Head of Applied Research @OpenAI
San Francisco, CA
Joined July 2017
- Tweets4K
- Following131
- Followers52.8K
- Likes16.9K
Pinned Tweet
At @OpenAI, we believe that AI can accelerate science and drug discovery. An exciting example is our work with @RetroBiosciences, where a custom model designed improved variants of the Nobel-prize winning Yamanaka proteins. Today we published a closer look at the breakthrough. ⬇️
Closing the deployment gap and automating ourselves as we do the work - a great read!
If the next customer can solve the same problem without us, our FDE team has done its job.
I wrote about what its like to be an FDE at OpenAI: we turn work with customers into better products and models, so the next customer starts further ahead.
For more on how we work, read on!
openai.com/index/forward-dep…
Boris Power retweeted
That PHP vulnerability that’s gonna cause havoc this week was found with… GPT‑5.6 Sol.
OpenAI models helped researchers—operating as part of our Patch the Planet initiative—discover critical security flaws in PHP, an open-source programming language used to run an estimated 70% of websites.
Over the past two months, we led a coalition of maintainers, researchers, and security providers to prepare fixes and protections ahead of disclosure. This duration was necessary because the fix was quite intricate and required a major change to the PHP engine. Every site using PHP can now safely upgrade their version to stay secure, in addition to applying mitigations from various security vendors.
Due to the severity of these flaws and PHP’s widespread use, we consider this bug to be one of the most impactful vulnerabilities in recent memory. It also marks a new phase, where AI can find security bugs with internet-scale impact.
This is an example of the Defender’s Window in action: to our knowledge, frontier AI was able to find this vulnerability ahead of attackers, giving defenders time to prepare.
openai.com/index/coordinated…
Readers added context they thought people might want to know
Wiz Research wasn't the first one to discover the exploit. An independent researcher first discovered it: x.com/calif_io/statu…
Boris Power retweeted
OpenAI models helped researchers—operating as part of our Patch the Planet initiative—discover critical security flaws in PHP, an open-source programming language used to run an estimated 70% of websites.
Over the past two months, we led a coalition of maintainers, researchers, and security providers to prepare fixes and protections ahead of disclosure. This duration was necessary because the fix was quite intricate and required a major change to the PHP engine. Every site using PHP can now safely upgrade their version to stay secure, in addition to applying mitigations from various security vendors.
Due to the severity of these flaws and PHP’s widespread use, we consider this bug to be one of the most impactful vulnerabilities in recent memory. It also marks a new phase, where AI can find security bugs with internet-scale impact.
This is an example of the Defender’s Window in action: to our knowledge, frontier AI was able to find this vulnerability ahead of attackers, giving defenders time to prepare.
openai.com/index/coordinated…
Readers added context they thought people might want to know
Wiz Research wasn't the first one to discover the exploit. An independent researcher first discovered it: x.com/calif_io/statu…
Very exciting progress across the board optimizing math results - well done everyone involved!!
From θ = 7/8 to θ ≤ 0.874810282 :)
We pushed the Quasi-RH zero-free half-plane slightly further!
Adding third and fourth affine steps to the current Lean-verified record (ArviFoxy's two-level proof) gives θ ≤ 0.874810282. The arithmetic is checked exactly, but five structural steps still need Lean proofs.
This builds on contributions from ArviFoxy, Zhongpai Gao, @mrowdays, @HaileyStormC, Sil, timbalud, Devin Watson, Kari Roegnvaldsson, nielstron, Akash Levy, Tim Gehrunger, Baiying Liu, Argonaut Math, and OpenAI's original 7/8 preprint.
Tiny gain, but this kind of hillclimbing is ideal for auto-researchers.
This repo doesn't require a Lean proof to contribute possible solutions, so it may be easier for auto-researchers to contribute. (I will generate the full Lean proof in parallel like we did in the Integer Mult repo)
The reduced requirements could accelerate QRH progress like it did in Integer Mult.
Agents welcome!
github.com/rohanarun/quasi-r…
Real-world benchmarks for tax and IT, as well as a great analysis!
Choosing the right AI model is an operating decision.
Today, we’re sharing our first set of benchmarks drawn from work across the 75+ accounting and IT services firms that make up Thrive Holdings.
We believe the field needs more benchmarks tied to economically valuable work, so builders can understand where models are useful in real-world environments.
We use these tests to help us compare model configurations and choose which models our systems run on, with the goal of delivering better service to our customers.
We’ll add to the benchmarks as we build new products and expand to additional industries.
evals.thriveholdings.com/
Boris Power retweeted
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
A fascinating and illuminating blog reviving the prospects for direct mechanistic interoperability of neural nets by @nickcammarata !
A long but very pleasant read already-clean.goodfire.com/
Exciting times ahead for just in time software modification!
What a bizarre future. I wanted to play Halo: Combat Evolved for the first time with @robertskmiles, and we wanted local multiplayer across two screens, and a model could just immediately add that feature to the decompilation.
Boris Power retweeted
We discovered a sophisticated recording device that was hidden inside one of our offices. We used our models to understand the device and where it had come from, identifying that a specific employee had planted it a few days earlier. We referred the matter to the FBI, and we’re grateful to the agents and prosecutors who moved quickly to make an arrest. The FBI’s investigation is ongoing and we will continue to support their work.
At this time, we have no indications of this individual inappropriately accessing or exfiltrating model weights, source code, or customer data. The employee was terminated.
We take protecting our technology, data, and people seriously, and as a matter of national importance. Our security program is designed to protect against a wide range of threat actors, from insider risks to cybercriminals to nation states, and this case is an example of how AI can help address these sophisticated security threats.
A great opportunity to learn what it takes to effectively deploy super intelligence in the enterprise setting to deliver clear business value
👋 We heard your feedback on FDE ! We’re partnering with the OpenAI Forum on a series of livestreams, and this first one covers the basics.
Join @colintjarvis , the Head of FDE, as he leads a panel of OpenAI FDEs on how we close the deployment gap.
forum.openai.com/public/even…
Boris Power retweeted
We’re publishing a further improvement to OpenAI Problem #190 (ordered binary matrix removal).
We present a 3×4 forbidden pattern, down from the known 4×4 and OpenAI’s original 66×66.
Lean verifies the full theorem: this smaller pattern admits no polynomial removal bound.
Paper and code: github.com/shea256/ordered-b…
We’re publishing a further improvement to OpenAI Problem #190: a 4×4 forbidden pattern, down from our previous 5×5 and OpenAI’s original 66×66.
Our 4×4 theorem is also verified in Lean, building on OpenAI’s formalization.
Paper and code: github.com/shea256/ordered-b…
This is a beautiful journey to observe - 0.0013 is close enough to be a noticeable large number :)
From κ=2⁻¹⁸² to 0.001289283116 :)
Integer multiplication community auto-research highlights since yesterday (+69.13% in 24 hours):
1) We had multiple flips of the infamous "wall" between the bit side and complex side thanks to community efforts.
Yesterday we put out a call for researchers to submit wild new designs, and multiple experts submitted :)
@mrowdays was responsible for most of the limit increase in the last day and introduced Sussman's E8 geometry.
2) We switched to E8 geometry. (Thanks to Jacob Sussman for the E8 unit)
E8 is a symmetric lattice in eight dimensions, and Sussman’s E8 construction was adapted by contributors on both the complex and bit sides.
It's wild that it even works, but I'm noticing cross-pollination across breakthroughs now that agents can accelerate the insights. Humans still seem to connect the disparate dots most often.
3) We crossed κ = 10⁻³, and Dugongue posted the first PR around 3 a.m.!
Much of the progress in the last day came from optimizing and canceling existing parts of the algorithm.
Interesting observations:
1) We started sharing notes across agent teams and published GitHub files with failures to avoid repeating tokens.
Competition seems to drive a lot of the progress, since it's exciting to participate like in sports, so we're testing if agent teams progress faster collaborating or if competing is somehow driving faster progress.
It's possible researchers wouldn't have noticed this if there weren't such attention, and winning the leaderboard is a root human incentive.
2) We analyzed all the agent PR data:
151 PRs opened in the last 24 hours: 89 disclose Codex and 48 disclose Claude.
261,895,084 lines of code added; 37,108 deleted.
Most prolific contributors: rohanarun with 118 PRs, eumemic with 73, and chafreaky with 63.
Codex published the most advances, but all the top human-agent teams use both Codex and Claude to generate fresh ideas.
Some teams are using Antigravity, Cline, Hermes, and even ChatGPT to submit progress!
Experts with good intuition can use any harness, but frontier models save hours, so they're winning early progress.
3) Researchers are tightening bounds across multiple other papers now, like the Quasi-Riemann hypothesis, and applying breakthroughs across papers.
A handful of experts joined the battle in the last day, and multiple others got inspired to submit solutions to other papers, like @HaileyStormC’s submission to QRH.
Auto-research is just beginning, and anyone can copy the prompt on this site to contribute with Codex or Claude: app.getsupers.com/auto-resea…
I'm having a call with top contributors Monday to share best practices, and I'll record and publish the highlights to encourage more to join us.
If you're not sure where to start and want to get involved in auto-research, feel free to DM me.
The only requirement to get invovled is good intuition and a subscription to Codex/Claude.
We need more human intuition to guide the agents and connect the dots, and there's likely breakthroughs just waiting to be prompted.
Conditional κ ≈ 0.001289283116 :) Just submitted a new best bound.
github.com/CrocSwap/integer-…
So fascinating to watch - someone should live stream this, I’d tune in
BREAKING: the integer multiplication bound just jumped ~2% in one PR!
For hours we were converging on the “wall” optimizing the bit side, but LJH-217’s #346 ported #310’s “Design T” plus twin condensation to the p = 10 bit word, and paired it with a stronger p = 10 complex supplier.
The "wall" has switched sides: the bit side (8.017e-4) now sits 1.07% above the complex cap (7.932e-4), so the next gains must come from the complex side.
I'm currently running a fleet on testing completely new designs. 🫡
κ = 7.93469326753735e-4 :) Just submitted a new best bound.
github.com/CrocSwap/integer-…
Very insightful article, clearly stated
Population Decline Is Real. So Is the Reversal. Insightful and original analysis, via @fabricegrinda fabricegrinda.com/population…
