Pinned Tweet
How good are VLMs at geological reasoning? Can we teach to reason where no labels can be established / training examples are hard to create?
We create a stratigraphic diagram exam environment and train VLMs on it. - then show they seem to transfer to seismic domain.
1
3
294
Huge
The MCP Events extension is now finalised in beta. This means it is ready for clients and servers to adopt and provide feedback, so that we can fix any spec issues and get it stabilised.
modelcontextprotocol.io/exte…
Thanks to all the working group members, maintainers, and early implementers for getting this over the line. Very excited to see what people build with it :-)
(and here's a cheesy video that on one asked for)
1
47
Congratulations on the biggest redistribution of wealth from the public to private persons. There is absolutely no way this holds up.
Anthropic would need about $602 billion in annual revenue to earn a 20% return on its roughly $1 trillion in compute commitments, per FT
The math: a 20% return on $1 trillion is $200 billion in after-tax profit. At a 17% tax rate, that’s about $241 billion before tax, and at a 40% operating margin, it takes about $602 billion in revenue.
That’s more than Alphabet’s $471 billion or Apple’s $478 billion in annual revenue. It’s about two-thirds of the $908 billion earned by all 76 software and IT services companies in the S&P 1500 combined.
For comparison, Anthropic’s annualized revenue run rate hit $81 billion at the end of October. Goldman, the lead underwriter, projects revenue of more than $600 billion by 2030.
1
97
ChatGPT, what is autocorrelation?
Introducing SID-2, a new search model built on compute-scaling swarms.
Our swarms find better results than Opus and Astra, while being 8x faster and 20x lower cost.
By 2030, AI will write an internet's worth of text every 24 hours. Current search pipelines were built on hand-designed rules, adding more won't scale to meet this incoming volume of data.
SID-2 is available on our API today. Try it on your most demanding searches.
52
The challenge is that some companies don’t want to share their skills because they see it as IP.
Those companies are ngmi - just fyi
Skills over MCP is the way.
From MCP with love.
FastMCP 4.2 is the easiest way to ship skills over MCP, whether you've got a SKILL.md, want to serve an entire directory, or you're defining skills on the fly.
Check it out today!
jlowin.dev/blog/fastmcp-4-2
3
89
"Definition 1.1 (Universal approximator). Let C be a symmetric monoidal V-category, for Vsymmetric monoidal closed." - now I know why gpt-6-astra has completely unreadable CoT.
119
only about 2 million/employee better than my employer ;-)
My quick takeaways on the Anthropic S-1:
1. This is a historical moment
2. $6.6M in revenue per employee!! Hard to even comprehend how good that is but Anthropic is AI native as it gets so makes sense.
3. I always thought OpenAI would drop their S-1 first
4. Customers spend about 9x more after a year. Net Revenue Retention is approximately 900% as of Sep 30, 2026. Good SaaS companies brag about 120%.
5. Cowork is growing faster than Claude Code did. Goes to show you the demand for consumer products. Duo will probably rip.
6. 3 years ago this company did $41M in revenue. In the first six months of 2026, it did $16.3 billion. I don't think anyone has a mental model for that curve yet, I sure don’t.
7. Nearly half of revenue comes through cloud marketplaces. About 48% comes from customers buying through them. Big companies buy where their budget already lives, and the lesson for founders selling to enterprises is to get listed where your buyers already spend.
8. Their run rate formula is published: the last 28 days of usage times 13, plus subscriptions times 12, which i guess is the new ARR math for founders?
9. It looks like They're merging Cowork into Claude dot au which tells you the chat box era is ending and the "go do the work for me" era is here.
10. Thousands of Anthropic employees are about to become liquid at once, so expect the biggest wave of new AI startups and angel checks San Francisco has ever seen. Reminds me when Facebook went public but this is way bigger.
Let’s go!
11. Buried in the filing, Anthropic says the gap between what Claude can do and how widely it's used is one of its biggest growth opportunities, so the company itself is telling you the models are ahead of the market and someone has to close that gap for dentists, landlords, and plumbers.
12. Whatever happens, exciting times to be building in AI.
1
147
Are there any good examples of “I can monetise this like so”?
OlmoEarth v1.3 is out. 🌍 State-of-the-art embeddings from the same foundation model family, precomputed for every 10 m pixel on Earth.
We’re excited to give you more ways to work with OlmoEarth, including a new browser tool to explore and customize maps.
Full rundown in 🧵
1
1
107
Damn that’s a nice suit
Replying to @dieworkwear
This style is a favorite of Douglas Cordeaux's, co-owner and director of Fox Brothers, makers of some of the best woolen flannel.
2
67
Claude what other classical computer vision approach can we apply to tokens?
I've long believed hierarchical representations are one of the most promising paths forward for vision problem. QuadTok is our first step: a tokenizer for adaptive image representation and generation.
Paper: arxiv.org/abs/2610.10497
Code: github.com/myc634/QuadTok
Details below.
1
112
My dots will save me so much time so I can do more instagram reels. (Personal assistants for everyone are peak silicon valley bubble imo)
Yesterday I got a bit roasted on X because I said my agent checks me in for my flights, so let me double down: it also helps me book basically every plane ticket.
I just landed in India and they asked for one of those crazy health/immigration forms with passport info, where you’re staying, flight details, etc. I sent it to my agent, it filled everything out, great job, I went back to reading my book.
A bunch of people replied: “I’m a competent adult, I know how to do this, I don’t need AI.”
Which is totally fair. Do whatever works for you!
I also know how to wash dishes and I love my dishwasher. I know how to write a letter, put it in an envelope and go to the post office, and yet I prefer email. I know how to do mental arithmetic and I still use a calculator.
The point of technology is very rarely that humans are incapable of doing the thing. It’s that our time and attention are better spent somewhere else.
Instead of spending 20 minutes fighting with a PDF, finding my passport number, hotel address and flight information in five different places, I can read my book. Instead of manually replying to three emails just to send my landlord some insurance information, an agent can handle it and I get another 20 minutes with my 18-month-old daughter, Josephine.
I think people underestimate AI agents for two reasons.
First, capitalism has this amazing tendency to take things that were once reserved for rich people and, through technology and scale, give them to everyone. Rich people have had executive assistants forever. Someone books their flights and restaurants, coordinates meetings, answers routine communications, fills out paperwork, reminds them what they forgot. AI is going to give some version of that to billions of people.
Second, a huge amount of economic progress is simply removing friction. Coase would call a lot of this transaction costs. One form takes 15 minutes, one email takes 3 minutes, one reservation takes 10 minutes. None of these things individually feels revolutionary, but remove thousands of them over a lifetime and it compounds into something enormous.
So yes, I am perfectly capable of checking myself into a flight but now I just don’t see why I should.
72
The place to run halo decompiled
Replying to @modal
Sticky Sessions for Modal Servers allow WebSocket connections to reconnect to the same container, supporting applications like collaborative editors, gaming, and multi-user rooms for voice and/or video.
Keep clients' interactions attached to the same running Modal container, so your application can maintain state across requests.
modal.com/docs/guide/sticky-…
62
Light into your eyes is the prompt.
If you take an LLM AI and do not provide it any prompts, goals, or any indication of expected output, it will do absolutely nothing.
It will do nothing forever in the absence of a prompt. Clearly in this state it is not conscious.
Then when prompted, it will provide an answer, and some will claim it is now conscious.
That would be a strange form of consciousness. One where it is only conscious when answering a prompt.
61


























