AI watcher with an analyst’s eye. Measuring how AI agents perform at work. Tech takes. Games for fun
Joined August 2016
- Tweets2.9K
- Following941
- Followers246
- Likes5.9K
Tony Stark barely wrote a line of code. He told JARVIS what he wanted, then tested the suit by flying it.
That’s vibe coding, plus the evaluation step most people skip.
23
When you pay for a premium AI, how do you know a cheaper one didn't answer?
@AnthropicAI has built a way to prove it, mathematically.
Each answer comes with a small proof showing which model wrote it, without revealing the model itself. Think certificate of authenticity, not watermark.
The only earlier attempt on a model this size produced a 93.6 GB proof that took about a day to check.
Anthropic's FlashProve proofs are about 19 MB and check in seconds.
The catch: it still costs 70–95x more than just running the model, and only works for short answers.
Would you pay extra to know which AI really answered you?
1
2
56
Amanda Askell shapes how Anthropic's Claude behaves. She isn't a computer scientist. She's a philosopher.
She trained in philosophy at Oxford and NYU, after starting out in fine art.
She's the lead author of the constitution that shapes how Claude thinks and handles hard questions.
Labs are spending billions on chips. One of the most important jobs went to someone trained to ask what's right.
Would you trust an AI more if you knew who shaped its values?
1
2
97
In the race for compute, safety research has become an afterthought.
Anthropic is going the other way, training new researchers to work on it full time.
Should the other labs be spending as much on safety people as they do on GPUs?
Applications are now open for the Anthropic Residency, a paid, fixed-term research program for academics, technical founders, and staff-level engineers who want to make AI safe and beneficial for everyone. No ML background is required.
Apply by November 22: alignment.anthropic.com/2026…
1
24
Hundreds of models, zero humans
Glif's (@heyglif) new AI modeling agency casts AI-generated models and creates the commercial end to end.
The hard part won’t be generating them. It’ll be making them resonate.
A commercial has to make you feel something, and the person on screen isn’t real.
Would you buy a product advertised by an AI model instead of a human one?
🎥 @heyglif
1
6
312
Elon Musk has an “Idiot Index”, revealed in Walter Isaacson’s biography of him.
I just finished reading the book and I can’t stop thinking about it.
It’s the cost of a finished part divided by the cost of its raw materials. If a part costs 10x its materials, someone has some explaining to do.
A physics degree tells you what something is made of. An economics degree tells you what it should cost. He uses both every time he looks at a part.
At least someone put their university degree to use.
What’s the idiot index of your job’s tools?
4
76
Google just ran an AI prompt in orbit, and the first request was a haiku.
They just rang the opening bell on the race for data centres in space.
Will we have a fully functional data centre in space by 2027?
A ‘hello world’ milestone for Project Suncatcher 🚀
After launching our prototype satellite a few weeks ago, we’ve now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth - a haiku!
3
43
The Future of Design youtu.be/xNqs_S-zEBY?si=OKK1… via @YouTube
We got a Trillionaire YouTuber before GTA 5 and GTA 6
72
Fired on a Friday. Back in charge within days.
Andrew Garfield plays Sam Altman in Artificial, a new film about the 2023 OpenAI boardroom drama. I expect it to make that comeback look less like luck, and to humanise him a bit.
Will it do for Altman what The Social Network did for Zuckerberg?
1
142
Every brick on this building was laid by @BuildMonumental's robots.
A factory is the same part in the same place, every time. A building site is curved walls, different bricks and uneven ground.
Humans still finish the job. The question is whether humanoids take that part next.
Robots building with robots, or robots building with people?
Our robots built this in the city of Almere: curved walls, different types of bricks and bonds, indented text.
Real world deployment, insane complexity, 100% of every brick done by our robots, with another 15 houses to come in the coming weeks.
1
1
62
OpenAI is using its large consumer business and growing enterprise business to fund heavy discounts.
Anthropic reportedly ends enterprise discounts (about 15%) once committed tokens run out: renegotiate or pay list price.
OpenAI gives customers extra time and more aggressive discounts.
Some customers are moving. Replit, until recently one of Anthropic's top five customers, has made OpenAI its default, citing cost and performance.
My read: OpenAI can afford it. A $122B round this year, a large consumer subscription base, ads that hit a $1B run rate in under 200 days, and annualised revenue near $50B (net).
Can enterprise-first AI companies compete on price with one that has a billion users behind it?
2
1
59
I just put a baby dragon in my living room.
@Simulon uses your current environment and puts something that isn't there to your room.
It is different from Genie 4. Genie creates simulated environments from scratch.
You can also/ turn the image into a video with a prompt and AI rendering then blends it into the scene with lifelike lighting, shadows and reflections.
What would you put in your living room?
🤖 Made with AI
Things that aren’t there can now see and hear you.
Reality-Aware Rendering is now live in Simulon. Place anything in the real world, act out the scene, and it responds to you.
We made this short in an afternoon. Watch to the end to see what was really there.
Available now on iPhone. How it works 👇
6
320
At the 2026 Robot Warrior Challenge held on 17–18 October in Yizhuang, China, wheeled four-legged robots crossed water and jumped pits. A humanoid fell over on a ramp.
Still, the progress is wild. In 2025, humanoids kept falling over in Beijing’s first robot half-marathon. In 2026, one beat the human world record.
The real challenge is limited real world training data on folding, gripping and cleaning.
What will a home robot do well first: cooking, cleaning, or folding laundry?
36
Fascinating work from @GoogleDeepMind
Genie 4 is a significant leap over its predecessor, with more human-like behaviour, native audio and higher visual fidelity.
The real breakthrough? Real-time interaction combined with Gemini-powered reasoning—enabling AI to interpret scenes, plan actions and adapt as environments evolve.
This opens up enormous possibilities for experimentation at scale.
Potential applications span robotics, autonomous vehicles, gaming, filmmaking and immersive learning.
We’re moving beyond AI that generates worlds towards AI that can interact with and reason within them.
Could world models be one of the missing pieces in the pursuit of AGI?
This is Genie 4, our most realistic and interactive world model yet. 🌐
Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.
1
1
36
Mistral wants to own the whole AI stack: models, data centres, cloud.
Its target is 200 MW across Europe.
Big Tech will spend $700B+ on capex this year alone.
The TWIST: Microsoft is now buying compute from Mistral in a multi-billion dollar deal signed this July.
When hyperscalers fund AI labs, they usually take equity. This deal added none. Microsoft keeps only the small stake it bought in 2024.
That leaves Mistral owning the data centres Microsoft is helping fund.
So, Microsoft is its competitor, its distributor and its customer at the same time.
Can you build independence on infrastructure your biggest rival helps pay for?
The next contract renewal could be interesting.
23
Since the S-1, a lot of the discourse has bordered on ridiculing @AnthropicAI .
But read the risk factors. They're unusually blunt: Anthropic admits it has "significantly underestimated" its compute needs before, and warns that capacity could end up idle with cash already committed.
Revenue isn't the problem: $16.3B in H1.
The bet is ~$450B in compute commitments.
Did Anthropic pick the right moment to go public?
Anthropic Warns of Existential Risks, Liability and Slowing Sales bloomberg.com/news/articles/…
36