Aravind Srinivas explains why Perplexity is building a browser: it's the only way to create AI agents with enough control over multiple apps, especially on iOS. Their goal? Agents that can book travel, buy things, and act as a personalized assistant for those who can't afford a human EA. He adds that this is the long-term vision. "Anyone who's saying agents will work in 2025 should be skeptical."

New breakthrough quantum algorithm published in @Nature today: Our Willow chip has achieved the first-ever verifiable quantum advantage. Willow ran the algorithm - which we’ve named Quantum Echoes - 13,000x faster than the best classical algorithm on one of the world's fastest supercomputers. This new algorithm can explain interactions between atoms in a molecule using nuclear magnetic resonance, paving a path towards potential future uses in drug discovery and materials science. And the result is verifiable, meaning its outcome can be repeated by other quantum computers or confirmed by experiments. This breakthrough is a significant step toward the first real-world application of quantum computing, and we're excited to see where it leads.

OPENAI ROADMAP UPDATE FOR GPT-4.5 and GPT-5: We want to do a better job of sharing our intended roadmap, and a much better job simplifying our product offerings. We want AI to “just work” for you; we realize how complicated our model and product offerings have gotten. We hate the model picker as much as you do and want to return to magic unified intelligence. We will next ship GPT-4.5, the model we called Orion internally, as our last non-chain-of-thought model. After that, a top goal for us is to unify o-series models and GPT-series models by creating systems that can use all our tools, know when to think for a long time or not, and generally be useful for a very wide range of tasks. In both ChatGPT and our API, we will release GPT-5 as a system that integrates a lot of our technology, including o3. We will no longer ship o3 as a standalone model. The free tier of ChatGPT will get unlimited chat access to GPT-5 at the standard intelligence setting (!!), subject to abuse thresholds. Plus subscribers will be able to run GPT-5 at a higher level of intelligence, and Pro subscribers will be able to run GPT-5 at an even higher level of intelligence. These models will incorporate voice, canvas, search, deep research, and more.

Self-improvement 🔄 SIMA 2 can teach itself new skills, learning through trial-and-error, based on feedback from Gemini. Getting better the more it plays – without additional human input.

iykyk arxiv.org/pdf/2509.24527

AI Progress since GPT-3.5 OpenAI seems to be slowing down with GPT-5 Anthropic incredibly steady progress Google had it's breakthrough with Gemini 2.5 Pro

Zuck's message to Meta staff post-earnings -- Even if we're wrong about our AI investments, we're not going to go bankrupt. (via @alexeheath) sources.news/p/chatgpts-quie…

GOOGLE AND ANTHROPIC ARE REPORTEDLY IN TALKS ON CLOUD DEAL WORTH TENS OF BILLIONS Anthropic is in discussions with Google $GOOGL about a deal that would provide the with additional computing power valued in the high tens of billions of dollars The plan, which has not been finalized, involves Google providing cloud computing services to Anthropic - Bloomberg Google stock is up 3% in after hours on the news 🟢🟢🟢

“ our research team did something unexpected and quite amazing and we think it will be very very worth the wait, but needs a bit longer. “ Care to elaborate ?

META attempted to buy Ilya Sutskever's  Safe Superintelligence, and also attempted to hire him, according to reporting tonight by CNBC.

working nights and weekends

Keep thinking.

Microsoft $MSFT said it is pursuing a more powerful form of AI called “superintelligence” Mustafa Suleyman, chief of the Microsoft AI group, will lead what the company is calling the MAI Superintelligence Team that will target hypothetical milestones that are even more ambitious than artificial general intelligence “If AGI is often seen as the point at which an AI can match human performance at all tasks, then superintelligence is when it can go far beyond that performance,” (Source Bloomberg)

I'm back at the top of ARC-AGI with my new program. I use @grok 4 and multi-agent collaboration with evolutionary test-time compute

I agree, although I believe most matrix multiplies for AI are done in two dimensions. Anyway, I stand by my statement. Matmuls are an effective piece of mathematical machinery, but the mechanics of calculating them are headache-inducing. Indeed, it's exactly their computationally cumbersome nature that's propelling the data center boom. That point seems beyond argument.

Gemini 2.5 pro is the first model to ever get this image correct

Introducing Claude Sonnet 4.5—the best coding model in the world. It's the strongest model for building complex agents. It's the best model at using computers. And it shows substantial gains on tests of reasoning and math.

Noam is right

So, all the models underperform humans on the new International Mathematical Olympiad questions, and Grok-4 is especially bad on it, even with best-of-n selection? Unbelievable!

right now is the time where the takeoff looks the most rapid to insiders (we don’t program anymore we just yell at codex agents) but may look slow to everyone else as the general chatbot medium saturates

Introducing Tinker: a flexible API for fine-tuning language models. Write training loops in Python on your laptop; we'll run them on distributed GPUs. Private beta starts today. We can't wait to see what researchers and developers build with cutting-edge open models! thinkingmachines.ai/tinker

Tomorrow

🧩 Thinking Machines’s mission is to make AI models predictable instead of inconsistent. For example, ask ChatGPT the same question a few times over, and you’re likely to get a wide range of answers. This has largely been accepted in the AI community as a fact — today’s AI models are considered to be non-deterministic systems— but Thinking Machines Lab sees this as a solvable problem. Thinking Machines published a solid blog yesterday and says, that AI randomness mostly comes from how GPU kernels get stitched and scheduled during inference, and that tighter control here can deliver deterministic inference. That would give enterprises repeatable answers and make reinforcement learning cleaner by cutting label noise from slightly different outputs. A GPU kernel is a tiny program that runs the math on the graphics processor, and when thousands run in parallel, small things like the order of adds, thread timing, atomic updates, or library choices can nudge numbers in different directions. Floating point math is not perfectly associative, so changing the reduction order or mixing precisions tweaks the last few bits, which then cascades through layers into a different token choice. The proposed fix is an orchestration layer that locks down kernels, seeds, algorithms, and execution graphs, pins math library versions, and enforces a repeatable schedule across runs and machines. That kind of guardrail usually trades a bit of throughput for stability, but it pays off when scientists or auditors need the exact same answer on rerun. The lab also points to using reinforcement learning to tailor models for businesses, which gets easier when the reward signal comes from consistent model outputs instead of a fuzzy mix. They plan frequent research posts and code releases under a series called Connectionism, and a first product aimed at researchers and startups is said to be coming soon. Given a $12B valuation, the big question is whether they can prove determinism at scale without hurting speed and cost. The view here is that kernel-level determinism is a practical, engineering-first path, and if they show identical outputs across nodes and GPUs with minimal slowdown, that would count as a real win for production LLMs. In July, Murati said Thinking Machines Lab will launch its first product within months for researchers and startups. What it is, and whether it boosts reproducibility, is unknown.

🚨 Gemini 3.0 Pro - ONE SHOTTED I asked it for windows web os as everyone asked me for it and the result is mind blowing , it even has python in terminal and we can play games and run code in it Google really cooked here , source code and prompt in comment box

🧠 How can we equip LLMs with memory that allows them to continually learn new things? In our new paper with @AIatMeta, we show how sparsely finetuning memory layers enables targeted updates for continual learning, w/ minimal interference with existing knowledge. While full finetuning and LoRA see drastic drops in held-out task performance (📉-89% FT, -71% LoRA on fact learning tasks), memory layers learn the same amount with far less forgetting (-11%). 🧵:

My pleasure to come on Dwarkesh last week, I thought the questions and conversation were really good. I re-watched the pod just now too. First of all, yes I know, and I'm sorry that I speak so fast :). It's to my detriment because sometimes my speaking thread out-executes my thinking thread, so I think I botched a few explanations due to that, and sometimes I was also nervous that I'm going too much on a tangent or too deep into something relatively spurious. Anyway, a few notes/pointers: AGI timelines. My comments on AGI timelines looks to be the most trending part of the early response. This is the "decade of agents" is a reference to this earlier tweet nitter.app/karpathy/status/188254… Basically my AI timelines are about 5-10X pessimistic w.r.t. what you'll find in your neighborhood SF AI house party or on your twitter timeline, but still quite optimistic w.r.t. a rising tide of AI deniers and skeptics. The apparent conflict is not: imo we simultaneously 1) saw a huge amount of progress in recent years with LLMs while 2) there is still a lot of work remaining (grunt work, integration work, sensors and actuators to the physical world, societal work, safety and security work (jailbreaks, poisoning, etc.)) and also research to get done before we have an entity that you'd prefer to hire over a person for an arbitrary job in the world. I think that overall, 10 years should otherwise be a very bullish timeline for AGI, it's only in contrast to present hype that it doesn't feel that way. Animals vs Ghosts. My earlier writeup on Sutton's podcast nitter.app/karpathy/status/197343… . I am suspicious that there is a single simple algorithm you can let loose on the world and it learns everything from scratch. If someone builds such a thing, I will be wrong and it will be the most incredible breakthrough in AI. In my mind, animals are not an example of this at all - they are prepackaged with a ton of intelligence by evolution and the learning they do is quite minimal overall (example: Zebra at birth). Putting our engineering hats on, we're not going to redo evolution. But with LLMs we have stumbled by an alternative approach to "prepackage" a ton of intelligence in a neural network - not by evolution, but by predicting the next token over the internet. This approach leads to a different kind of entity in the intelligence space. Distinct from animals, more like ghosts or spirits. But we can (and should) make them more animal like over time and in some ways that's what a lot of frontier work is about. On RL. I've critiqued RL a few times already, e.g. nitter.app/karpathy/status/194443… . First, you're "sucking supervision through a straw", so I think the signal/flop is very bad. RL is also very noisy because a completion might have lots of errors that might get encourages (if you happen to stumble to the right answer), and conversely brilliant insight tokens that might get discouraged (if you happen to screw up later). Process supervision and LLM judges have issues too. I think we'll see alternative learning paradigms. I am long "agentic interaction" but short "reinforcement learning" nitter.app/karpathy/status/196080…. I've seen a number of papers pop up recently that are imo barking up the right tree along the lines of what I called "system prompt learning" nitter.app/karpathy/status/192136… , but I think there is also a gap between ideas on arxiv and actual, at scale implementation at an LLM frontier lab that works in a general way. I am overall quite optimistic that we'll see good progress on this dimension of remaining work quite soon, and e.g. I'd even say ChatGPT memory and so on are primordial deployed examples of new learning paradigms. Cognitive core. My earlier post on "cognitive core": nitter.app/karpathy/status/193862… , the idea of stripping down LLMs, of making it harder for them to memorize, or actively stripping away their memory, to make them better at generalization. Otherwise they lean too hard on what they've memorized. Humans can't memorize so easily, which now looks more like a feature than a bug by contrast. Maybe the inability to memorize is a kind of regularization. Also my post from a while back on how the trend in model size is "backwards" and why "the models have to first get larger before they can get smaller" nitter.app/karpathy/status/181403… Time travel to Yann LeCun 1989. This is the post that I did a very hasty/bad job of describing on the pod: nitter.app/karpathy/status/150339… . Basically - how much could you improve Yann LeCun's results with the knowledge of 33 years of algorithmic progress? How constrained were the results by each of algorithms, data, and compute? Case study there of. nanochat. My end-to-end implementation of the ChatGPT training/inference pipeline (the bare essentials) nitter.app/karpathy/status/197775… On LLM agents. My critique of the industry is more in overshooting the tooling w.r.t. present capability. I live in what I view as an intermediate world where I want to collaborate with LLMs and where our pros/cons are matched up. The industry lives in a future where fully autonomous entities collaborate in parallel to write all the code and humans are useless. For example, I don't want an Agent that goes off for 20 minutes and comes back with 1,000 lines of code. I certainly don't feel ready to supervise a team of 10 of them. I'd like to go in chunks that I can keep in my head, where an LLM explains the code that it is writing. I'd like it to prove to me that what it did is correct, I want it to pull the API docs and show me that it used things correctly. I want it to make fewer assumptions and ask/collaborate with me when not sure about something. I want to learn along the way and become better as a programmer, not just get served mountains of code that I'm told works. I just think the tools should be more realistic w.r.t. their capability and how they fit into the industry today, and I fear that if this isn't done well we might end up with mountains of slop accumulating across software, and an increase in vulnerabilities, security breaches and etc. nitter.app/karpathy/status/191558… Job automation. How the radiologists are doing great nitter.app/karpathy/status/197122… and what jobs are more susceptible to automation and why. Physics. Children should learn physics in early education not because they go on to do physics, but because it is the subject that best boots up a brain. Physicists are the intellectual embryonic stem cell nitter.app/karpathy/status/192969… I have a longer post that has been half-written in my drafts for ~year, which I hope to finish soon. Thanks again Dwarkesh for having me over!

GPT-5 singlehandedly destroyed the fast takeoff narrative and killed the stock market

Btw, since people don’t seem to know this, you can literally spawn subagents in Claude Code just by asking.

Yann LeCun: I'm not interested in LLMs anymore - they're the past. The future is in four more interesting areas: machines that understand the physical world, persistent memory, reasoning, and planning.

Keep thinking.

We've made progress on the AI safety problem of detecting and reducing "scheming": - Created evaluation environments to detect scheming - Observed current models scheming in controlled settings - Found deliberative alignment (openai.com/index/deliberativ…) decreases scheming rates These are some of the most exciting long-term AI safety results to date, and there's still a lot of work left to do. Looking forward to seeing further work done in this space. Research done in collaboration with @apolloaievals: openai.com/index/detecting-a…

Mark Zuckerberg has successfully lured away three OpenAI researchers. Lucas Beyer, Alexander Kolesnikov and Xiaohua Zhai have joined META's superintelligence lab.

anthropic.claude-3-7-sonnet-20250219-v1:0 Claude 3.7 Sonnet is Anthropic's most intelligent model to date and the first Claude model to offer extended thinking - the ability to solve complex problems with careful, step-by-step reasoning. Anthropic is the first AI lab to introduce a single model where users can balance speed and quality by choosing between standard thinking for near-instant responses or extended thinking or advanced reasoning. Claude 3.7 Sonnet is state-of-the-art for coding, and delivers advancements in computer use, agentic capabilities, complex reasoning, and content generation. With frontier performance and more control over speed, Claude 3.7 Sonnet is the ideal choice for powering AI agents, especially customer-facing agents, and complex AI workflows. Supported use cases: RAG or search & retrieval over vast amounts of knowledge, product recommendations, forecasting, targeted marketing, code generation, quality control, parse text from images, agentic computer use, content generation Model attributes: Reasoning, Text generation, Code generation, Rich text formatting, Agentic computer use

Newsletter: I am sick and tired of everybody pretending that generative AI is the next big thing. The media is complicit in accepting fantastical nonsense - both in the numbers put out by OpenAI and the silly jobs created by Anthropic - it has to stop. wheresyoured.at/reality-chec…

I was laid off by Meta today. As a Research Scientist, my work was just cited by the legendary @johnschulman2 and Nicholas Carlini yesterday. I’m actively looking for new opportunities — please reach out if you have any openings!

1/N I’m excited to share that our latest @OpenAI experimental reasoning LLM has achieved a longstanding grand challenge in AI: gold medal-level performance on the world’s most prestigious math competition—the International Math Olympiad (IMO).

Utterly confusing, funny and sad

I don't need to sign up to ChatGPT to understand it's limitations & strengths, I have read dozens of research papers by now on the limitations of decoder-only transformer models that power OpenAI models. And I am confident these models aren't intelligent.

*APPLE FINALIZING PACT THAT WOULD PAY GOOGLE ROUGHLY $1B A YEAR *GOOGLE GEMINI AI MODEL TO HELP RUN SIRI FEATURES DUE IN 2026 *APPLE TO USE 1.2 TRILLION PARAMETER GOOGLE MODEL TO POWER SIRI

Human engineers for example can predictably design software, planes, cars, computers etc an artist can predictably produce art & a cook predictably, cook. Generative models on the other hand? Good luck getting them to do what you asked for.

Introducing T5Gemma: the next generation of encoder-decoder/T5 models! 🔧Decoder models adapted to be encoder-decoder 🔥32 models with different combinations 🤗Available in Hugging Face and Kaggle developers.googleblog.com/en…

Today Thinking Machines Lab is launching our research blog, Connectionism. Our first blog post is “Defeating Nondeterminism in LLM Inference” We believe that science is better when shared. Connectionism will cover topics as varied as our research is: from kernel numerics to prompt engineering. Here we share what we are working on and connect with the research community frequently and openly. The name Connectionism is a throwback to an earlier era of AI; it was the name of the subfield in the 1980s that studied neural networks and their similarity to biological brains. thinkingmachines.ai/blog/def…

We’re releasing GPT-5-Codex — a version of GPT-5 further optimized for agentic coding in Codex. Available in the Codex CLI, IDE Extension, web, mobile, and for code reviews in Github. openai.com/index/introducing…

it is 2025 and there are people who believe that accurately predicting the next token does not require understanding the underlying reality that created that token btw

Next version insanely better is the plan

Anthropic just sent the next model, codenamed Neptune V6, to red teamers and launched a 10-day challenge with extra bonuses for confirmed universal jailbreaks

BREAKING 🚨: Eggs Egg Prices have now collapsed 86% since the start of March 🥚🐔📉

Sonnet 4.5 is out! It’s the most aligned frontier model yet; a lot of progress relative to Sonnet 4 and Opus 4.1!