Skip to content

Podcast

The Next Big Breakthrough Will Be AIs Learning on the Job

Dwarkesh Podcast

Source ↗ ← All highlights
  • Research Bet On Massive RL For AGI
    • Labs are betting that training AIs on millions of verifiable RL tasks will yield general problem-solving agents.
    • This assumes scale can overcome deficits like sample inefficiency and lack of continual learning by amortizing training across many sessions. (Time 0:00:00)
  • Verifiable Domains Still Need Grindable Simulators
    • Verifiability alone isn’t enough; a domain must be grindable with deterministic, replayable simulators to enable massive parallel RL training.
    • Computer use lags because sites and apps aren’t easily cloned for thousands of identical rollouts without heavy engineering. (Time 0:02:24)
  • Sample Efficiency Is Essential For Real World Skills
    • Many real-world skills can’t be trained with replayable environments because outcomes are sparse, delayed, and require real-world interaction.
    • Sample efficiency is crucial since you can’t run thousands of parallel, resettable rollouts for business, politics, or markets. (Time 0:04:25)
  • Deployment Is Wasted Learning Opportunity
    • Most inference compute (~30–50%) currently produces no learning since deployment sessions aren’t distilled back into model weights.
    • Real-world deployment reveals the most valuable data about usage, mistakes, and tacit organizational knowledge that models aren’t yet learning from. (Time 0:07:50)
  • Weights Compression Versus Context Cache Limits
    • Human learning compresses experience into weights; models relying on large context caches can’t scale that way.
    • Gradient updates are sample inefficient, so current online learning works only when millions of users repeat identical objectives. (Time 0:09:03)
  • Using Mercury To Automate My Bill Pay
    • Dwarkesh uses Mercury to automate invoice processing by routing contractor emails to a dedicated account.
    • Mercury scans invoices, extracts fields, and creates draft payments for him to review, eliminating manual bill tracking. (Time 0:10:58)
  • Distill Session Learning With OPSD
    • Use on-policy self-distillation (OPSD) to distill session-learned behavior back into base model weights.
    • OPSD trains the base model to match a veteran teacher’s per-token probabilities, giving denser supervision than sparse RL rewards. (Time 0:12:52)
  • OPSD Balances Sparse RL And Dense Supervision
    • OPSD preserves RL’s sparse, targeted updates while providing denser token-level supervision, avoiding naive full-session supervised replay.
    • This prevents overwriting unrelated capabilities by focusing updates only on what’s necessary to match the teacher’s performance. (Time 0:14:06)
  • Dreaming As Test Time Training
    • Dreaming means the model builds internal simulators to rehearse skills and generate massive additional training samples.
    • If accurate, dreaming becomes a fourth scaling axis: test-time training where the model writes environments and trains against them. (Time 0:15:34)
  • 2027 Vision Of On The Job Learning
    • A 2027 scenario: RLVR yields competent agents that gather real-world experience and then distill weekly learnings back into base models.
    • Combined with longer effective context windows, continual learning from deployment becomes the primary driver of improvement. (Time 0:17:27)