Theo frames a conversation with Sam Altman around whether AI models can reliably learn new tools, frameworks, and APIs after training, or whether they remain stuck around the conventions they were trained on. His core worry is that AI coding tools may freeze today’s software stack in place, while Sam’s answer made him more optimistic that models will soon be able to explore a new environment once and then use it reliably.
Watch on YouTube ›Get the market thesis, key claims, assets, contradictions, and follow-up questions from any financial video — then unlock a version personalized to your portfolio, watchlist, and favorite speakers.
Theo’s thesis is that the biggest practical limitation of current AI coding tools is not raw intelligence, but poor adaptability to new software primitives. He worries that if models only know how to use frameworks, languages, and libraries as they exist today, then the ecosystem could ossify around today’s defaults and make it harder to adopt better tooling later. He brings that concern directly to Sam Altman, and Sam’s answer is presented as cautiously optimistic: OpenAI is aiming for models that can be shown something totally new, explore it once, and then use it reliably. From there, Theo spends most of the video arguing why this problem is real. He uses a runtime-vs-compiler analogy to say models are effectively “compiled” from training data rather than continuously learning. …
Near term, the tradeable setup is still around compatibility-first developer tools: AI agents work best when new software looks like old software. The immediate risk is that newer frameworks keep getting routed back to legacy patterns unless explicit docs or retrieval tools are added.
Over the next few months, the likely path is gradual improvement in model adaptation, but only if memory, retrieval, and codebase-aware agents keep reducing the gap between training data and current APIs. If that fails, adoption should continue to favor drop-in replacements over novel abstractions.
The structural question is whether AI becomes a force that accelerates software evolution or one that unintentionally standardizes it. If models never learn new tools well enough, software design will drift toward compatibility with what models already know; if they do, the pace of framework innovation could expand rather than contract.
Current AI models cannot learn new information after their training is complete — they are like compiled programs, not runtimes.
The speaker analogizes models to compilers (capabilities cemented at compile time) vs runtimes (can add new functionality dynamically), arguing models are frozen after training.
Sam Altman expects models to be able to learn new tools and technologies on the fly within the next couple of years.
Sam Altman directly states that the milestone of models rapidly learning new things is 'a next couple of years thing' and doesn't feel 'very far away'.
AI models currently struggle with newer frameworks like Effect, Convex, TRPC, and Tailwind V4, often reverting to older versions or needing access to the full codebase to function correctly.
The speaker cites specific examples of model difficulties, including the Effect devs recommending giving the model the entire codebase, and Tailwind V4 causing models to revert to V3.
Are we making foundations out of the technologies as they exist right now that are going to be harder to swap in the future? Do you think we'll be able to steer the models enough to get them to use new things or are we just done improving the technologies we build on now?
Sam Altman says that if we use models correctly, they are like a general purpose reasoning engine. He believes we are moving in the right direction and that within a couple of years models will be able to learn new skills and adapt to new tools and technologies even faster than humans. He says a milestone will be when a model can be presented with something totally new, have it explained once, and then super reliably use it.
Why do context and specialized instructions sometimes make coding agents worse rather than better?
The speaker argues that every extra instruction or tool description consumes context that could have been used for the task itself, so the model is steered more by external text than by its own built-in capabilities. They frame this as specialization coming at the cost of generalization and say that over-instructing can make the model worse at the core job.
How do memory systems like Beads help coding agents across runs?
The speaker describes Beads as a graph-based issue tracker and memory layer for agents, intended to store information and make it available later so the agent can keep track of tasks and needed information across runs. They present it as promising but say they have not used it personally yet.
Unlock the full claims, asset map, scores, related transcripts, follow-up questions, and AI chat — shaped around your portfolio, watchlist, favorite speakers, and risks.
