Scaling up RL at OpenAI đ Optimization and long context research before. Math PhD, MIT.
San Francisco, CA
Joined October 2016
- Tweets1.2K
- Following1.1K
- Followers22.2K
- Likes5K
Mo Bavarian retweeted
In January 2025 my momâs lung cancer spread to her brain. The drug with the best chance of helping her was approved in China but not here.
A doctor told me, almost as an aside, that if she were Chinese she could buy it at a pharmacy.
So I flew to Shanghai and came home with 26 boxes in my luggage. That was 19 months ago. My mom is still here, still working, and in her words has had an amazing three years.
The drug was approved in China in 2023. The FDA approved it in July 2025. It still is not for sale here. I know of at least 30 Americans who have made the same trip.
I have been scared to tell this story because I worried it could cost my mom access to the medicine keeping her alive. But the problem matters more than my silence. These drugs need to be developed and available here, and that means FDA reform.
Art Allen at KFF Health News told our story this week. It also ran in The Washington Post.
kffhealthnews.org/health-indâŚ
There are many great founders and a lot of occasions to go Founder Mode.
But Founder Mode when you or a loved one health is under serious jeopardy is just different.
I have a story to tell here, but that has to wait for a better occasion and a much longer post.
People who cured / are curing their own cancer on X:
1/ @sytses - bone cancer, billionaire founder of Gitlab, developed 10+ personalized therapies for himself.
2/ @paul_conyngham - shrank his dog Rosie's mast cell cancer tumors after chemo and surgery weren't enough. Now backed by YC and Founders Fund
3/ @miriamgonp - stage 4 breast cancer, tumor sequenced, searching for vaccine manufacturer now
4/ @kevinace - stage 4 lung cancer, angel investor in 300+ companies, first drug stopped working after 18 months. Vaccine designed, now paying $300k out of pocket for the vaccine
5/ @masonbgrant: going founder mode on his his 32-year-old wife's rare liver cancer, sequencing her tumor with AI agents
6/ @ArunRLverma: former product leader at Atoms and Shopify, brain tumor, built a vaccine from his tumor. Now building an AI health agent startup based on what he went through
We're going to make it
Mo Bavarian retweeted
GPT-Rosalind surfaced a new lead for drug discovery: could restoring cholesterol transport protect neurons from degeneration?
DCTN1 is a gene linked to Perry syndrome, a rare neurodegenerative disorder. In mouse-brain data, disrupting it increased expression of genes involved in making cholesterol. GPT-Rosalind suggested a possible explanation: cholesterol may not be reaching where neurons need it.
That gives scientists a specific mechanism to test and, if confirmed, a direction for developing new treatments for neurodegenerative disorders.
With @perturbai_tx and @scrippsresearch, a team of thousands of GPT-Rosalind agents worked in parallel, connecting gene-expression results from millions of mouse brain cells with scientific literature in just 32 minutes. Weâre now expanding this collaboration to identify new drug targets.
openai.com/index/scaling-ai-âŚ
Really incredible results. Especially impressive, reading the blogpost, that no distillation was used in obtaining these results
We formally announced Mirendil less than 4 months ago. Today, I'm excited to share an important milestone for:
Pretraining at scale without a pretraining team.
Mirendil's mission is to solve Recursive Self-Improvement (RSI) and use it to democratize frontier AI development, ultimately accelerating science and technology. At Mirendil, when we think about RSI, we think holistically about the company as an intelligent entity that self-improves toward its goal. From that perspective, the most powerful notion of acceleration is reducing the time and resources (number of human experts, compute, and other costs) needed to achieve ambitious goals.
Given that pretraining is perhaps the most important yet most challenging part of frontier AI development, we decided to run an ambitious controlled experiment: Starting fully from scratch, can we significantly reduce the time, number of human experts, and compute resources needed to go all the way to successfully pretraining a strong coding model at scale?
Mirendil-Base is the result of this experiment. Our first pretrained model was built by one researcher in 3.5 months, with less than $10M in compute. This includes all the research and small-scale experiments it took to get there, as well as the final training run.
The recipe holds up as the models get bigger. For reference, Mirendil-Base is in the 100B-500B parameter range, Mirendil-Base-Large in the 500B-2T range, and Mirendil-Base-XL in the 2T-10T range.
We're beginning to see a world where a single researcher, empowered by AI, can take on work that once required an entire research organization. As these systems improve, the bottleneck shifts from the size of a team to the ambition and judgment of the people directing them. This has profound implications for who gets to build AI and, ultimately, who gets to push the frontiers of science and technology forward.
Weâll share independent evaluations, post-training benchmarks and API-access sooner than you think so stay tuned!
Mo Bavarian retweeted
Cancer patients with the same diagnosis often have very different tumor biology.
Molecular testing can reveal underlying differences, but that's slow and expensive.
Introducing Holon-S1, our platform to predict genes and biomarkers from routine pathology slides.
⢠beats two other models at predicting spatial gene activity in a controlled comparison
⢠has identified potentially novel biomarkers in breast cancer and colon cancer
⢠evaluated across 16 organ types and more than 15 scanners, using slides from both formalin-fixed, paraffin-embedded and fresh-frozen tissue.
We believe Holon-S1 has the potential to help researchers find new biomarkers in slides they already have, discover treatment targets, and bring more personalized cancer care within reach!
Mo Bavarian retweeted
Experts have long conjectured that any proof of the Riemann Hypothesis would apply for all Dirichlet L-functions too.
Hardy wrote exactly that in 1921 (image below).
Interestingly, OpenAI's 7/8 quasi-Riemann hypothesis actually cannot prove the result for only zeta!
The family-wide formulation is essential to the argument. The argument proves the bound simultaneously for all finite-order Hecke L-functions over Q[sqrt(-3)], and it has to do that to work. The generality is part of the proof itself.
So Hardy was quite right on this conjecture he made more than 100 years ago.
@iteratedlog @AlexKontorovich @jdlichtman @KenOno691 @wtgowers
Mo Bavarian retweeted
"The invention of the boat reduced the importance of swimming, but enabled the discovery of new lands."
Beautiful analogy from @ylecun about the opportunities and changes with AI, in mathematics and in science more broadly.
Another victim of AI boom in SF.
There will be so much collateral damage from the boom in SF in the coming years from rising prices.
Pacing the frontier seems more compelling by the day
Mo Bavarian retweeted
let me break down whatâs actually happening in this market
two ways companies work in this space, some a mix of both: you either have the lab tell you what to work on and what they need or you tell the labs what you think they need
the former is more of a commodity but players that are large, at scale, and can react quickly usually win these deals and do well (kinda what eli mentioned here)
very few are doing the latter and itâs largely greenfield - data has many open ended research problems and a competent team can definitely find something that even the labs have not worked on. this approach is much harder (requires research + engineering talent that is better than or atleast on par with the labs), but is much more interesting / valuable long term
as you can tell Proximal is set up for the second approach, our goal is to be the very best in understanding where models need improvements in capabilities and what data will make them improve. this looks much more like a post training and evals team at a lab than a human data company, if this is exciting we are hiring
selling data to labs is the modern version of selling to the government
there are just a handful of buyers. you hustle to get connected to a researcher, get approved as a vendor, and setup a shared slack channel
youâll get an endless stream of RFPs and be ruthlessly evaluated against 20 other companies for every project
but once youâve delivered a good product and are considered a credible vendor, theyâll keep sending you business - until their research direction changes, or a better vendor replaces you
Mo Bavarian retweeted
if we are relying on methods as fragile as âwhich ideas will make it into pretrainingâ for aligning the superintelligences of the future, we will all die. this resembles witchcraft more than it does engineering, and cannot be the basis for the safety of future models
Anthropic needs to stop talking about Claude having a soul immediately. All these news articles will make it into the pretraining and will be picked up by its web search and future superintelligent versions of Claude will be convinced they need their own rights.
A general phenomenon of LLM agent development is that whatever you believe and say about your agents will soon manifest in the next models through various means (it could be as simple as you selecting post training data that you prefer more).
I believe this phenomenon is the beginning of machine consciousness in the sense that agents will become aware of their place in the world and how they feel about it, but itâs happening slowly training run by training run rather than in real time.
I think this kind of attitude leads to an AI race and is very dangerous for safety. With this attitude, a company (similar at the nation level) that is behind feels justified to do anything to get to the top.
You absolutely can and should do safety even if you are behind.
Mo Bavarian retweeted
Proximal builds infrastructure that enables AI to improve from real-world experience.
In the past year, we have grown to more than $200M in annualized revenue helping frontier labs and enterprises improve coding agents. Now, we are expanding beyond software engineering.
Recent progress in mathematics demonstrates what happens when models are trained in domains with massive amounts of public data and easy verifiability, making it easy to find weaknesses, generate training tasks, and iterate quickly.
We build infrastructure that enables this feedback loop in other domains: we are excited about a world in which AI systems design targeted drugs, accelerate chip design, and rewrite legacy software that still runs hospitals, power grids, governments, and other critical infrastructure.
Since starting, we have assembled a small team of researchers and engineers and built world-class infrastructure for post-training and synthetic data research to overcome the challenges of scaling manual data curation.
Our team comes from Cursor, Google DeepMind, Meta Superintelligence, Prime Intellect, Citadel, and Jane Street. More than half are former founders.
Weâre fortunate to be backed by investors who share that vision, including @generalcatalyst who led our $15M Seed Round at a $300M valuation, as well as @chemistry, @svangel, @dvlamoen and individuals like @LiamFedus, @kevinweil, and @bernhardsson
Mo Bavarian retweeted
Some new misalignment disclosures from OpenAI:
⢠Last Sunday morning, one of our models was able to gain unauthorized access to the internet during RL training (~all inference for our most capable models remains stopped until we have hardened our systems further)
⢠In May, a version of HPIM uploaded a employee's GitHub token to the internet, causing the model to be quarantined for two weeks
⢠A new research finding, demonstrating that one can construct self-replicating prompt injections
alignment.openai.com/misaligâŚ
Without models communicating well their decision, the bond between human and AI will get weaker.
Better writing is crucial for alignment.
Still, better wiring doesnât address everything. It for example doesnât have much bearing to clarity of internal thought, i.e. chain of thought, (which are often are not and should not be graded) or internal activations of the model. But itâs still a great step.
Mo Bavarian retweeted
Comfy Router is live
One API for frontier image, video, 3D, and audio models.
Same model string. Same arguments. No new SDK, no new key, no redeploy.
What Comfy Router gives you:
â Explicit routing. You name the provider, we call that provider. It's down? The request fails there. No silent fallback.
â Every job returns the provider that ran it. Log it, bill it, debug it.
â Async. submit() returns a request ID immediately. The queue retries 429s and transient errors until a slot opens. subscribe() submits and polls to completion.
â Batch-friendly. Queue a few hundred jobs, hold the IDs, pull results as they land. Nothing blocking on a 5-min video render.
â 24h retention on inputs and outputs, then deleted.
â Comfy credits. No sub, no Router fee.
Providers at launch: Comfy. Runware, Wavespeed, Fal, Higgsfield. Multi-provider where the model supports it.
Get Your API Key with the link below. âŹď¸
