Continual learning lock-in
If models keep learning on the job, switching labs looks like firing an employee. That is a future market-structure claim, not a measured one.
Dwarkesh Patel’s picture is an office, not a benchmark. You have been using one lab’s model for months. It has picked up how your company writes, decides, and files. To switch is not to change a vendor. It is to fire the person who has the context and hire a new intern.
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.
Dwarkesh Patel
The bottleneck he named earlier is why the intern stays dumb. Context is not memory — a model cannot just grow a larger cache as it meets more users. Weights are not cheap: the moment you write experience into the weights, you give up the sample efficiency of in-context learning. Humans, in his analogy, do not keep every observation on the tip of the tongue. They chisel intuitions back into the weights. Today’s systems sit inside organisations, see the work, and throw it away at the end of the session.
He also states the other side. Deficits like data inefficiency and the lack of continual learning “can just be steamrolled if we just scale training more,” the way every named problem in natural-language processing collapsed under compute. The essay rejects the steamroll. It does not refute it. No second source has settled the point.
The industrial-organisation sequel stays on the same voice. If deployment is training, the returns to being ahead accelerate, and a four-month internal gap — Anthropic using Mythos since February and shipping it in June — becomes, in Patel’s word, suicidal. Labs may tell enterprises that refuse training access they cannot have the best models. Personalized full-weight updates favor large orgs: the optimal inference batch for a sparse model like DeepSeek V3, he says, is more than 2,400 concurrent sequences. A single user at batch size one is more than two orders of magnitude worse. Merging those per-user forks back into a base model is deferred as “more technically challenging.”
Ryan Greenblatt, on a related Patel conversation, puts full automation of AI research and development “perhaps somewhere around 2031. 2030,” and the beats-all-humans-on-the-job milestone around 2033. Those dates are his.
A markdown-and-memory agent that matches human whole-job competence would undercut the first step. Enterprises that multi-home with no switching-cost penalty would undercut the lock-in. Cheap adapters at batch-size one would undercut the batching economy. None of those tests have been run. The moat is a narration of an untested future.