Something on my mind a lot recently is this scene from the first episode of Westworld:
I remember watching this when it came out thinking that it was so unrealistic as to be comedic. I very distinctly remember thinking “Any complicated AI being would no be able to be debugged just by talking to it. That’s so dumb, you would obviously need actual real debugging tools. You’d be writing code, etc. Stupid!”
10 years on, now I’m not so sure.
Right now, there are two(-ish) groups of people that use AI. There are the default chat-interface people we’ll call Group A that use whatever OpenAI or Anthropic serves them in a web UI, and then there are the agent/harness people using things like Claude Code or Codex or Pi or whatever. Let’s call them Group B.
There is a tacit assumption right now that, like most all technology before, expertise and mastery (or a desire for it) drives a portion of Group A to Group B. You could call group B “Expert Users”.
However, model advances seem to upset this status quo, advancing at a rate that seems to actually push Group B people back to Group A, as Group A’s default case gets stronger, and the mastery Group B developed is subsumed by the advances as to render them useless. Model enhancements seem to obviate the need for previously required workflows that were used to get better agent behavior.
For example — when I ask Claude to look for something, I do not tell it explicitly to search the web. However, it does search the web, as it knows broadly what I want based on what I asked it.
When I say “I” here, I mean not only literally me-the-person, but also anything an agent may have access to: my computer documents, memory files, browsing history etc. When I ask an agent a question, I’m asking it with my whole visible-to-claude self.
Because of this, the model already does a really good job doing the thing I asked it to do. It has my previous context, it’s own memory on my computer, project context, etc.
And this is obviously not just me, but any version of an “I” here, meaning this same capability is available to anyone else. I maintain that the best way to work with the AI is to just ask it stuff, so people that continue to just ask the model things are able to leverage all the AI advances without needing to really do anything extra. It just works.
“Just ask it things” seems to be the only strategy that scales, and is immediately available to both expert and novice users equally. The asking here can be specific or well informed (think scientific/medical research, etc.), but the doing of the model increasingly becomes unrelated to the input, or more that it knows how to do what you ask of it, regardless of what you ask.
This stands apart from “expert AI use advice” by presumably AI-use experts of the last few years, with much of it geared towards trying to coax the model to behave in a certain way. In fact, I think the idea of getting the model to “perform” (i.e,, an expert’s work) is just becoming so much less of a thing because of model advances.
Karpathy recently talked about AI-use cohorts, and described the general “expert” cohort like this:
Now we get to professional use of frontier-grade LLMs in math & code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries… This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.
Notice of the passiveness of the word “see” here. I don’t think this is a knowing wink to what I’m saying by him, but I think the idea of “witnessing” is a similar dynamic to what I’m describing.
He goes on to talk about a higher tier of use that is even further divorced from model use, and is characterized almost solely by model access:
And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work.
This is all to say that someone’s ability to leverage a model for some outcome seems to pale in comparison to simply having access to a great model that you can just ask things of that will do what you want. And knowing what to ask is not about model-use skill, but more about knowledge of the domain you’re asking the model about.
I think of it like this: Imagine there is some spectrum of “all possible intelligence”, with the dumbest things on the left, and universe-understanding-at-atomic-level on the right (some may call this “God”):
An interesting thing here is that we don’t really know where humans are compared to all possible intelligence, in part because we don’t really know what the upper bound of “understanding” can be. If someone happens to know everything about our reality and Earth, they would be very smart based on our current frame of reference. But what if it turns out there are in fact many realities and many millions of Earth-like planets? That one being/person could be aptly described as dumb with a frame of reference that encompasses everything.
I think for our own sanity though, let’s assume that humans as is are in the middle of the pack:
If we assume that million-worlds-knower is on the graph, but not as intelligent as the Godhead, they may go somewhere here:
Now, let’s imagine the smartest possible human. Someone who intimately knows everything there is to know about everything we know we can know now, with immense ability for taking in new knowledge. Lets also imagine this “smartest human” is The Best Human Person With AI. They know every possible way to interact with a model to get a good result, etc.
Combined, that’s almost certainly more the actual smartest person now, but no where near the million-worlds-knower. The “smartest possible human” barely registers as any movement on the X-axis:
Compared to the span of all possible intelligence, our smartest human isn’t even close.
Now, let’s imagine that, the further you go to the right on the X axis, by definition your ability to know and reason about intelligences to the left of you grows exponentially the further you are from it. Said differently, a Godhead can understand inert matter more than the inert matter can understand itself. Inert matter has no ability to understand or conceive of a Godhead.
That is the extreme example, but imagine how this works with a sufficiently advanced AI. We already know, based on benchmarks, that a frontier model’s general intelligence is greater than any human in terms of it’s workable knowledge (and maybe intelligence as well, given the recently published math results from OpenAI). Not only this, but our ability to reason about how LLMs now work is also far from a given, spawning a discipline in an of itself (Interpretability).
This puts LLMs now likely to the right of us. Tack on the compounding advances of a few years, maybe reaching superintelligence, and I suspect a sufficiently advanced AI would be probably about half way between us and the million-worlds-knower (I suspect that the limits of available compute, even at planetary scale, prevents AI from becoming God. For now).
It follows then, from my coarse, pseudoscientific diatribe, that a superintelligent AI would be able to understand what any human wants when it interacts with it. Or if not that, at least infinitely able to better understand the intention of what a human is asking of it.
Going back to that sense of an “I” being your sort-of “personal metadata”, a sufficiently advanced intelligence will be able to rapidly build up a picture of you, your needs, desires, preferences, etc., and serve you what you are looking for no matter how you ask it.
This is the “seeing” of the Karpathy quote from above. We experience our desired action, no matter our input. AI-use “Experts” have no leg to stand on.
This position puts me a bit at odds with this (popular) recent post. Its conclusion is:
The useful distinction is between wielding an instrument with skill and delegating work without adequate understanding or oversight. That distinction applies to AI, and it gives us something more practical than declaring that an entire class of software cannot be a tool.
The author spends most of the post rebuking Rich Hickey’s talk about AI, which I think is correct, but the conclusion here I think is missing that the understanding-ness of use doesn’t have to implicitly come from the human. Or assuming that the human using the tool is implicitly the one that best-positioned to know how their own problem should be solved (or what approach should be taken). Or if nothing else should be expected to instruct the AI how to do the task asked of it.
Given what I talked about above, I just don’t think this is the case, or if it is, it will not be for much longer. The skill feels, to me, that it will be irrelevant in short order.
What will still matter is asking the right questions, informed by a problem space, etc. This sketches at notions of “taste” that I think others have echoed as still relevant in the future, and I agree here.
I talk to people at work frequently about the idea of “AI model advancements” as something like “buy the index” investment advice. There was an industry wide push for putting SKILL.md files in repos about 8 months ago, to help models better understand codebases. I’m not sure this matters anymore, or more that “making the tool itself be more AI-legible” outperforms skills for evals, when benched against later models that are just simply better at figuring out what to do based on the default data they are given. “Just ask the [better] model things”, like buying the index, is just incredibly hard to consistently beat.
To be clear, and to wrap this up, I don’t think this is necessarily a bad thing. I think a lot of the crazy harness engineering stuff of the last few years has been fun to watch but also a lot of noise. I get a lot of use out of AI by having a relatively “dumb” process, and I find that I never run into token spend issues, reasoning loops, etc., and consistently get good results. I may even call myself an AI Expert (for now). But I would characterize this less about knowing individual tool use things, and more just an openness to try stuff, and follow my own taste to what I think is good. I care less and less about my meta-interactions with AI, and instead try to work on the paved path of just asking it things. If something doesn’t work, I just ask it again (which scales — I can ask an AI model something 1000x times and then sort through the produced solutions myself. The asking is trivially easy to automate!).
And my hope is that, as the walls to “expert” use continue to be torn down, others can see these same benefits. Naïve uses becoming the “best” way to use AI is actually the success outcome here. If good results from AI requires people to be well-versed in various tools and techniques, it undermines AI as a good solution as it puts up knowledge prerequisites for engagement. If instead people can get amazing results with little upfront work, that’s great! It democratizes the benefits of AI to people who don’t read all the blogs, Github repos, Karpathy posts, etc. It allows everyone to get the best possible output from a model. Everyone can be an expert.
Published on October 18, 2026.