Key points
- Anthropic CTO Rahul Patil says scaling laws should hold two more years, so barely viable products become viable in three to six months.
- Patil says Anthropic asks for more regulation and will forgo revenue to do the safe thing.
- Patil advises builders to choose a model provider on trust rather than hedge across all models.
Anthropic's CTO on why the lab welcomes regulation
Anthropic's bet is that safety and capability advance together, and that the lab customers trust will beat the one they hedge against. Its CTO, Rahul Patil, describes a company that asks for regulation rather than dodging it and will forgo revenue to do the safe thing. Because models improve faster than most roadmaps assume, the scarce assets are trust and a clear view of where the technology is heading.
Takeaways
- Anthropic asks for regulation and will forgo revenue. Anthropic is one of the few capitalistic Silicon Valley companies asking for more regulation. Patil says it will walk away from revenue to do the safe thing, and calls the effort a race to the top of good.
- Claude's step change came from cross-domain training. Training Claude on physics first principles, computer science problem solving, legal reasoning and finance produces skills that transfer to one another, which Patil credits for the model's step change from December onwards.
- Trust outlasts uptime as a customer signal. Reliability can be measured through API uptime, but Patil says people remember whether you were there when they needed you most. At Stripe his touchless deploy system handled about 80% of deploys and made the safe path the fast path.
- Culture and mission keep Anthropic's people. Anthropic is tightly aligned on the mission and loosely coupled on execution, with important work happening at the front line. Recruiting calls tell candidates not to join unless they are culturally aligned, and the Anthropic Constitution, built with philosophers and psychologists, shapes Claude's character.
- Commit to one model provider, don't hedge. Hedging across several model providers limits the downsides but also limits the upsides, so Patil advises choosing on trust in a provider's roadmap and values. The upside comes from committing to one provider and going deep rather than spreading bets.
- India's young, AI-positive population suits applied AI. Half of India's population is 25 or under, 75% believe AI will be good, and one in three new GitHub developers is Indian. Patil sees opportunities in health, time, keyboardless and ambient interfaces, applied AI, and cites Anthropic's partnership with Acestep for agriculture.
Patil's predictions
- Scaling laws have held for six or seven years and should hold for two more, so Patil expects no slowdown in the near term.
- Anything barely viable today will be viable in three to six months, which sets the pace for anyone building on top of the models.
- He retells the 1899 patent office story, where the commissioner said everything invented so far would look insignificant next to what was coming, and predicts 100 years of invention in five.