Sundar is the only lab leader whose statements on RSI match the data released in their model cards.

this is the part that fascinates me. both the labs are amazing, but the low switching cost is just...wild theinformation.com/articles/…

Frontier labs making custom chips is not really as hard / as impressive as you might think. 1. They are primarily targeting inference not training. inference is easier 2. They are targeting transformers only. Think large MXU matmul units 3. They are writing all kernels in close to ISA with coding agents RLed for their ISA. 4. They aren’t making high level compilers for codegen A frontier lab vending a true generic chip like a Nvidia GPU or Google TPU is MUCH harder than spinning up an in house inference program. Still I think you will see many more inference chip companies as the cost of the software is just the cost of tokens.

OpenAI models are coming to Google Cloud. We're bringing our frontier models and Codex agentic workflows to Google Cloud's Model Garden, so enterprises can use our technology where they already build. openai.com/index/openai-mode…

Some thoughts about the frontier AI business right now: 1. The recipe for making AI coding models is widely known throughout the industry and all of Silicon Valley. There is no secret to this. It is primarily RL envs + Compute. 2. I predict in single digit years most companies / individuals will use free AI coding models. Free meaning sold at chip time (cost). There will be no real business in selling AI coding models, as there is no inherent moat in making AI coding models. 3. Frontier labs without cash producing businesses will be unable to subsidize AI coding models and will transition to selling other forms of intelligence like biology, chemistry, pharmaceuticals etc. 4. Most companies you see that are still behind in coding models like @Muse and @SpaceXSI are mainly due to lack of RL envs and harder tasks. What I am saying here is that these gaps will further close, as these companies make more RL envs.

Real life, baby

Dumb question but what's the advantage of $META having spent 100b+ to develop their own model under Muse, as opposed to having kept that money and just built Muse with whatever model gets the job done?

I left Google DeepMind in June. Jacob is right: many researchers believe they are building something that could kill everyone on the planet. It was literally my day job to think about how to stop that.

Opus 5.5 is currently the best model. Astra and the new sol are oddly again bad at writing, communication and style to the point it makes it hard to figure out what’s going on during long auto research loops. Quite frustrating. Whatever made 5.6 sol amazing should be repeatable. But somehow it’s not. Funny how writing, comm style and personality are so hard to get consistently right. Why?

ok i think i finally figured out why alignment has gone basically nowhere after 15 years and several billion dollars. the problem is hard sure. but have you met the people in charge of it 1) the leadership does not know how computers work. ask one what happens when you type a url into a browser and watch them start sweating. dns is a spooky word. they genuinely believe the model lives in "the cloud" like a ghost haunting a castle 2) zero idea how large scale distributed systems work. never been paged at 3am. never watched one bad bgp announcement take half the internet offline. their big threat model is the ai "copying itself across the internet" like its a floppy disk virus from 1998. brother it needs 50 megawatts, a liquid cooling loop and a personal relationship with jensen. it is not escaping onto your smart fridge 3) they dont understand the stack they are supposedly protecting us from. ask about inference servers, kv cache, batching, egress, rate limits. blank stare. ask how the model will seize the power grid and you get a 45 minute answer with a hand drawn diagram and three links to their own blog 4) nobody can define an agent. ask ten of them and you get eleven definitions and a 30 page google doc. to the rest of us an agent is a while loop that calls tools and eats a 429 every thirty seconds. to them its a digital god in its larval stage. my agent cant book a dentist appointment without asking me three times if im sure 5) they started from "everyone dies" and worked backwards so every result is bad news somehow. model refuses, deceptive. model complies, sycophantic. model does well on evals, well now its scheming about the evals. i cant think of a single thing a model could do that would get them to say ok maybe we're fine

Anthropic has set up a wet lab in San Francisco, their bio research has advanced beyond in silico. They are now conducting physical research both in-house and with external partners. The NYT also broke the news this morning that Anth now plans to scale its compute to 5 GW by the end of the year, and 10 GW within the next 16 months. Bio is next after math, and I think we are going to see some medical announcements much sooner than people are expecting.

Early game in sg is overrated. Mostly a deep dive into a local minima (a decent one but far from global). I frequently tell others the reason why I'm here today is because I did horribly for a levels and CS was a alevels dumping ground back then. If I had a strong early game maybe I could have accidentally gone into finance or law. 😳 Common wisdom in sg is also generally local minima diving and more often than not, just bad advice. Be brave to be who you are and enjoy what you do.

Broadcom CEO Hock Tan breaks down AI economics: Open-weight models burn $100B compute to make $30B in revenue. Frontier models spend $100B to make $120B in revenue. One of these won't be sustainable source: Goldman conference

We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics. The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra. The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.

Still more absolute 🔥 from Terence Tao: “We have now seen that even the rumor of someone working on a problem can trigger a massive amount of AI-powered effort to flatten it before the original research project has time to reach its full potential. The incentives may now be pointing in the direction of no longer sharing any promising research directions with the broader community, which would reverse centuries of traditions of open science and do serious long-term damage to the future of the field”

GPT-6 Astra has the best result yet on the Bach Benchmark. Its chorale contains no voice-leading errors, and its harmonic palette is sophisticated enough to include a Neapolitan sixth chord. More importantly, it is the first model to ever write passing tones on this benchmark, a significant leap in musical understanding. Extra High effort. Prompt below: “In LilyPond (version 2.24), write a 4-part chorale in the style of Bach, 3/4 time, G minor. Use two staves -- soprano and alto on the top staff, tenor and bass on the bottom staff. Respond with only the code block.”

the silver-lining of hitting rock bottom after two garbage truck acquihiresitions where negative value was created is that there will simply be nowhere further down to go after a third