ExpertsFMThe next big breakthrough will be AIs learning on the job
Dwarkesh Podcast · June 26, 2026
5 Selected Lessons:
reasoning_pattern · 2 heuristic · 1 case · 1 framework · 1
'Dreaming' — simulated rehearsal as a fourth scaling axis
Source
Dwarkesh Podcast — The next big breakthrough will be AIs learning on the job (June 26, 2026) · at 15:34 ▸
When this applies
when real-world data/practice is scarce or slow but you can build an approximate simulator of the environment
The takeaway
When real samples are the bottleneck, substitute cheap simulated practice for scarce real practice by building a model-based simulator — the harder the domain is to model faithfully, the weaker this substitution works.
Why it works
Simulated rollouts let a system experience orders of magnitude more practice per unit of wall-clock time by trading compute for scarce real data, as demonstrated by EfficientZero running many simulated games per real game step.
One level up
Compute can sometimes substitute for data if you can model the environment well enough.
Where it stops holding
This stops working when the domain is too complex or open-ended to build a faithful simulator of, e.g., the real world versus a game of Go.
Detail
approach: Model spends compute building an RL environment representing what it's witnessing in the real world, then trains against it before acting.
benchmarks: {"precedent":"EfficientZero playing dozens of simulated games per real game step, beating a sample-limited novice human"}
rationale: If this works, it becomes a fourth axis of scaling alongside pre-training, RL, and inference-time compute.
In their words
for each step in the real game, Efficient Zero is playing dozens of simulated games in its head. In a similar way, future LLMs might be able to consume far less real world data while practicing endlessly against environments that they build for themselves.
Amortized training cost justifies local inefficiency
Source
Dwarkesh Podcast — The next big breakthrough will be AIs learning on the job (June 26, 2026) · at 0:00 ▸
When this applies
when deciding whether a technique that is locally inefficient can still work because its cost is paid once and amortized across many future uses
The takeaway
Before dismissing an approach as inefficient, check whether that inefficiency is paid once (amortized) rather than repeatedly at the point of use.
Why it works
Training is a one-time cost spread across billions of deployment sessions; what matters for users is in-session performance, which can keep improving even if the training process itself stays sample-inefficient.
Where it stops holding
This stops working when the amortized training doesn't actually transfer the needed skill because the training distribution fails to cover the real deployment distribution.
Detail
approach: Separate training-time efficiency from deployment-time competence; judge a paradigm by in-session capability growth, not training sample efficiency.
signals:
- agents solving more ambitious problems over longer time spans
- capability gains tracked across RL training generations
rationale: If deployment behavior keeps improving despite training inefficiency, the inefficiency may not be the real constraint on reaching the goal.
In their words
They think that if we train AIs to accomplish millions of verifiable tasks across thousands of diverse RL environments, then we will have basically built AGI.
Non-replayable real-world domains break standard RL mass-rollout training
Source
Dwarkesh Podcast — The next big breakthrough will be AIs learning on the job (June 26, 2026) · at 2:13 ▸ · at 5:26 ▸
When this applies
when a skill's feedback loop requires months or years of real, non-repeatable interaction with the world rather than fast resettable trials
The takeaway
For skills whose outcome signal is slow, singular, and non-replayable, mass-rollout RL training is structurally inapplicable — a different, sample-efficient learning mechanism is required.
Why it works
Credit assignment (figuring out exactly what action caused success) normally requires perturbing and replaying many parallel trials from the same state; the real world doesn't allow resets, so standard RL-style mass rollouts cannot be applied to these skills.
One level up
Match the learning mechanism to the replayability of the domain's feedback loop, not just to whether outcomes are verifiable.
Where it stops holding
This stops working as an objection once sample-efficient online/continual learning techniques mature enough to learn from single, sparse real trajectories.
Detail
approach: Identify whether a target skill's outcome can only be observed via slow, irreversible real-world action; if so, expect RLVR-style training to fail to cover it.
signals:
- outcomes take months/years to resolve
- no deterministic simulator exists
- can't isolate which action caused the result
rationale: Reset-free, non-stationary environments are a known open problem in RL, and real-world entrepreneurship/politics/law are extreme examples of this.
In their words
The rollout here requires interacting with the real world, and you can't recreate it from just within the data center.
Grindability test: verifiable isn't enough, you need replayable rollouts
Source
Dwarkesh Podcast — The next big breakthrough will be AIs learning on the job (June 26, 2026) · at 2:13 ▸ · at 5:26 ▸
When this applies
when deciding whether any skill domain can be mass-improved through trial-and-error practice at scale
The takeaway
A domain only scales with practice-based learning if you can cheaply and faithfully replay the same starting conditions many times in parallel — verifiability is necessary but not sufficient.
Why it works
Fast capability gains require running thousands of parallel, identical, resettable trials; verifiability alone tells you success can be measured, not that you can cheaply generate many comparable attempts to learn from.
One level up
Structural bottlenecks (ability to practice) can matter more than measurement bottlenecks (ability to score) when judging how fast a capability will improve.
Where it stops holding
This stops working once a faithful simulator of the domain can be built, removing the replay constraint even for previously 'ungrindable' domains.
Detail
rule: To predict RL progress speed in a domain, check for parallel, deterministic, replayable rollout capability, not just verifiability of outcomes.
threshold: Thousands of identical parallel container copies (as in coding) vs. none available (as in live websites)
rationale: Models are extremely sample-inefficient during training, so without mass parallel practice the sample-efficiency gap can't be closed.
In their words
it is not enough for a domain to be verifiable. It also has to be very grindable in the sense that you have to be able to run lots of parallel rollouts against a deterministic and replayable simulator.
Cursor's tab model: online learning needs a shared objective across millions of users
Source
Dwarkesh Podcast — The next big breakthrough will be AIs learning on the job (June 26, 2026) · at 8:43 ▸
When this applies
when building an online-learning system and wondering whether weight updates can be personalized per-user or per-session
The takeaway
Per-instance personalization via weight updates currently requires either massive shared-objective scale or a different (in-context) mechanism; a single session's data generally can't safely drive a weight update.
Why it works
Gradient updates are sample-inefficient, so safely updating weights currently requires huge aggregated volumes of identical-objective feedback; a single user's or session's data is not enough to update weights reliably.
Where it stops holding
This stops working once new techniques (e.g., OPSD) let a single session's distilled signal update weights without needing millions of aggregated examples.
Detail
situation: Cursor wants its tab-completion model to keep improving from real usage.
decision: It online-learns one fixed objective (predicting accepted edits) across all users simultaneously.
outcome: At 400 million requests a day, this works, but no model has shown online learning of genuinely different things for different users.
lesson: Current successful online learning is narrow and shared, not personalized, because weight updates need huge volume.
In their words
the Cursor tab model online learns by predicting the same exact objective for over 400,000,000 requests a day.