

✓Researching reasoning @OpenAI | Co-created Libratus/Pluribus superhuman poker AIs, CICERO Diplomacy AI, and OpenAI o-series 🍓 reasoning models


I used GPT-6.2 Astra to decipher Raymundo Machaín's letter to Juan Silvano Godoi, written in 1878 - which, to my knowledge, has never been deciphered before. Significantly, GPT-6.2 Astra deciphered this letter without having access to any key. Godoi, a Paraguayan politician, had previously fled the country after being involved in a conspiracy to assassinate Paraguay's previous president, J.B. Gill. Later in 1878, Godoi would buy a ship and 700 guns and sail up the Paraguay river in a bid to overthrow the country's new president, Candido Bareiro. This letter, which was written to Godoi by his accomplice, Raymundo Machaín, relates to the shipment of guns for this expedition. Only a few sensitive phrases in this letter were written in a cipher (the rest is just written in Spanish). Thanks to GPT-6.2 Astra, we now know that: 1. The man working with Godoi in Montevideo to equip the ship (also referred to in the letter as "the Englishman") is "Brown" - likely, Eduardo Brown, who, according to other correspondence, had been working with Godoi on equipping the expedition. 2. The man who arrived in Montevideo and offered Machaín money was "Bravo" - likely Francisco Javier Bravo. The cipher reveals that Bravo showed Machaín the "complete and accurate list" (presumably of the weapons being purchased) and offered Machaín money to "dispatch the [ship] and pay the crew". 3. Machaín says that "Folle", a shipping firm, has refused to handle the shipment of guns without being presented "formal instructions", which he needs from Godoi. 4. Brown apparently doesn't trust Bravo: "Brown is afraid Bravo will play us a dirty trick", the cipher reads. And: "Brown wants everything put in his name so that he has the right to reclaim it if they try to take it away." How Astra solved this cipher: - First, it just guessed at three words in the letter based on the surrounding context: "Machaín gave us clues in the way he switched between plain Spanish and cipher. He wrote “me mostró una” (“he showed me a…”) and then five encrypted signs. Lista (“list”) was a guess. Elsewhere, “sin antes recibir” (“without first receiving…”) was followed by thirteen signs, and Machaín used the word instrucciones in clear Spanish later on the same page. A third guess was factura (“invoice”)." - Then, the model figured out the key: moving the written characters, in order, and up and down the Spanish alphabet by −7, −11, +13, +8, +4, fit all of the above words perfectly. - The model had a "breakthrough" when it realized that the cipher had restarted the key at every word, even when a few words were encrypted together without spaces. This let GPT-6.2 Astra completely decipher the letter.

GPT-6.2 Astra is amazing. For $20/month, you can now access a model that surpasses top human intelligence in some domains. I'm most excited to see what scientific discoveries people are able to make using it.

Final standings in the charity poker tournament after @sequoia's poker bot tournament: 1. Noam Brown (@polynoamial) of both poker bot and AI fame 2. @MichaelBarrySF from @factory (which I'm on the board of!) 3. @JasonKoon who is a professional player and also a bot pioneer Phenomenal event masterminded by @abhishekm1636 The young MIT/Harvard/CMU/Caltech/Stanford kids are still cracked!

Bragging rights for best poker player in AI were on the line at the No Limit Charity Cup in SF. An honor to have poker legends @JasonKoon and @Mister_Keating at the table. Across from them: founders and competitors from @Etched, @Factory, @appliedcompute, @DoorDash, @fal, @trajectorylabs, @mercor, @getserval, @neros_tech, @hyperbolic_labs, @p0, @SableHQ, @simile_ai, @conwayinc, @OpenAI and @citsecurities. Poetic that @polynoamial, the man behind the first superhuman pokerbots, took home the trophy. All proceeds to @khanacademy. Thanks to all our competitors and partners. See you next year.

Honestly quite impressed with the methodology on this benchmark. @dwarkesh_sp want a job designing evals at @OpenAI?

Every winter, @MIT has run a 1-month class (6.9630) in which students make poker AIs and compete against each other for cash prizes. Coding agents have accelerated that process so much that it's now possible to build a poker bot in an *hour*. Also, I'll be at the finals!

Only 2.5 years ago, AI labs would highlight their frontier LLM performance on grade school math benchmarks (GSM8K). I can't wait to see where we'll be in another 2.5 years.

Last month a large team of agents used over 100 billion tokens to resolve the Navier-Stokes Millennium Prize problem. Now we're using a similar approach for red-teaming to find and fix infra vulnerabilities.

This phishing email made it past Gmail’s filters, but my dot caught it and warned me.

LLMs have crossed an important threshold: surpassing top human experts on some research problems. Capabilities have been steadily improving for years, but in scientific discovery there is a big difference between a model just below top human performance and one just above it. That helps explain why the surge in math results feels so sudden. LLM capabilities remain jagged, and there are many ways they are still worse than humans. But as models continue to improve, I expect breakthroughs in other domains to follow, leading to a new era of scientific discovery.

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


@polynoamial Results are even more impressive at max! 6.1 sol outperforms Opus 5.5 while using 4x fewer tokens.

I’m happy that @OpenAI presents model evals this way. We should measure intelligence as a function of cost.

Was testing out dots over the weekend and it figured out how to save me ~$500/yr in recurring charges. Most impressive part was when it connected to customer service and handled a whole texting convo on my behalf.