Thoughts on AI progress and why AI labs' actions hint at a worldview in which AI models will continue to fare poorly at generalization and on-the-job learning
Why I'm moderately bearish in the short term, and explosively bullish in the long term — What are we scaling? X: @sriramk , @_simonsmith , @dwarkesh_sp , @emollick , @dwarkesh_sp , @dwarkesh_sp , @mikeknoop , @bobmcgrewai , @dwarkesh_sp , @dhruvboruah , and @arikagan_ X: Sriram Krishnan / @sriramk : really good read from @dwarkesh_sp on current state of AI. also worth thinking about how to think about “timelines” broadly - and why in my personal view, AGI is not a very useful concept when thinking about how these systems develop in the near future. https://www.dwarkesh.com/... Simon Smith / @_simonsmith : Huge respect for Dwarkesh, and he sparks some great conversations, but one thing I disagree with is how general most human employees are. Over a 25-year career I've managed dozens of people and worked directly with hundreds. The idea that the average person can learn to do any Dwarkesh Patel / @dwarkesh_sp : I totally buy that AI has made you more productive. And I buy that if other lawyers were more agentic, they could also get more productivity gains from AI. But I think you're making my point for me. The reason it takes lawyers all this schlep and agency to integrate these models [image] Ethan Mollick / @emollick : Interesting post & agree AI has missing capabilities, but I also think this perspective (common in AI) undervalues the complexity of organizations. Many things that make firms work are implicit, unwritten & inaccessible to new employees (or AI systems). Diffusion is actually hard [image] Dwarkesh Patel / @dwarkesh_sp : @deredleritt3r I think we might actually agree on the state of models. For example, the reason I haven't hired a research assistant is that I think it would be hard to find someone who is such an improvement on these models in answering my questions that the greater hassle and latency of Dwarkesh Patel / @dwarkesh_sp : Robert Conquest had 3 famous laws of politics. The first: “Everyone is conservative about what they know best.” In the same vein: “Everyone thinks that the median job is far easier to automate than their own job.” Mike Knoop / @mikeknoop : Great post. It's unfortunate that shifting AGI goalposts is associated with luddism. Pointing out flaws and building theories is how to drive progress — in fact it's a strong bull signal as we get more humans studying the real issues, increasing the likelihood we solve them. Bob McGrew / @bobmcgrewai : “Models keep getting more impressive at the rate the short timelines people predict, but more useful at the rate the long timelines people predict.” Good post. Dwarkesh Patel / @dwarkesh_sp : New post: Thoughts on AI progress (Dec 2025) 1. What are we scaling? [image] Dhruv Boruah / @dhruvboruah : The power/speed of humans to learn & process the million changes in the context window is just incredible. Less memory constraint is when we will get to AGI as @karpathy mentioned! Great analysis by the boss @dwarkesh_sp . I got nanobanana to give me a summary for my brain to [image] Ari Kagan / @arikagan_ : I don't see why this has to be true. Isn't it at least possible that RL is a necessary step to build the scaffolding for learning on the job? Sure, humans don't rehearse every software task, but we also don't pop out of the womb fully baked either! [image]
Context & Ripple Effects
This debate sits between arguments for a longer path to broadly capable AI—such as Karpathy’s decade-scale AGI framing—and evidence that capability gains do not automatically translate into reliable real-world work. Related coverage on software agents similarly identified a “messiness tax” in real-world engineering tasks.
The post matters less as a forecast than as a challenge to the premise that benchmark progress equals workplace autonomy. Responses from Mollick, McGrew and Kagan sharpen the distinction between models improving quickly and organizations being able to deploy them effectively.
First-order effects
- The discussion shifts attention from headline model capability to generalization and on-the-job learning as the binding tests for AI agents in workplaces.
- Labs and enterprise buyers face greater pressure to demonstrate reliable performance in messy, organization-specific workflows rather than infer readiness from broad model progress.
Second-order effects
- AI product teams are likely to emphasize scaffolding, reinforcement learning, evaluation and human oversight to close the gap between a capable model and a deployable worker-like system.
- Buyers may favor narrower, well-instrumented automation projects over generalized-agent rollouts, extending the advantage of firms that can supply workflow context and integration.
Third-order effects
- If this gap persists, AI adoption may resemble a gradual industrialization process: major gains arrive through redesigned processes and complementary systems, not a single transition to general-purpose digital labor.
- The central competitive question would shift from who trains the strongest general model to who can make models dependable within particular work environments; the pace remains uncertain because the debate includes claims that reinforcement learning could change that constraint.
The trend: AI is moving from a race to demonstrate broad capabilities toward a harder contest to operationalize dependable agents inside real organizations.