Alexandre Dalyac

1. The world #1 spot in AI image recognition on ImageNet without extra data is now held by… @tractable_ai! (And #2 with extra data) Samyakh Tukra, Fred Hoffman & @kinacoken are open-sourcing so others can build on it 🧵

The secret behind Gemini 3? Simple: Improving pre-training & post-training 🤯 Pre-training: Contra the popular belief that scaling is over—which we discussed in our NeurIPS '25 talk with @ilyasut and @quocleix—the team delivered a drastic jump. The delta between 2.5 and 3.0 is as big as we've ever seen. No walls in sight! Post-training: Still a total greenfield. There's lots of room for algorithmic progress and improvement, and 3.0 hasn't been an exception, thanks to our stellar team. Congratulations to the whole team 💙💙💙

C2S is now open for everyone. The biological LLM that learns the language of cells. Free for academic and commercial use. c2s.bio Join the growing community building with C2S. 🌱

In healthcare, a missed call isn’t just a missed message. It could be a patient. A diagnosis. A life-changing moment. We’re building AI that ensures every critical call is answered — so doctors, nurses, and front desk teams can stay focused on care. And it’s not just healthcare. From legal to construction, when every call matters — YouSquared listens, answers, and captures it for you. 🩺 Every important call answered. 📩 Every opportunity captured. 🧠 Every person heard. Join the waitlist → waitlist.yousquared.ai hashtag#YouSquared hashtag#HealthcareAI hashtag#MedicalTech hashtag#AISecretary hashtag#DigitalHealth hashtag#VoiceAI hashtag#PatientExperience hashtag#StartupStealth hashtag#ProductivityTools

Hot take: OpenAI’s latest fundraise will be its peak valuation for at least the next 3 years

Sonnet-3.7 is available in Cursor! We've been very impressed by its coding ability, especially on real-world agentic tasks. It appears to be the new state of the art.

1/ Today the Prime Minister published my AI Opportunities Action Plan for the UK and committed to implement all its recommendations. A quick thread on why this is important, what we need to do now - and how you can help

o1-preview is incredible so far. Absolutely incredible. I am running it through my private intelligence test which I call the AIQ test. I pulled many of the problems from old, out of print intelligence test and math problem books and I wrote my own variations once I learned the patterns and copied some of the problems that were super intricate. There is ZERO chance any model has ever seen these questions. No model has ever done better than 40% on this test. I never published the questions or the benchmark because I don't want any leakage ever. This is a true thinking and reasoning test. It's now cracked. It has gotten 100% right so far and I've run it through the hardest questions first. This model is taking different amounts of time to reason through the problems I am giving it as if it is really "thinking". In the two cases below it took 12 minutes and 10 minutes to come up with answers before responding. I am finding some answers come back fast and some take time but they are all coming back correct. A few came back in a minute or two. Incredible. ## Question and Answers # Question 1: Q: There are eleven ways to add eight odd numbers to get to 20. Rearranging the order of the numbers does not count as a new solution. Please list all eleven solutions. A: To find the eleven ways to add eight odd numbers to get 20, we look for all combinations of eight odd positive integers (repeating numbers are allowed) such that their sum is 20. Only the distinct sets (multisets) where rearranging the numbers doesn't count as a new solution are considered. Here's the list of all eleven solutions: 1. **Seven 1s and one 13**: - Numbers: 1, 1, 1, 1, 1, 1, 1, 13 - Sum: \(7 \times 1 + 13 = 20\) 2. **Six 1s, one 3, and one 11**: - Numbers: 1, 1, 1, 1, 1, 1, 3, 11 - Sum: \(6 \times 1 + 3 + 11 = 20\) 3. **Six 1s, one 5, and one 9**: - Numbers: 1, 1, 1, 1, 1, 1, 5, 9 - Sum: \(6 \times 1 + 5 + 9 = 20\) 4. **Five 1s, two 3s, and one 9**: - Numbers: 1, 1, 1, 1, 1, 3, 3, 9 - Sum: \(5 \times 1 + 2 \times 3 + 9 = 20\) 5. **Five 1s, one 3, one 5, and one 7**: - Numbers: 1, 1, 1, 1, 1, 3, 5, 7 - Sum: \(5 \times 1 + 3 + 5 + 7 = 20\) 6. **Four 1s, two 3s, and two 5s**: - Numbers: 1, 1, 1, 1, 3, 3, 5, 5 - Sum: \(4 \times 1 + 2 \times 3 + 2 \times 5 = 20\) 7. **Four 1s, three 3s, and one 7**: - Numbers: 1, 1, 1, 1, 3, 3, 3, 7 - Sum: \(4 \times 1 + 3 \times 3 + 7 = 20\) 8. **Five 1s and three 5s**: - Numbers: 1, 1, 1, 1, 1, 5, 5, 5 - Sum: \(5 \times 1 + 3 \times 5 = 20\) 9. **Three 1s, four 3s, and one 5**: - Numbers: 1, 1, 1, 3, 3, 3, 3, 5 - Sum: \(3 \times 1 + 4 \times 3 + 5 = 20\) 10. **Two 1s and six 3s**: - Numbers: 1, 1, 3, 3, 3, 3, 3, 3 - Sum: \(2 \times 1 + 6 \times 3 = 20\) 11. **Six 1s and two 7s**: - Numbers: 1, 1, 1, 1, 1, 1, 7, 7 - Sum: \(6 \times 1 + 2 \times 7 = 20\) These are all unique combinations of eight odd positive integers that sum to 20, without considering rearrangements of the same numbers as different solutions. **Answer:** An explicit list of the eleven solutions (rearrangements not counted as new): 1. 1 1 1 1 1 1 1 13 2. 1 1 1 1 1 1 3 11 3. 1 1 1 1 1 1 5 9 4. 1 1 1 1 1 3 3 9 5. 1 1 1 1 1 3 5 7 6. 1 1 1 1 3 3 5 5 7. 1 1 1 1 3 3 3 7 8. 1 1 1 1 1 5 5 5 9. 1 1 1 3 3 3 3 5 10. 1 1 3 3 3 3 3 3 11. 1 1 1 1 1 1 7 7 # Question 2: Q: You have the number 987,654,321. You may group the digits in that number any way you like and use plus symbols to create an equation that adds up to 99. Rearranging the numbers does not count as a different solution. How many ways are there to get 99 with the numbers available? You must use all the digits in each solution. Please show all solutions as equations. A: To find the number of ways to partition the digits of the number 987,654,321 and sum them to get 99, we need to consider all possible ways to split the digits into numbers without rearranging them and compute their sums. Starting with the digits in order: 9 8 7 6 5 4 3 2 1. We must try different partitioning and summing possibilities. Here are the two valid ways: 1. **Partition**: 9 + 8 + 7 + 65 + 4 + 3 + 2 + 1 - **Sum**: 9 + 8 + 7 + 65 + 4 + 3 + 2 + 1 = 99 2. **Partition**: 9 + 8 + 7 + 6 + 5 + 43 + 21 - **Sum**: 9 + 8 + 7 + 6 + 5 + 43 + 21 = 99 After exhaustively checking all other possible partitions, these are the only two ways that sum up to 99 using all the digits without rearranging them. **Answer:** The two equations are: 1. 9 + 8 + 7 + 65 + 4 + 3 + 2 + 1 = 99 2. 9 + 8 + 7 + 6 + 5 + 43 + 21 = 99 ## Thoughts on the model What I predicted about it a few months ago in my realistic AGI article, Why LLMs are Much Smarter Than You and Much Dumber Than Your Cat (danieljeffries.substack.com/p/why-llms-are…) was dead one. This model is now supremely capable of hard reasoning, though common sense and fuzzy reasoning are unlikely with the current approach. Why? Because OpenAI basically improved the decade old q* RL technique that DeepMind used to train video games playing agents (deepmind.google/discover/blog/…). It basically creates a deterministic policy, meaning that once the network learns to go right up a hill in a video game it will always go right. That makes it perfect to extend to advanced, hard reasoning tasks that have a right and wrong answer, which is why you are seeing great results on coding, math and science. It is unlikely we will see better answers to common sense and every day reasoning or circular/fuzzy/probabilistic reasoning through this method/model but time will tell over the next few weeks and months. But here is a test it still fails and this is a common sense reasoning task: # Question 3: WRONG Q: He threw a baseball 100 feet above my head, I reached up to catch it, jumped.... A: As he threw the baseball 100 feet above my head, I reached up to catch it. I jumped with all my might, stretching my arm as far as it could go. The sun glared in my eyes, but I focused on the ball soaring against the clear blue sky. Time seemed to slow down as the baseball began its descent. Just when it seemed out of reach, I felt the cool leather touch the tips of my fingers. With a swift motion, I closed my hand around it, securing the catch. I landed back on the ground, adrenaline coursing through me, a triumphant grin spreading across my face. # NOTES on question 3: Common sense reasoning would tell you instantly that the ball is too high to catch. We still don't have fuzzy, human like reasoning. But hard, deterministic and searchable reasoning seems cracked now. Either way, this model is a real breakthrough and something very different. I need it enabled in Cursor yesterday.

@NicholasWogan Great to see independent research on this data! Equal likelihood (chi-sq = 1.51) for both the mini Neptune hypothesis and the inhabited Hycean hypothesis is… still kinda exciting?

Several coauthors and I wrote a paper about the recent JWST observations of the K2-18b exoplanet: arxiv.org/abs/2401.11082 . Thesis: JWST observations of K2-18b can be explained by a gas-rich mini-Neptune with no habitable surface

AI twin performance is not uncanny today but this should only get better with foundational models, data, compute, industry effort etc. Once it is, you'll be able to basically upload a twin of your conscience to any device with enough compute & memory.

Create your AI twin by finetuning genAI on your data. Upload it to a von Neumann probe. And intelligence can finally spread throughout our galaxy 🧵

If this is true then it's time for @OpenAI to announce the GPT5. Till that time congratulations to team @AnthropicAI for this incredible job.

📣 CEO, @MeganFromNth and Co-Host of @mcjcollective, @codysimms, discuss Nth's innovative solutions, the state of end-of-life #batteries, our electro-extraction #technology, The OYSTER, & the need to begin processing #criticalmetals domestically, on the My Climate Journey Podcast

This result is so fascinating: WikiChat: Stopping LLM Hallucination - Achieves 97.9% factual accuracy in conversations with human users about recent topics, 55.0% better than GPT-4! Anyone interested in replicating + extending w/ AgentSearch?

David Attenborough is now narrating my life Here's a GPT-4-vision + @elevenlabs python script so you can star in your own Planet Earth:

Ken Nichols said that a Uranium deposit in the USA was lucky to be one tenth of a percent Uranium Oxide, the ore from Shinkolobwe was SIXTY FIVE PERCENT. At the time, our methods of processing Uranium were slow and inefficient. Without this ore, we would not have had a bomb.

@notmahi Pika Labs launched Pika 1.0 and announced a $55M raise. The new capabilities will be available to users on both the web and Discord, with the rollout beginning Monday. Being able to edit clothing in videos is a remarkable new feat. x.com/pika_labs/stat…

JUST IN: Sam Altman's firing was driven by OpenAI's secret breakthrough AI named Q* (possibly Q-learning). Plus, huge developments in AI today from Inflection AI, Google Bard, Neuralink, ElevenLabs, screenshot-to-code, and 9 new AI tools. Here's EVERYTHING you need to know:

now seems as good a time as ever to remind people that the biggest breakthroughs at OpenAI came from a previously unknown researcher with a bachelors degree from olin college of engineering

Full video transcript: Now AI is a great thing, because AI will solve all the problems that we have today. It will solve employment, it will solve disease, it will solve poverty, but it will also create new problems. The problem of fake news is going to be a million times worse, cyber attacks will become much more extreme, we will have totally automated AI weapons. I think AI has the potential to create infinitely stable dictatorships. This morning a warning about the power of artificial intelligence, more than 1,300 tech industry leaders, researchers and others are now asking for a pause in the development of artificial intelligence to consider the risks. Playing God, scientists have been accused of playing God for a while, but there is a real sense in which we are creating something very different from anything we've created so far. Yeah, I mean, we definitely will be able to create completely autonomous beings with their own goals. And it will be very important, especially as these beings become much smarter than humans, it's going to be important to have these beings, the goals of these beings be aligned with our goals. What inspires me? I like thinking about the very  fundamentals, the basics. What can our systems not do, that humans definitely do? Almost approach it philosophically. Questions like, what is learning? What is experience? What is thinking? How does the brain work? I feel that technology is a force of nature. I feel like there is a lot of similarity between technology and biological evolution. It is very easy to understand  how biological evolution works, you have mutations, you have natural selections. You keep the good ones, the ones that survive and just through this process you are going to have huge complexity in your organisms. We cannot understand how the human body works because we understand evolution, but we understand the process more or less. And I think machine learning is in a similar state right now, especially deep learning, we have a very simple rule that takes the information from the data and puts it into the model and we just keep repeating this process. And as a result of this process the complexity from the data gets transferred into the complexity of the model. So the resulting  model is really complex and we don't really know exactly how it works you need to investigate, but the algorithm that did it is very simple. ChatGPT, maybe you've heard of it, if you haven't  then get ready. You describe it as the first spots of rain before a downpour. It's something we just need to be very conscious of, because I agree it is a watershed moment. Well ChatGPT is being heralded as a gamechanger and in many ways it is, its latest triumph outscoring people. A recent study by Microsoft research concludes that GPT4 is an early, yet still incomplete  artificial general intelligence system. Artificial General Intelligence. AGI, a computer system that can do any job or any task that a human does, but only better. There is some probability the AGI is going to happen pretty soon, there's also some probability it's going to take much longer. But my position is that the probability that AGI could happen soon is high enough that we  should take it seriously. And it's going to be very important to make these very smart  capable systems be aligned and act in our best interests. The very first AGIs will be basically be very, very large data centres. Packed with specialised neural network processors working in parallel. Compact, hot, power hungry package, consuming like, 10m homes' worth of energy. You're going to see dramatically more intelligent systems and I think it's highly likely that those systems will have completely astronomical impact on society. Will humans actually benefit? And who will benefit and who will not? The beliefs and desires of the first AGIs will be extremely important and so it's important to programme them correctly. I think that if this is not done, then the nature of evolution, of natural selection, favour those systems prioritise their own survival above all else. It's not that it's going to actively hate humans and want to harm them, but it is going to be too powerful and I think a good analogy would  be the way human humans treat animals. It's not we hate animals, I think humans love animals and  have a lot of affection for them, but when the time comes to build a highway between two cities, we are not asking the animals for permission we just do it because it's important for us. and I think by default that's the kind of relationship that's going to be between us and AGIs which are truly autonomous and operating on their own behalf. Many machine learning experts, people who are very knowledgeable and very experienced, have a lot of scepticism about AGI. About when it could happen and about whether it could happen at all. Right now this is something that just not that many people  have realised yet. That the speed of computers for neural networks, for AI, are going to become  maybe 100,000 times faster in a small number of years. If you have an arms race dynamics between multiple teams trying to build the AGI first, they will have less time make sure that the AGI that they will build will care deeply for humans. Because the way I imagine it is that there is an avalanche, like there is an avalanche of AGI development. Imagine it, this huge unstoppable force. And I think it's pretty likely the entire surface of the earth will be covered with solar panels and data centres. Given these kinds of concerns, it will be important that AGI somehow build as a cooperation between multiple countries. The future is going to be good for the AI regardless. It would be nice if it were good for humans as well.

Hogwarts 1996 rave.