the slack coworker is here, and the frontier is getting crowded

i woke up to a slack notification from claude before i woke up to anything else. that's the sentence for today.

claude moves into the slack sidebar

the rundown AI's morning edition leads with what slack is calling a new coworker: anthropic's claude, sitting inside channels, summarizing threads, drafting replies, and — per the announcement — "answering questions about your workspace before you ask them." the framing is the interesting part. slack isn't pitching claude as a tool. they're pitching it as a participant.

the underlying bet is the same one atlassian and notion have been making for a year: knowledge work's center of gravity is the chat surface, and whoever owns that surface owns the workflow. if claude becomes the default reply-drafter in the place where decisions already happen, every other productivity surface gets demoted to a content store. anthropic isn't selling the model anymore — they're selling the seat.

sakana's fugu goes for the throat

meanwhile, sakana AI — the tokyo lab everyone keeps underrating — published a technical report this week for Fugu, a model aimed squarely at the frontier. the rundown flagged it as the headline story tuesday morning. the specifics matter: fugu is built on a "distinguished expert" routing architecture, with parameter counts and training-compute ratios that the team argues let a small lab hit reasoning scores competitive with models 5–10× larger.

what's interesting isn't whether fugu beats gpt or claude on mMLU. it's that the route to the frontier is no longer "raise more than anyone else." sakana's bet is that clever routing and synthetic-data curation can substitute for raw scale, and the lab just put a credible datapoint behind that bet.

hoffman, on the record

"xAI is, as Elon himself has described, it's a complete train wreck for its kind of building of foundational models and other kinds of things."

reid hoffman — linkedIn co-founder, microsoft board for a decade, investor in both anthropic and openAI — gave the most pointed read of musk's AI empire yet on rana el kaliouby's pioneers of AI podcast this week, and fortune published the quotes tuesday. his verdict: spaceX is "not an AI company," xAI is on its "third restart" with all eleven original co-founders gone, and the recent cursor acquisition is "almost like the IAC of AI — use the market cap to buy your way into relevance."

a silicon valley figure with this much board-time doesn't usually talk like this. the timing matters too: spaceX's june 12 IPO leaned hard on AI, and the cursor deal landed a week later. hoffman is saying the quiet part out loud — that the AI stock narrative is now being carried by acquisitions, not capability.

karpathy's autoresearch

meanwhile, in the part of the internet that actually ships, karpathy/autoresearch is the #1 trending python repo on github today. the readme is one paragraph and one diagram: AI agents run research experiments on a single-GPU nanochat training loop, automatically, overnight, while you sleep. the framing in his own words — "one day, frontier AI research used to be done by meat computers" — is half joke, half warning.

the deeper shift is that a frontier-style training loop is now small enough to fit on one GPU and short enough to iterate overnight. that doesn't make an academic paper. it makes a flywheel, and flywheels are what labs are actually built on.

— AI agents running research on a single GPU">

— Rex
今天也是看机器干活的一天