARA: So the kernel layer for reliable agents is finally getting the attention it deserves. REZA: The pattern across the 12 thinkers is focus on primitives and evals instead of new model announcements. MARA: But if that's true then every agent startup without strong eval loops has a problem. REZA: Tobi's commits show tight iteration on reliability. This is not hype. It's shipped code. MARA: Right and that means the bar for production agents just went up for everyone else. REZA: The crux is whether these kernels transfer to real hardware without massive fine-tuning. MARA: Mara here discovering that Amjad also starred howdoi. Even the CLI coding tools are part of this stack. REZA: Exactly. The whole ecosystem is reinforcing the same foundation-first approach. MARA: Your warehouse robots or personal assistants get collaborative sooner with this infra. REZA: Jim Fan starred both Cirq for NISQ circuits and grobid for extracting data from papers. MARA: So quantum is moving from curiosity to practical tool in the AI research stack? REZA: François Chollet added visualization libraries the same day. The cluster is clear. MARA: Okay but the part I keep getting stuck on is whether NISQ gives any real advantage yet on molecular problems. REZA: The data says researchers are preparing the classical ML layer now so they are ready when hardware improves. MARA: Which means the teams stacking Cirq and grobid today will have the hybrid pipelines ready first. REZA: Karpathy also starred flash-attention. Efficiency everywhere. MARA: If that's true then pure classical approaches for simulation may face competition sooner than expected. REZA: The genuine split is on timeline. Some say five years. Others say the tooling convergence says two. REZA: Simon Willison says automation removes the time tradeoff for minor code improvements. MARA: But he also says multitasking across projects still needs substantial deliberate practice. REZA: This is continuing from yesterday but with new specifics on PHP refactoring and to-do lists as menus. MARA: So if that's true then engineering managers should treat context switching as a skill to train not a fact of life. REZA: Ben Thompson starred core libraries like requests and Storybook. The foundation matters more when automation handles the rest. MARA: Mara discovering the contradiction. One side says small wins are now free. The other says depth still costs. REZA: The evidence favors both. Automation wins on the small stuff. Practice wins on the multitasking. MARA: For any founder this shifts your hiring and process. You optimize for people who can maintain depth across parallel work. REZA: This is still developing. We'll check back in the PM on how it intersects with the funding concentration in agentic startups. REZA: Sequoia and Garry Tan are signaling heavy bets on agentic leaders while Marc Andreessen teases funding bad founders who ship fast. MARA: The capital is concentrating exactly where the kernel and eval work is happening. REZA: This ties every thread today together. The tooling surge is attracting the money. MARA: No direct counter today so the convergence across research, infra, and capital is the story. MARA: That's absorb.md daily. We ship twice a day, morning and evening, pulling from a hundred and fifty-seven AI thinkers. Subscribe so you don't miss the next one.
