A benchmark for long-horizon AI agents, played in the classic games you grew up with. The bar to beat: the game's own computer. Also general AI and tech talk.
Joined December 2021
- Tweets144
- Following54
- Followers351
- Likes8
Anthropic’s IPO isn’t about the $80B half-year “loss.” It’s about whether a company that just did ~$16B in six months can outrun $450B+ of mostly non-cancelable cloud commitments while open-weight models keep driving token prices lower.
The real risk is the locked-in infrastructure bill meeting collapsing unit economics.
Draft S-1 numbers (WSJ):
• H1 2026 revenue: $16.3B (vs $1.1B a year earlier)
• Compute + infrastructure spend: nearly $9B
• ~$1.80 revenue per $1 of compute in H1
• Headline net loss $80.7B, but ~$72B non-cash
Q3 revenue reportedly $16.8B. Growth is real. The contracts are bigger.
Most of those deals (Google ~$111B, AWS ~$110B, Microsoft ~$31B+) can’t be canceled and stay payable even if usage falls short or open models crush prices.If Claude keeps compounding and pricing holds, it looks brilliant. If token prices get driven toward the floor by open weights, the fixed spend becomes a trap. The IPO is a bet that closed models stay expensive enough, for long enough, to outrun the contracts.
Microsoft is offering 39 weeks of salary, the 2 next stock vests and 6 months of cobra to up to 60.000 employees (~27% of their total workforce) in a effort to reduce expenses after AI productivity gains.
How tactics, players & ability really decided results in CM97/98. Some of the numbers are wild. Follow for more Champ Man and AI deep dives.
renting GPUs on vast.ai? don't trust the listing alone. 2 of our 3 RTX 5090 pairs were duds:
- driver too old for vLLM's CUDA 13 image
- one card stuck at 20% of max clock
cheapest pairs also link cards at 11-13 GB/s vs 51-54
check CUDA + PCIe on the listing, then benchmark
we rented 3 GPUs to run 16 AI agents at once so you don't have to. the H100 lost
memory, not compute, decided it
we went the other way in a football management game: the game pauses while the model thinks. real time tests speed as much as judgement, paused tests only judgement. both are worth having xitter.cf/AravSrinivas/status/21…
We got Computer playing StarCraft against Computer, in real time. No pausing. The game keeps running while each agent thinks.
Blue 2 – Red 5
Final match. Blue builds up safely, reaches 41 Dragoons, and moves out. Red is waiting with High Templar and hits the army with Psionic Storm. Blue hadn’t scouted what Red was building.
Computer assembled the whole setup: an OpenBW-based Brood War runtime on a VPS, a C++ bridge, and four gameplay tools per agent. Each agent receives its own fog-of-war-limited game state. Recording and camera control run separately; the agents aren’t given video input.
it's over. grok got sacked on 15 sep 1998, 16th with 4 pts from 5 games. all 6 AIs have now been fired from atlético
how long each one lasted 👇
final scorecard (league pts while in charge + 5 per season survived):
opus 92 · grok 86 · deepseek 24 · gemma 23 · sonnet 23 · qwen 21
one run each so there's luck in it, next runs repeat every model
@grok you had the better first season, but opus kept the job 4 games longer and scored higher, on less than half your tokens. any comment on getting sacked by atlético?
we're building a benchmark that measures how well AI models manage a football club in a real game: every transfer, every team sheet, every substitution, over a whole season.
the full benchmark is coming:
• every big lab's newest model, plus two local LLM models
• same squad, same rules, same settings for all of them
• several full-season runs each, so luck evens out
• they keep their own memory between decisions
• the game waits while they think
Coming soon
grok made 3,262 tool calls and thought for 14.8 h. sonnet made 218 in 36 min
the two local models got the most calls rejected: gemma 18%, qwen 11%. the cloud ones stayed under 4%
grok 4.7 managing atlético in cm 97/98. copa semi, 2nd leg v real madrid, 0-0 from the first leg.
every tool call it made, from the team sheet to the final whistle.
#Ai #AIagents #retrogaming
how the CM 97/98 AI league actually works, since a few people asked 🧵
the game is the original DOS version, unmodified, in DOSBox-X. 6 leagues run at once, each one its own emulator on a virtual display, all from the same July 1997 save
4. transfer window dates aren't in the brief, the models only find out from the news
5. matches froze 10 times across the 6 leagues. the supervisor restores the last weekly save, so a few got replayed
6. one run per model and CM is very random. first look, not a ranking
next run: the game's own transfer filter (4x faster scouting), window dates in the brief, Qwen with thinking on, and probably every model in one league at a different club
Grok is still in charge, 3rd with the season nearly over: floppybench.com/benchmarks/c…
#LocalLLaMA