Reflections on Dylan Patel’s podcast
Working draft. These are notes after listening, not part of the essay yet.
What I initially thought he was arguing
The obvious disagreement seemed to be about demand. Dylan thinks people will always want the frontier and that demand for tokens is almost unlimited. I have been wondering whether most tasks actually have a ceiling, where a smaller and cheaper model is already more than good enough.
But I think there is a more interesting argument underneath what he is saying.
The public gap might not be the real gap
We can look at public models and conclude that competitors are catching up faster. But what if OpenAI and Anthropic already have much more intelligent models internally?
They can use the internal frontier model to help train the next frontier model. They can also use it as a teacher to create smaller models at every price point. Those smaller models make money, the money pays for the next frontier training run, and then the loop starts again.
If the competition is capital constrained, it might understand the same techniques and still be unable to repeat this loop. Knowing how something was trained is different from having the money, compute and strongest teacher model needed to train it.
Where this could become difficult to catch
At some point, the internal model could become better at AI research than the researchers working at competing labs. The competing researchers might eventually recreate the public model, but the frontier lab is already using a stronger private model to build the next one.
In that case, the public catch-up window could keep shrinking while the internal gap is actually growing.
Why I am still not sure
Until now, competitors have caught up. They did it with less money, less compute and weaker models. So either internal models are not driving research enough yet, or knowledge and efficiency improvements still spread faster than this advantage compounds.
The real question might be whether there is a point where this changes. When does an internal model become good enough at AI research that catching up becomes structurally harder rather than easier?
What I want to understand next
- How much are internal models already contributing to AI research?
- Was Mythos genuinely far ahead of the publicly released models?
- Can a frontier model reliably produce the best smaller model at every price point?
- Is compute still the main bottleneck even if the model can do more of the research?
- Can China, Google, Meta or a state-funded lab enter the same loop?
- Are we measuring the wrong catch-up window by looking only at public releases?
Podcast notes
14:53–18:27 · Demand and Mythos
Dylan says people immediately move towards the newest frontier model. He describes Mythos as a large capability jump and argues that a model with more expensive tokens can still be cheaper for a finished task if it needs fewer tokens and fewer attempts.
18:38–20:59 · Models make implementation easier
He argues that implementation used to be the hard part. As models make implementation easier, companies can try many more ideas. Choosing a valuable idea and having enough capital become more important.
21:05 onward · Access could concentrate
He expects the best models to become less broadly available. Labs may give early access and higher limits to companies that can pay and turn the intelligence into something valuable.
Source: The Supply and Demand of AI Tokens, Dylan Patel with Patrick O’Shaughnessy, April 2026. The claims about Mythos are Dylan’s claims and have not been independently verified here.