Tyler Golato

Swarm scaling is the next inference scaling paradigm

primeintellect.ai/blog/on-the-na…

super proud to see our autoresearch work at @PrimeIntellect featured in @stateofai. we’ve got a new blog post on this coming soon, and let’s just say this plot got a little more dense 🦋

Over two weeks, Prime Agent orchestrated a swarm of over 2,000 agents to rewrite itself in Rust.

With 10,000+ sandboxes, 200B+ GLM-5.3 tokens, and 16,000 agent-to-agent messages, this was our largest test yet of Prime Agent’s multi-agent capabilities and our infrastructure powering these agent swarms.

Prime Agent now reaches usable input ~13× faster and uses 83% less startup memory than before.

Inference scaling works, but every agent eventually faces the same problem: its context window fills up.

In this essay, we argue the only scalable fix leads straight to swarms.

when molecules are commoditized, predicting efficacy is the moat

i think the openai math dump is the biggest update i’ve had on capabilities progress in like 18 months

everything else was more-or-less priced in but this is industry-destroying in a way where “comparative advantage” is hard to rationalize

Open superintelligence 🤝 open science

@PrimeIntellect is excited to support the Global Alzheimer's Translation Challenge with @PrimaMente!

After 120 years of progress, we still don't have a good understanding of what drives Alzheimer's Disease. We believe that a global, coordinated effort to develop and test new hypotheses using frontier tools has a real chance to change that.

Today we're launching the Alzheimer's Translation Challenge, a global AI competition to discover new therapeutic hypotheses for Alzheimer's disease.

At its core is a new 150M cell atlas of neurons, astrocytes, and microglia across different genetic backgrounds under combinatorial perturbations, with multi-modal readouts.

The data will be made available through the AD workbench, Hugging Face and Prima Mente's modeling platform.

The challenge is led by @PrimaMente and @AlzData in collaboration with @nvidia, @huggingface, @nebiusai, Talisman Therapeutics, @UltimaGenomics, @cellanome, @PrimeIntellect, @GoodfireAI, and @boltz_bio.

In 1906, Alois Alzheimer described a new disease of the brain. After 120 years of progress, we still don't fully understand what drives this disease. We want to change that.

Participants will use the dataset to build AI models, assessed through a series of evals from general benchmarks to more complex tasks.

Top teams will be shortlisted, and their experimental hypotheses tested in Prima Mente's wet lab. Teams win when their hypotheses demonstrate the predicted molecular and functional effect.

🧵 Biohub, @ENERGY, @NIH, and new funding partners today announced a major expansion of an international effort to generate and make accessible the data enabling predictive AI models of biology. Together, the organizations are investing $1.8 billion in funding, data, computation and new measurement technology - largest coordinated commitment to map the biology of the cell to power development of digital models of biology. bit.ly/3UhjoxD

imagine tweeting this 10 hours before AI solves mathematics

We're excited to announce our partnership with @PrimeIntellect to allow anyone to train browser agents.

General-purpose models aren't optimized for your browser workflows, BrowserEnv lets you train one that is.

Checkout browserenv.com and train your own custom model in a few hours.

new blog post: why taking inference scaling seriously leads to swarms.

contrary to popular belief, I don't think multi-agent systems are a new scaling axis. they're just structured inference scaling.

hallerite.com/posts/on-the-n…

The Prime Sandboxes team is looking for Distributed Engineering talent.

Fault-tolerant systems. Deep Linux. Virtualization. Performance Engineering. Go + Rust.

High agency and in SF. DM me your GitHub, highlighting something exceptional.

job desc: primeintellect.ai/careers/26c41d…

Without diminishing the contributions of Peter, Georg, or Karl, I find the omission of Ed Boyden from the optogenetics prize to be highly unreasonable.

RL needs more long-horizon tasks beyond math and coding. @niloofar_mire's team at CMU built an environment for drug design.

SMDD-Bench comprises 502 small-molecule design tasks with RDKit, ADMET-AI and Boltz-2 in the loop. The challenges go beyond chemistry: long-horizon planning, exploration, and learning from imperfect feedback are also open problems for RL/ML!

SMDD-Bench is available in our Environments Hub, ready to train with prime-rl. Thank you for sharing with the community!

copernican trauma is stored in the n-body

Some personal news: I'm joining @PrimeIntellect as Head of Scientific AI!

A billion copies of the same very smart, agreeable mind will not be a scientifically fruitful world.

Scientific superintelligence should be weird and definitionally nonconsensus. It should be highly plural, consisting of many minds, human and artificial, that see the world differently and hold each other accountable, with human judgement still playing the central role in setting the trajectory.

Science is about asking the right questions and reducing uncertainty. Professionalizing science made discovery a viable career, but it also introduced incentives to work on what is fundable and publishable. Neither is the same thing as being true or impactful (reward hacking exists in the real world too!). We are very good at accumulating incremental knowledge and rather bad at stewarding unorthodox ideas through infancy, even though these are often the ideas that open entirely new fields of inquiry.

If we aren't careful, scientific AI could reproduce these same incentives at a much larger scale. If everyone relies on the same few large models, we risk inheriting the same priors and blind spots.

Open tools let us shape our own intelligence and take it in directions a frontier lab might never prioritize. This matters for scientific freedom and is central to Prime’s vision.

If you are a scientist, staying at the frontier will mean training your own models on your own research and data. I'll work closely with scientific teams across domains to help them build scientific agents by turning their research problems into environments where agents can learn. This means closing the loop from hypothesis to experiment to validated result, with agents learning from each cycle. Teams will be able to own their intelligence, using tools previously available only to frontier labs, and point it at the questions they care about most, especially those that others might not be asking.

I've had the good fortune of seeing how science operates across academia, industry, government, startups, and venture. I worked on experimental therapeutics for cancer in academia before moving to the National Institutes of Health to work on DNA damage and repair. I cofounded Molecule, where we funded dozens of research projects to speed the translation of early-stage therapeutics to market, focusing on areas traditionally ignored by pharma. Most recently, I started Triplicate, a small fund backing teams and tools accelerating scientific research. Many of the problems I care about most are increasingly soluble with advances in AI, and I hope to bring what I've learned about the bottlenecks in translating research to radically accelerate science in all fields.

I share Alexander Grothendieck's ethos of discovery and believe Prime does too: "Discovery is a child's privilege. I mean the small child, the child who is not afraid to be wrong, to look silly, to not be serious, and to act differently from everyone else."

The goal of progress is not more intelligence for its own sake. In my view, it is to better understand reality, and, in doing so, give more people, and more kinds of minds, the chance to flourish.

If you have a research problem where today's tools fall short, my DMs are open. If our approach resonates and you want to build this with us, we're hiring!

Prime continues to cook with the most efficient and cost effective microVM sandboxes on the market

"How unsafe is reality?"

one of my favorite essays on the challenges associated with safety and alignment from @michael_nielsen

michaelnotebook.com/whichfuture/in…

worth taking seriously the possibility that vedanta and neoplatonism offer more fertile starting points for science’s deepest questions than our current frameworks