medium convictionactive · updated 2026-08-15T00:00:00.000Z

Continual learning → deployment-as-training → lab lock-in moat

Once models learn on the job, usage compounds capability, switching costs become employee-firing-like, and inference batching of personalized weights favors large orgs — a cloud-like margin moat the labs currently lack.

The chain

Whole-job competence requires writing experience into weights; session-to-session markdown is not continual learning (saxophone-student analogy).

dwarkesh-patel in 2026-08-15-dwarkesh-podcast-8-predictions-for-the-era-of-continual: "I don't think you can have AIs that perform whole jobs as competently as humans if they are forced to just write markdown files from session to session."

When deployment is training, returns to being ahead accelerate and labs must ship the smartest model immediately (Anthropic's 4-month Mythos internal gap becomes suicidal).

dwarkesh-patel in 2026-08-15-dwarkesh-podcast-8-predictions-for-the-era-of-continual: "When deployment becomes part of training, the returns to being ahead in the AI race accelerate. Because if you have the best model and more people are using your AI ... your model will become even smarter."

dwarkesh-patel in 2026-08-15-dwarkesh-podcast-8-predictions-for-the-era-of-continual: "Anthropic has reportedly been using Mythos internally since February, but it only shipped the model to the public in June. In the regime with actual continual learning, this kind of thing would just not be possible."

Session-to-session improvement creates switching costs like firing an experienced employee, letting labs demand cloud-like margins; enterprises that refuse training access may be locked out of the best models.

dwarkesh-patel in 2026-08-15-dwarkesh-podcast-8-predictions-for-the-era-of-continual: "If you want to change the AI that you're using, you basically had to fire an employee that has accumulated months of context on your organization and you replace them with a very fresh, very unexperienced new intern."

dwarkesh-patel in 2026-08-15-dwarkesh-podcast-8-predictions-for-the-era-of-continual: "the labs may say that any enterprise that refuses to let them train on the sessions can't have access to the very best models."

Personalized full-weight updates have huge inference-batching economies (optimal batch >2,400 sequences for a sparse model like DeepSeek V3), so large orgs serve weight-forks efficiently and individuals suffer >100x worse compute efficiency.

dwarkesh-patel in 2026-08-15-dwarkesh-podcast-8-predictions-for-the-era-of-continual: "the optimal inference batch size for a sparse model like say, deep seq v3 is more than 2,400 concurrent sequences being generated at once. ... A large company with lots of employees and agents ... can very efficiently serve their weight fork, whereas an individual user who's only running a batch size one may suffer more than two orders of magnitude worse efficiency."

What would falsify this

Step 1: A markdown/memory-scaffold agent matches human whole-job competence in a controlled eval.
Step 3: Enterprises multi-home models with no switching-cost penalty after on-the-job learning ships.
Step 4: LoRA/adapters make personalized weights cheap at batch-size 1.

Contradictions / tensions

Essay is a narration of an untested future; no empirical confirmation that continual learning works.
Merging per-user weight forks back into a base model is explicitly deferred as 'more technically challenging.'

Implications

Tradeable: winner-take-more among frontier labs (private) and the cloud analog (AMZN/GOOGL/MSFT) already earning high margins on switching costs. Coding-agent lock-in (Cursor etc.) is the near-term public surface.
Sequel to continual-learning-as-next-breakthrough — that page is the bottleneck; this page is the industrial-organization once the bottleneck breaks.
Safety/regulatory regimes that assume a train-then-deploy freeze become archaic — a policy risk, not a ticker.
ryan-greenblatt in 2026-08-15-dwarkesh-podcast-ryan-greenblatt-what-happens-once-ai-can: "I expect full automation of ARD, perhaps somewhere around 2031. 2030, and then getting to the beats. All humans on the job milestone. Maybe I expect median around 2033." Dated timeline on the *automation-of-AI-R&D* step that turns deployment-as-training into a lab lock-in race.
ryan-greenblatt in 2026-08-15-dwarkesh-podcast-ryan-greenblatt-what-happens-once-ai-can: "R and D is a type of task at which the AIs are especially good because both the companies are trying really hard to make their AIs good at R and D. And it's the kind of domain, it has a lot of nice properties ... pretty verifiable." Verifiability as the reason AI-R&D automates before most other jobs.

Companies

Dwarkesh Patel

Concepts

Continual learning as the next breakthrough Proprietary coding data as a frontier moat Train-then-deploy safety regimes become archaic under continual learning

Open questions

none