chuplung

Anthropic's platform team on a counterintuitive idea for building AI agents: not all tokens are equal. everyone's lever is the same - spend more tokens, get a better result. they asked: what if you give tokens jobs instead? so they split the work. some tokens execute. some advise the executor. some grade it against a rubric. some "dream" - review past runs and write lessons to memory for next time. then the key test: hold the token budget fixed across all of them. if tokens were fungible, every strategy should score the same. they didn't. on a financial-analysis benchmark, plain executing hit a perfect answer 42% of the time - the smarter strategies, up to 75%. the kicker on cost: to brute-force one perfect answer, executing burns ~1.8M tokens. advise and grade get there for a fraction - same budget, better jobs. ~15-min talk, free. Anthropic on why *how* you spend tokens beats *how many* ↓

Jess Yan (Claude Managed Agents, Anthropic): "we set agents tasks overnight. we wake up and the backlog is resolved and bugs are squashed." she also said: "I talk to Claude more than I talk to my colleagues." one example: 4,000 companies on a waitlist full of duplicates. she spun up an agent in 30 minutes - it cleaned the list, scored each company, sent daily invites to the best ones. "the limit is no longer our personal capacity. it's how much we can delegate at once." bookmark this ↓

Dario Amodei (CEO of Anthropic) said something nobody wants to hear. "we could have 5-10% GDP growth and 10% unemployment at the same time. never happened before. but it's not logically inconsistent." high GDP always meant lots of jobs. AI breaks that assumption. also from the interview: → Anthropic revenue: $0 → $100M → $1B → $10B in three years → co-work built in a week and a half almost entirely by Claude → "software is going to become cheap. maybe essentially free" bookmark this ↓

Jack Clark (co-founder of Anthropic) thinks Claude will start training itself by 2028. "Claude 10 builds Claude 11. It designs the architecture, does the research, runs the training. We step back entirely." what that means: → last 5-6 years of AI progress compressed into 2-3 years → then compressed again → humans out of the development loop his 7-month-old will be in kindergarten when this happens. bookmark this ↓

Anthropic added Claude to Slack. 65% of all PRs in their product org are now written by it. Boris Cherny (creator of Claude Code) has a Tag session running for a month. every day it checks data, fixes bugs, opens PRs. he just watches them come in. what makes it different from Claude Code: → you don't open it. it's already in the channel watching → multiplayer - whole team guides it, not just one person → remembers everything. tell it once, it never forgets → self-schedules work days or weeks out "I just got tired of tagging it. so I told it to always respond. now it just has my back." bookmark this ↓

Greg Brockman (co-founder of OpenAI) on the future of AI interfaces: "you want almost no interface. you want no product. just talk to something that goes and accomplishes goals for you." buttons, modes, toggles - that's the machine forcing you to speak its language. the goal is the opposite. two other things he said: → 230 million people use ChatGPT for health questions every week. patients doing what doctors used to gatekeep → his wife has several health conditions. says he doesn't know how they'd manage without AI "bring the machine closer to the human. not the human to the machine." bookmark this ↓

A videographer asked Codex to edit videos in Premiere Pro. Codex couldn't - so it built itself a Premiere Pro extension, installed it, then used it to do the edits. nobody told it to do that. Andrew Amersino (Codex product lead, OpenAI): → 90% of all OpenAI uses Codex. not just engineers. everyone → "implementation is no longer the expensive part. it's taste" → same app in November would have failed. only the models changed between then and February "the job is no longer building. it's curation." bookmark this ↓