meet @mostik_ai! what happens when you put 12 PhDs in one room for four months? first place on the ARC-AGI leaderboard, which I can't say much about while the competition is still running. and this, which I can. everyone's arguing about whether open models will catch up to frontier models. we think it's the wrong question. here's the one we pose: why does a frontier model have to generate your answer at all, when the only thing you need from it is the reasoning? we do this by enabling models to communicate in latent space. through our protocol, hidden states pass straight from a frontier model into a small one running on your infrastructure -- no text between them, and neither model is fine-tuned. two models from different families, sharing reasoning, both left untouched. how do we know it works? we tested it on a setup where a 753B model reads the problem, and a 4B edge-class model writes the answer. with this approach, we get results 80% as accurate as the frontier model, but at 20x faster performance. we're committed to preventing frontier model lock-in and are already partnering with inference providers to accelerate open-weight adoption. we've done this between 15 of us, in four months, 12 PhDs and a Fields medalist, backed by @generalcatalyst WIRED has the first external account of the company and the work: wired.com/story/russian-star… full writeup, the setup, and all the numbers: mostik.ai/read-more

Speech-to-text is solved. Unless you have an accent. Over 1 million people from 180 countries have spent 580,000 hours speaking with Fluently AI tutors while learning a new language. Nearly all of them have an accent, and transcribing their speech is still hard. So we benchmarked streaming speech-to-text systems on spontaneous, accented English from our learners. Leaderboard and methodology in the thread. English is the easy one. Our learners study 14 other languages too. Working on tough speech problems? DMs are open.

The largest math proof in human history is now computer-verified. Today, Axiom announces a complete Lean auto-formalization of the Classification of Finite Simple Groups, the "periodic table of symmetry." About 15,000 pages, built by over 100 mathematicians across generations.

the Exponential Time Hypothesis is false: 3-SAT can be solved in 2^o(n) time, and @OpenAI and @AnthropicAI got there independently and simultaneously! an internal Claude model found its algorithm with no human input while OpenAI’s model found a different proof allegedly the same day Claude’s bound only beats brute force past ~10^10000 variables (so nothing changes in practice), and human verification is just starting (for OpenAI by Impagliazzo and Paturi, who proposed ETH in 2001) Ryan Williams put ETH at 70% likely in 2019, and 800+ papers have results resting on it 😬

reading this new latent communication paper, out last week: arxiv.org/html/2610.26840v1 their setup is very practical: when one article gets 18 questions, a model reading it as text processes the whole article 18 times, whereas here one model reads it once (before any question) and compresses it into a fixed 384 KiB packet, and then a different model answers every question from that packet (so the receiver reads ~100 tokens instead of ~5,900, and the article’s questions go 3x faster) they checked that accuracy drops 20 points in case of swapping in another article’s packet, so the receiver does use it (on datasets it wasn’t trained on, plain text still wins by ~15 points though) and a 9B encoder is barely better than a 0.8B one, but for reading comprehension that makes sense as the answer sits in the text at @mostik_ai we’re betting on the case where the big model’s computation is the part worth sending

one more from Essentialism: do the right few things superbly well, and the contribution follows as a CEO my few things are clear by now: find the right people, keep them excited about the problem, and turn the vision into goals they can own everything else ideally belongs to someone on the team who’d do it better than me (which with @mostik_ai superb team is most things 😄)

a week of flights means a lot of audiobook hours, so this is book two: Essentialism by @GregoryMcKeown, which I’d been meaning to get to for ages the idea I liked most is the 90 percent rule: score every option, and anything below 90 is a no, so a 70 doesn’t get to be a maybe it’s pretty much how I’ve built the team at @mostik_ai 🙂 if someone isn’t a clear yes, they’re a no (with only 15 people, every hire is a noticeable share of the room)

12 PhDs and a Fields medalist vs a home pizza oven to be fair, our research turns out better than our crusts, but it was very tasty 😄

almost done with The Great CEO Within, and the last thing I underlined is the energy audit: mark what in your week gives you energy and what drains it, delegate the drains, aim for 75-80% of your time on the first kind the uncomfortable part as a CEO is that the energizing stuff (for me, research and whiteboards) isn’t always what the company needs from me that week, so I read it as a delegation exercise more than permission to do only the fun parts 🥲

another one I liked from the Mochary book: whoever raises an issue brings a proposed solution in writing, and “I don’t know” isn’t allowed, one have to guess for a research team the guess is the most useful part because “something’s off with the eval” can sit in a chat for a week, while “I think the eval leaks answers through the prompt template” tells you what to check first and prompts the team to come up with other hypotheses

Big news: 2027dev is now Firegoat! ❤️‍🔥🐐 For the next chapter of our company, @firegoat_ai felt like a name that reflects what we're here for. Our mission is to empower teams to ship agent-ready products with confidence and ease. We're already working with incredible agent-first developer teams at @browserbase, @e2b, @mastra, @clerk, @daytonaio , @sazabi & more. The best part: it's only the beginning. We're launching exciting things this week - stay tuned! Aaand we're hiring exceptional backend+infra/product engineers 👀 DMs are open

still listening to The Great CEO Within, and the chapter on buy-in lists three ways to make a team decision: decide and announce, put a straw man on the table and let people attack it, or start from a blank page and aim for consensus with researchers I’d pick the straw man almost every time! people who’d never volunteer a half-baked idea will happily take one apart, so you get all the objections in one meeting decide-and-announce only works if nobody in the room knows more than you, which is rarely my situation in @mostik_ai 😂

I’ve had a rule for a while: the third time I type the same prompt, it becomes a Claude skill @mattmochary’s version in The Great CEO Within is even stricter: "if you do it or say it twice, write it down" after last week’s @mostik_ai offsite I’m switching to his -- a week in one room is the best bandwidth a team gets, but by Monday most of it lives only in people’s heads, so anything that came up twice goes into a doc now

two new papers checked what happens to agent’s bill when you cut the amount of text it works with, and the results are quite counterintuitive in the first, a compression tool cut 38.4% of the text the agent gets back from its tools, and the bill went up(!) by 6.8% about 87% of the cost was the model processing the long context it keeps carrying from step to step, and the text the tool trimmed was only about 3.3% of the bill arxiv.org/abs/2607.12161 in the second, across nearly 35,000 agent runs, setups that used a third of the tokens took up to 1.8x longer to finish arxiv.org/abs/2609.32961 so for anyone running agents (and who doesn't today?😄), the number to track is cost per finished task, and most of that cost comes from context getting passed around and read again @mostik_ai helps to solve that bottleneck of today's models that hand work to each other as text (so each next model has to read everything again from scratch), but if the first model passes along what it has already computed, that rereading isn’t needed

barefoot by the ocean in October, a very effective reset after a full day of arguing about geometry 😄

AI is getting cheaper faster than almost any technology before it @EpochAIResearch measured it nicely and they found that getting the same quality of answer costs about 47% less every quarter, so roughly 13x cheaper per year e.g. last year o3 scored 75% on GPQA Diamond (hard science test) at about 30 cents per question, and GPT-5.6 Luna now gets the same score for $0.0004, which is 725x cheaper in under 18 months so any AI cost estimate from July is already about 2x off it’s also why we built @mostik_ai around combining models: a small, cheap model borrows what a big model understands, so you pay big-model prices only for the part that needs a big model epoch.ai/publications/the-pl…

new paper from USC, UC Irvine and Intel Labs on reusing a small model's prefill inside a big one: the small model reads the prompt, a linear map translates its KV cache into the big model's format, and the big model recomputes only part of the positions Qwen3-0.6B prefilling for Qwen3-14B (23x apart in size) with 30% of positions recomputed: GSM8K goes from 71.6% with the map alone to 90.2%, against 95.1% for the 14B on its own, and prefill runs up to 3.04x faster the map alone can leave the big model worse off than the small one: Llama 3.1 8B into Llama 3.3 70B scores 81.0%, below the 8B by itself (82.7%), and recomputing 40% of positions brings it to 92.0% section 2.2 is worth reading on its own: choosing which positions to recompute by attention score does worse than choosing at random, because the 14B puts 72% of its attention on the start token and 24% on the chat template and instruction, which leaves 4% for the actual question (at a 20% budget the attention-based selector doesn't pick a single question token) so attention weight mostly tells you where the sink and the template are arxiv.org/abs/2610.11358

whiteboard by day, the same arguments by evening, just with better snacks 🍷

.@OpenAI put 722 AI-written math manuscripts on GitHub last week: 372 result families out of roughly 4,000 problems their models attempted that's a huge result, especially when you remember that not so long ago models couldn't tell whether 9.9 or 9.11 is bigger 😄 still, I'm with the mathematicians who want something else from AI: the research I admire finds the one insight that makes a problem simple and fits the whole argument into 10 elegant pages (rather than 50+ pages of clever slop that's hard to follow and even harder to check is a different kind of output, and it's still far from how real mathematicians work) I'm lucky to watch great mathematicians do this up close every day, so I know what the bar looks like I hope models get there one day, and I'd love to help them get there ✨ github.com/openai/math

full brainstorm mode with most of the team on-site, and the bandwidth in one room is unreal, roughly the gap between text and latents but for humans 😄 Stas Smirnov joined too, so every whiteboard argument now has a final boss

no arXiv today, letting my brain reset in front of something with zero tables in it 🎨