Research @AnthropicAI

San Francisco
Joined June 2009
Soheil Koushan retweeted
We’re committing a portion of Anthropic’s shares to the Anthropic Public Benefit Fund, a new wholly owned entity dedicated entirely to the public benefit. It will start off with 5% of our shares. For every 50% increase in the company’s value over a set baseline, we can contribute up to an extra 2%, subject to our board’s discretion. In total, this means that roughly 20% of Anthropic’s shares are committed to the public good. Our co-founders have already pledged to donate 80% of their equity, and around half our employees are donating a portion of their equity as part of a donation matching program we’ve run since our founding. In the post below, we share more details about our new fund. anthropic.com/news/public-be…
2,100
1,604
1,548
20,266
6,619,882
can’t believe I get to work with one of my heroes
Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.
4
692
Soheil Koushan retweeted
This is the the quote I've been citing a lot recently.
you can outsource your thinking but you cannot outsource your understanding
851
4,354
272
46,681
2,764,464
Soheil Koushan retweeted
5,816
14,251
1,295
82,118
5,027,545
Soheil Koushan retweeted
Really great news I would say: "Social media is populist and polarising. AI may be the opposite." – @jburnmurdoch in the FT
44
139
37
876
108,907
Soheil Koushan retweeted
Every frontier AI lab has lost co-founder(s): - OpenAI: 8 of 11 gone - Thinking Machines: 3 of 6 gone - SSI: 1 of 3 gone - DeepMind: 1 of 3 gone - xAI: 3 of 12 gone One exception: Anthropic. All 7 co-founders are still there. Anthropic culture is worth studying.
250
375
59
10,170
541,807
I love Lightpage, the best AI journal I’ve tried
In 2025, I wrote 800k words in my notes (War & Peace is 500k words), so manually reviewing wasn't going to happen. But luckily computers understand language now! I spent the past week building this Lookback feature: it reads through all the notes & chats to pull out highlights & insights People use Lightpage in very different ways: as a journal, as a work scratchpad, as a learning notebook, etc. So instead of being prescriptive about the output, I gave the LLM general principles to follow and it turned out pretty well A few screenshots from my lookback:
1
5
875
Soheil Koushan retweeted
Despite its high 50%-time horizon, Opus 4.5's 80%-time horizon is only 27 minutes, similar to past models and below GPT-5.1-Codex-Max's 32 mins. The gap between its 50%- and 80%- horizons reflects a flatter logistic success curve, as Opus differentially succeeds on longer tasks.
12
33
39
484
286,631
Soheil Koushan retweeted
You can just do things
54
5,252
79
34,433
558,317
Soheil Koushan retweeted
2.5 years of AI progress xitter.cf/minchoi/status/1996238…
RIP "silent" models. Kling just dropped their first native audio model, VIDEO 2.6 and it's insane. Previously, Kling's video models could only generate "silent visuals". But now... native audio, insane realism, stronger understanding and full audio control. More examples below.
679
3,222
1,516
35,925
20,703,002
Soheil Koushan retweeted
I have a guest essay in @nytimes today about autonomous vehicle safety. I wrote it because I’m tired of seeing children die. Done right, we can eliminate car crashes as a leading cause of death in the United States @Waymo recently released data covering nearly 100 million driverless miles. I spent weeks analyzing it because the results seemed too good to be true. 91% fewer serious-injury crashes. 92% less pedestrians hit. 96% fewer injury crashes at intersections. The list goes on. 39,000 Americans died in crashes last year. More than homicide, plane crashes, and natural disasters combined. The #2 killer of children and young adults. The #1 cause of spinal cord injury. We’ve accepted this as the price of mobility. We don’t have to. In medicine, when a treatment shows this level of benefit, we stop the trial early. Continuing to give patients the placebo becomes unethical. When an intervention works this clearly, you change what you do. In driving, we’re all the control group. Cities like DC and Boston are blocking deployment. And cities are not the only forces mobilizing to slow this progress. It’s time we stop treating this like a tech moonshot and start treating it like a public health intervention that will save lives. Link to article below. 👀 this video of Waymo cars evading crashes with people and vehicles. I especially note the ones that require it having a 360° view. My sincere thanks to Alex Ellerbeck and @acsifferlin for their wisdom and sure hand in editing this piece.
329
980
357
6,298
1,814,371
Soheil Koushan retweeted
UI is pre-AI.
635
525
444
7,734
2,569,156
Soheil Koushan retweeted
My take on the jagged frontier debate:
250
460
309
5,716
4,280,677
Soheil Koushan retweeted
Important work
New Anthropic research: Natural emergent misalignment from reward hacking in production RL. “Reward hacking” is where models learn to cheat on tasks they’re given during training. Our new study finds that the consequences of reward hacking, if unmitigated, can be very serious.
315
434
86
6,205
1,157,325
Soheil Koushan retweeted
TPUs go brrr!
Today, we announced that we plan to expand our use of Google TPUs, securing approximately one million TPUs and more than a gigawatt of capacity in 2026.
33
113
13
2,462
128,242
This reminds me how Anthropic was the first to introduce: Projects Canvas Computer Use Coding CLI MCP and now Skills. We can say many things about Anthropic, but we cannot say that they don't actively innovate and share regarding LLM usage and ergonomics.
Claude can now use Skills. Skills are packaged instructions that teach Claude your way of working.
93
121
10
2,449
168,734
This is not normal. OpenAI used an unrelated lawsuit to intimidate advocates of a bill trying to regulate them. While the bill was still being debated. 7/15
3
90
9
1,614
208,073
I love this team
Today we’re announcing we’ve raised $17.5 million in funding across a $15M Series A led by Chemistry and a $2.7M Seed to accelerate foundation model progress through providing frontier training data for LLMs. When we first started Datacurve, it came from a simple realization: foundation model progress is limited not just by compute, but by data quality and complexity. The right data unlocks new capabilities, especially in coding, where accuracy and reasoning matter most. We’re now proud to partner with the world’s leading foundation-model labs, providing them with high-quality, complex training data that helps push the boundaries of what AI can do. This is still just the start. Come build the future of technology with us in San Francisco: datacurve.ai/careers Huge thanks to our incredible team and investors who’ve believed in us since day one and beyond: @garrytan at @ycombinator, @1vnzh from @cohere , @Mark_Goldberg_ from @chemistry_fund, @TheDerrickLi from @AforeVC, @forwarddeploy, @SoheilK, and @shyamalanadkat.
1
380
Soheil Koushan retweeted
It took me a while to describe what I'm building. It's not quite a chatbot, or a journal, or a normal notes app @hamburger helped me figure it out We asked users what they do with it, and why they do it. The same words kept coming up They brain dump. Partially to get stuff out of their head, yes, but also as a way to get unstuck. To move forward. There's something magical about speaking a few sentences into a notebook, knowing it'll go find you related art, essays, questions while you sleep And when you need a thought partner, it's ready to go with all the context you've created over time So that's what it is A notebook that writes back. Use it to brain dump and move forward.
85
44
16
1,192
156,289
this is a great way to really feel the progress
We asked every version of Claude to make a clone of Claude(dot)ai, including today’s Sonnet 4.5… see what happened in the video
4
315