Dwarkesh Podcast

Dario Amodei on Dwarkesh Patel: Why the Exponential Ends

Why the big blob of compute predicts log-linear gains through 2025: AIME-tested RL and pre-training confirm the curve; SWE task breadth is the remaining gap.

DPDwarkesh PatelhostDADario Amodeiguest

Feb 12, 2026·2h 22m·Watch on YouTube ↗

🙋 FREQUENTLY ASKED QUESTIONS

Direct answers grounded in the episode transcript. Tap any timestamp to verify against the source.

1. 1

   What is Dario Amodei's Big Blob of Compute hypothesis?

   The Big Blob of Compute hypothesis says progress comes mainly from scalable inputs, not clever tricks. Amodei says he formed the view in 2017, when GPT-1 had just come out, and meant it as a general AI claim rather than a language-model-only claim. The inputs he names are raw compute, data quantity, data quality and distribution, training duration, scalable objective functions, and numerical stability or conditioning. Pretraining is one objective that scales, and RL goals can be another, including objective rewards in math and coding plus more subjective RLHF-style rewards. His update is that RL now appears to show the same kind of scaling as pretraining: more training on math contests, code, and a wider variety of RL tasks produces log-linear gains.

   ▸ 1:58 in transcript
2. 2

   Why does Dario Amodei say AI diffusion is fast but not instant?

   Amodei's diffusion point is that adoption can be explosive while still requiring organizational work. He rejects diffusion as a hand-wave for slow progress, but also rejects the idea that AI impact becomes instantaneous as soon as models improve. He points to Anthropic revenue as an example of a very fast curve, moving from zero to $100 million in 2023, $100 million to $1 billion in 2024, and $1 billion to roughly $9-10 billion in 2025, with more added in January. The reason it still takes time is practical deployment: enterprises need legal approval, provisioning, security, compliance, permission changes, rewritten internal tools, and change management. His model has two exponentials: one for model capability and one for diffusion into the economy.

   ▸ 20:24 in transcript
3. 3

   Why doesn't Anthropic buy more compute if AGI is close?

   Anthropic's compute caution comes from demand timing, not from disbelief in fast AI progress. Amodei says he has high confidence that very powerful models arrive within a few years, and a hunch that a country of geniuses in a data center could arrive in one to two years. The uncertainty is how quickly that turns into revenue. Data centers require forward commitments: if a company buys compute for 2027 assuming 10x annual revenue growth, but revenue is one year late or grows 5x instead, the company can go bankrupt. That is why he says responsible scaling is not mainly about spending the smallest amount. It is about writing down the spreadsheet, balancing the risk of too little capacity against the risk of ruin, and buying enough to capture strong upside worlds without betting the company on perfect timing.

   ▸ 47:16 in transcript
4. 4

   How can AI labs make money while training huge new models?

   Amodei's profit model separates profitable inference from losses caused by scaling the next model. He argues that an individual model can have positive economics because inference can carry high gross margins, but a frontier lab may still lose money while it spends far more to train the next generation. In his stylized example, a model trained for $1 billion could generate $4 billion of revenue and cost $1 billion to serve, but the company still loses money if it is simultaneously spending $10 billion on the next model. At industry equilibrium, firms cannot spend 100% of compute on training because they need revenue, and cannot spend 100% on serving because they would fall behind. Profitability depends on demand forecasting, gross margins, and the fraction spent on research during the scale-up phase.

   ▸ 58:49 in transcript
5. 5

   What is Claude's constitution supposed to do?

   Claude's constitution is meant to teach principles that generalize better than rule lists. Amodei distinguishes two questions: rules versus principles, and corrigibility versus intrinsic motivation. On the first, he says Anthropic has empirically found that principles make behavior more consistent, cover edge cases better, and help the model understand what it should aim to do. A rule list like individual prohibitions is harder to generalize from. On the second, he says Claude is still intended to be mostly corrigible: under normal circumstances it should do the task a user asks for. The limits are principled refusals around dangerous or harmful requests, such as biological weapons. He says the constitution can be iterated inside Anthropic, compared across companies like Gemini, and informed by broader public input, including past work with the Collective Intelligence Project.

   ▸ 2:05:46 in transcript

Answers are AI-generated from the transcript and may contain errors. Tap a question to verify against the source.