Pinned Tweet
hi, i’m dima. lived in finland for years, now in sf.
i write sci-fi short stories and co-founded an ai coaching company before chatgpt was a thing.
i’ll mostly post about what i’m building and why running ai costs so damn much.
if you’re building too, say hi
seems like OpenAI's new models only look good on the benchmarks
in irl GPT-6.2 Astra / Sol are far behind Opus 5.5 and Fable 5.2
and it's more expensive as well, GPT-6.2 Astra is 100% more expensive than Opus 5.5 and costs $0.20 more per million tokens than Fable 5.2
Some of you have been asking very politely. Some of you have just been asking.
GPT-6.2 Astra and Sol are now available.
Super efficient and built to perform where it counts—the real world.
openai.com/index/introducing…
1
1
71
Dima Syrotkin retweeted
anthropic literally admitted in its ipo filing that it has significantly underestimated its compute needs in the past
dude they have $900b in compute commitments and dario is telling jamie dimon they buy less than they think they'll need and pay spot prices if demand goes up
so yeah even they're guessing as well
anthropic literally admitted in its ipo filing that it has significantly underestimated its compute needs in the past
dude they have $900b in compute commitments and dario is telling jamie dimon they buy less than they think they'll need and pay spot prices if demand goes up
so yeah even they're guessing as well
Dima Syrotkin retweeted
Everyone is building sandboxes.
But a sandbox is only 10% of the infrastructure you need to run AI agents. You're left building the other 90% in-house: deployment, persistent state, monitoring, recovery, scaling
You're trying to sell agents. Somehow you're building a cloud.
Today we're launching Maritime out of YC F26 🚢
Maritime is a managed cloud for AI agents. We handle the infrastructure so you can focus on your product.
Dima Syrotkin retweeted
i see that many sites are now blocking the ai agents and its so fucking insane
i my opinion the ones that let agents in will end up as the default source, the ones that wall off will just quietly disappear from how people find things
the is the agentic era if you dont let agents use your site, you're gonna have a very bad time in the future
i see that many sites are now blocking the ai agents and its so fucking insane
i my opinion the ones that let agents in will end up as the default source, the ones that wall off will just quietly disappear from how people find things
the is the agentic era if you dont let agents use your site, you're gonna have a very bad time in the future
fable 5.2 and opus 5.5 aced every task in the new stack's coding tests, and fable cost almost 4x more.
but they ran at api defaults, and those aren't equal btw, fable runs on high effort and opus on medium unless you change it.
opus 5.5 gotta be the best model released this year
2
8
425
this may be true but the reason i think why they wanted him to go is that he was selling gpu's to china and the american government wanted it to be stopped
and traces of nvidia's blackwell chips were also found in the chinese ai models
this is a bigger reason than his support for open-source models
9
633
Dima Syrotkin retweeted
things you need to run opus 5.5 locally
- 72 blackwell gpus
- 13.4tb of gpu memory
- ~120kw of power, which is basically a gb200 nvl72, so
- liquid cooling
- $10k a month, which is roughly just the electricity for - 87,600 kwh
- a landlord who doesn't ask questions
you don't realise what it takes to run a model this size
things you need to run opus 5.5 locally
- 72 blackwell gpus
- 13.4tb of gpu memory
- ~120kw of power, which is basically a gb200 nvl72, so
- liquid cooling
- $10k a month, which is roughly just the electricity for - 87,600 kwh
- a landlord who doesn't ask questions
you don't realise what it takes to run a model this size













