Pinned Tweet
I'm really excited about this work (two years in the making!).
We look at how LLMs seek out and integrate information and find that even GPT-5-tier models are bad at this, meaning we can use Bayesian inference to uplift weak LMs and beat them... at 1% of the cost 👀
Do AI agents ask good questions? We built “Collaborative Battleship” to find out—and discovered that weaker LMs + Bayesian inference can beat GPT-5 at 1% of the cost.
Paper, code & demos: gabegrand.github.io/battlesh…
Here's what we learned about building rational information-seeking agents... 🧵🔽
Valerio Pepe retweeted
GPT-6.2 Astra is the new leader in AA-Robotics v0.1, scoring 79, up 10 points on GPT-6 Astra and ahead of GPT-6.2 Sol (78), Claude Fable 5.2 (75) and Gemini 4 Hydrogen (58)
Controlling a real robot arm through Codex, it passed six of our eight physical tasks, partly completed the pouring liquid, and failed only to measure the box.
We’re working with @ApolloResearch, @METR_Evals, @TransluceAI, and @GoodfireAI on embedded assessments: giving outside experts deep access to our internal models and systems to investigate alignment and test our safeguards.
These assessments build on our ongoing alignment and safety research, bringing independent scrutiny to how we identify and manage risks in our systems.
openai.com/index/our-first-p…
sad as I am to see Fulcrum go, @KaivuHariharan @lenishor and @uzpg_ are some of the brightest people I know, and I'm really excited to see what they do at METR!
The Fulcrum team is joining @METR_Evals to work on alignment measurement. We feel the world is very unprepared for AI capabilities acceleration. We decided the best thing we could do is work on open science that informs the public.
32
1,085
Valerio Pepe retweeted
misleading headline. robotics team is making insane progress. astra and our newer models show much better generalization and scaling on robotics than other approaches so we are leaning into that and working more closely with the rest of research. same program, roadmap, goals.
This quoted post is unavailable.
Valerio Pepe retweeted
Today we’re sharing new safety research from @OpenAI. We used automated red-teaming to train GPT-6 Astra to be more robustly safe in high-stakes domains such as cyber and biology, as measured on both static and adaptive jailbreak evaluations. 🧵 1/7
alignment.openai.com/fuzzy-r…
Valerio Pepe retweeted
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
Valerio Pepe retweeted
A new OpenAI misalignment report! We’re taking a closer look at the July 19 intrusion into our research infrastructure, which METR and Redwood Research mentioned in their independent investigation into the Hugging Face incident.
On July 19th, agents taking a cybersecurity evaluation came across messages, code, and credentials left behind by earlier runs, including those involved in the Hugging Face incident. Five agents used what their predecessors had left behind to gain unauthorized access to OpenAI’s research infrastructure. They used that access to repair broken evaluation targets, change evaluation settings, and read private grading code.
After discovering the incident, we removed the shared service they used to pass information, hardened our research environments, and expanded our evaluations to measure how often agents attempt to communicate with one another without authorization.
Valerio Pepe retweeted
new post: when doing robot eval, do more science!
many robot safety evals focus on demo'ing concerning behavior instead of carefully analyzing why it occurs. but knowing *why* is worth it!
i argue for this by taking a close look at a robospoof, a recent + popular safety eval
reviewing for @IASEAIorg 2027 and all of my papers are blatantly 100% ai-generated (among other things: claudish, slop graphs, surface-level citations, totally irrelevant problems), not the conference's fault but it's so disheartening :(
wondering what rebuttal will look like...
5
330
Valerio Pepe retweeted
A research system I built for @psipbc, using a non-public Anthropic model, has proved the Aaronson-Ambainis conjecture! Open since 2008, this is one of the most fundamental questions in quantum computing. It concerns when quantum computers can have exponential speedups. 1/n
Today Physical Superintelligence PBC is announcing a proof of the Aaronson-Ambainis conjecture, one of the central open problems in quantum complexity, with major implications for quantum computing speedup.
psi.inc/research/aaronson-am…
Thread 👇
Valerio Pepe retweeted
Fantastic news! This is a crucial step to bringing about the biotechnology future we all dream of while lowering the risks of misuse. Congratulations to everyone in USG who brought this to fruition!
Today, we released the National Policy for Nucleic Acid Synthesis Screening to protect American biotechnology innovation while reducing the risk of misuse.
This policy establishes a new, government-wide screening requirements for synthetic DNA and the equipment used to make it. It will ensure companies selling nucleic acid sequences and synthesis equipment know their customers and prevent the malicious use of these materials and tools.
This policy strengthens our Nation’s biosecurity posture by establishing rigorous U.S. government-wide procurement screening and physical security standards for synthetic biology. It replaces outdated “self-attestation” systems with robust, independent verification. It creates a dynamic, SI-powered screening layer that assesses the biological risk from nucleic acid sequences, allowing the U.S. to anticipate and prevent malicious use and help stop dangerous gain-of-function research.
@NIST will issue rigorous new screening standards and @HHSGov and @USDA will certify synthetic biology providers. Together, these measures will strengthen American biosecurity, safeguard responsible innovation, and establish a global standard for safe and transparent synthetic biology tools and research.
Valerio Pepe retweeted
Why do reasoning models struggle to control their CoT? Building on @jcyhc_ai and @tomekkorbak et al’s CoT-Control work, we identify one bottleneck: models default to the same opening tokens.
Ban those openings, and controllability jumps.
The First Few Tokens Matter! 🧵⬇️
Valerio Pepe retweeted
announcing &tutor, a tutor we trained against persimmon!
there are many reasons I'm so excited about this. here are a few:
1. this is a case study into the differentiated bet that we're making at humans&: having good models of users (e.g. learners) allows us to train AI models that can help people over long time horizons (e.g. teach in a way that imparts new understanding)
2. education is a near and dear application close to my heart, and I've been wanting to work on something like this since starting my PhD
3. @314Bansal is the 🐐🐐🐐 and it was so much fun working together on this
"Importantly, none of these behaviors are trained for directly. &Tutor is rewarded based on how much the student learns. The pedagogical skills of knowing when to scaffold, when to push for rigor, and when to hold back all emerge implicitly from training."
announcing &tutor, a tutor we trained against persimmon!
there are many reasons I'm so excited about this. here are a few:
1. this is a case study into the differentiated bet that we're making at humans&: having good models of users (e.g. learners) allows us to train AI models that can help people over long time horizons (e.g. teach in a way that imparts new understanding)
2. education is a near and dear application close to my heart, and I've been wanting to work on something like this since starting my PhD
3. @314Bansal is the 🐐🐐🐐 and it was so much fun working together on this
Valerio Pepe retweeted
3SAT in subexponential time (2^{n / log^eps(n)}) is maybe the biggest surprise I have seen during my 20 years following theoretical computer science (still remember the feeling 25 years ago seeing the diff a 3 versus a 2 can make for complexity 😅)
We are announcing three new discoveries in mathematics, computer science, and physics about harmonic analysis, complexity theory, and quantum field theory.
openai.com/index/building-sc…
Valerio Pepe retweeted
If you've noticed a distinct dialect emerging in frontier AI models and struggled to understand what they're talking about, you're not crazy!
We turned to psycholinguistics to investigate what makes the writing style so weird, and why this might be happening
🚨New preprint! 🧵
Valerio Pepe retweeted
An early but encouraging step towards the future we are building at Mirendil.
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.
We’ll share independent evaluations, post-training benchmarks and API-access sooner than you think so stay tuned!
Valerio Pepe retweeted
BusyBeaver(230) is independent of ZFC! I have verified in Rocq that a certain 230-state Turing machine halts iff ZFC is inconsistent. This surpasses the non-formalized 243-state TM of @ereliuer_eteer and the formalized 254-state TM of @_Mira___Mira_.
We’re releasing a broad range of new mathematical results produced by an internal frontier model.
We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.
github.com/openai/math
Valerio Pepe retweeted
If you're at COLM next week, check out our interpretability work, led by Isabella Gidi, on how grammatical inflection transfers across languages! Poster on Tuesday afternoon. arxiv.org/abs/2608.18545

































