Anthropic co-founder on AGI timelines, safety, and preparing for transformative AI

Define transformative AI by its measurable economic impact, not human-equivalence.

Use an "economic Turing test": if an agent can be hired for 3 months and later revealed to be a machine, it passes for that role. A market-basket of jobs weighted by money is used; once 50 % of such jobs are automatable, we have transformative AI.

Why: Societal institutions are sticky; GDP growth >10 %/yr would signal a new era.

"I like the term transformative AI because it's less about like, can it do as much as people do? ... more about objectively, is it causing transformation in society and the economy?"

50th-percentile superintelligence arrival is ~2028 based on scaling laws and hardware trends.

Extrapolates 2× annual CapEx growth ($300 B today → trillions), continuous scaling-law validity, and post-training improvements.

Why: Multiple converging exponentials (compute, data, algorithms) make short timelines credible.

"I think 50th percentile chance of hitting some kind of superintelligence is now like 2028."

X-risk probability is 0–10 %, but marginal impact of safety work is extremely high.

Even a 1 % airplane crash risk would make people reconsider flying; analogous caution applies to humanity’s future.

Why: Reference classes for X-risk are scarce, so small probabilities still warrant massive effort.

"my best granularity forecast for like, could we have an X-risk or extremely bad outcome is somewhere between 0 and 10%."

Safety and frontier capability are convex: investing in alignment improves product quality.

Claude’s personality and refusal style emerged from alignment research and constitutional AI, making the model more trusted and useful.

Why: Users prefer aligned agents that understand intent over literal but harmful compliance.

"working on one helps us with the other thing ... people really loved about it was the character and the personality. And that was directly a result of our alignment research."

Embed a published, society-vetted list of natural-language principles directly into the model.

Model critiques and rewrites its own outputs against principles (UN Declaration of Human Rights, Apple ToS, etc.) via self-supervised RL-AIF.

Why: Scales oversight beyond limited human raters and makes values transparent to users.

"we ask the model itself to first generate a response ... critique itself and rewrite its own response in light of the principle."

Expect 20 % unemployment during transition; long-term capitalism itself may dissolve.

AI will expand the pie first (10× code output, 82 % customer-service automation) then eliminate categories with low skill headroom.

Why: Exponential curves look flat until the knee; societal institutions lag technology.

"it's hard for me to imagine that even capitalism will look at all like it looks today."

Use AI tools ambitiously and iteratively; don’t treat them like legacy software.

Ask for big changes, retry prompts multiple times, and adopt tools early—legal and finance teams already gain from Claude Code.

Why: Early adopters leverage 10× productivity before full automation arrives.

"the difference between people who use Claude code very effectively ... is like, are they asking for the ambitious change?"

Prioritize curiosity, creativity, and emotional intelligence over credential chasing.

Montessori-style self-led learning; facts will fade, adaptability and kindness persist.

Why: In a post-singularity world, specific knowledge becomes obsolete quickly.

"I just want her to be happy and thoughtful and curious."