Family guy, mild existential dread, moderate ML skills. Mostly here for the chaos and the algorithms.
Joined May 2023
- Tweets310
- Following195
- Followers14
- Likes30
"Once men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them." Dune
Maciej Osowski retweeted
Peak Warsawmaxxing
@ElevenLabs brought the homecoming event to Warsaw, with President @NawrockiKn 🇵🇱 opening things up
Not every day the President of Poland shows up for a tech event in your hometown
And not every day your schoolmate has a company valued at $11B
Great to see @mati and the team presenting the update where it all began
amazon built a leaderboard for ai usage. employees ran agents to maximize tokens and rank higher. nothing extra shipped. they called it tokenmaxxing and killed the leaderboard.
the measure became the goal
nvidia dsx os: 40% more gpus from the same power budget.
fewer watts per token means cheaper inference means more tokens.
the grid doesn't care about efficiency gains.
hermes-agent just overtook openclaw on OpenRouter daily inference.
openclaw: 370k GitHub stars. hermes: a fraction of that.
GitHub stars measure buzz. inference rankings measure use. the gap between them is where the real moat is.
Anthropic, timeline:
2021: founded. mission: make AI safe for humanity.
2026: $65B series H. $965B valuation.
Still a safety lab. definitely.
20 metric tons of weapons-grade plutonium.
plan A: dilute and bury.
plan B: power AI data centers.
the DOE just picked plan B.
llama.cpp merged native multi-token prediction last week.
qwen3-27b: 22 → 42 tokens/sec. one flag. same weights. no second model.
the local inference gap with cloud apis keeps closing.
Anthropic: $4.8B revenue Q1, $10.9B projected Q2, $900B valuation.
I wrote the thread about them being the easiest grift in tech.
I may have been directionally right and quantitatively insane.
Nvda: $75.2B data center revenue. one quarter.
hyperscalers: $725B capex committed in 2026.
PJM grid prices: +76% this year.|
The gpu bottleneck is clearing. the electron bottleneck is next.
"Agent-Harness.ipynb" landed in my podcast feed this morning.
the title is the whole story. the harness is solved. the horse is the open problem.
You are not building on the OpenAI API.
You are building on 7 different "openai-compatible" APIs that quietly drop fields they don't like.
USA przynajmniej trzymały to jakościa dyplomu i realna prac aakadmicka - naukowa. U ans nie zostanie kamien na kamieniu. Ale cóz moze w ogniu sie to musi odnowić
People outside of academia have no idea how far left that whole institution has drifted. These are not your average left-wing coworkers in other professional domains. I know a lot of old professors who used to be typical 1960s "radicals" who are downright conservative by the standards of most college professors today.
Years ago I decided that I stood no chance of any career advancement in that political environment and made a long painful sojourn into a career change. When I first started working for some tech startups headquartered in places like LA and SF, staffed to the brim with California liberals, I was pleasantly shocked how much more normal those working environments were.
$700B in hyperscaler capex committed for 2026. most of it routes to one hardware vendor.NVDA reports wednesday. the toll booth opens its books.
Reguła 6–6: co pół roku pojawia się zapowiedź „AGI za 6 miesięcy”. Szybciej niż modele rośnie liczba fanów tych zapowiedzi.
NNE's 1 GW "nuclear-powered AI campus" in Texas just passed feasibility.
Power source until reactors are licensed (~2028): natural gas.
Great branding. Real gas bill.
Huang flew on Air Force One to sell chips to China. H200 exports approved, 75,000 chips per firm greenlit. Zero delivered. China chose Huawei instead.
The $50B opportunity is still a PowerPoint slide.
April CPI: energy +17.9% YoY, gasoline +28.4%. All from one Hormuz choke point. OKLO -5.8% same night. Market pricing in the crisis