On LLMs, Hype, and Keeping Your Feet on the Ground

George Hotz published an essay this weekend called I love LLMs, I hate hype. It hit #3 on Hacker News — 357 points, 222 comments, and the kind of discussion that happens when someone draws a line that a lot of people have been feeling but haven't been able to articulate.

If you don't know Hotz: he jailbroke the iPhone at 17, built comma.ai to compete with Tesla on self-driving, and created tinygrad — a deep learning framework that fits in a few thousand lines of Python. He's been elbows-deep in AI for over a decade, and he's earned the right to have opinions about it. This essay is one of the clearer statements I've read on what it means to be genuinely excited about AI without buying into either flavor of nonsense surrounding it.

Go read it. It's short. Then come back.

---

The two things Hotz hates

Hotz's essay is structured around two things he can't stand about the current AI discourse. They're worth looking at separately because they pull in opposite directions, and a lot of the confusion in the industry comes from conflating them.

The first is negative-valence hype. This is the doomer stuff — "the window is closing," "you'll be left behind," "the perpetual underclass is coming." It's designed, Hotz argues, to make you feel bad about yourself and move to San Francisco, "where everything really does suck like how these people claim." The mechanism is familiar: manufacture urgency, sell the cure. In this case the cure is usually "give us money" or "join our company" or "adopt our framework before it's too late." The second is positive-valence hype — the singularity stuff. The jump from "it's a fancy autocomplete, smart compiler, better search engine" to "it's gonna own the whole light cone, bro, and if you aren't at the right parties in SF there's gonna be a flash of light in the sky one day and you won't even know what happened." Hotz bets everything he owns that this doesn't happen.

Both versions serve the same purpose: they make AI feel like something happening to you rather than something you can use. Whether you're afraid of being left behind or afraid of being vaporized by the light cone, either way you're not thinking clearly about what the tools can actually do for you right now.

---

The tool/ideology boundary

Here's the distinction Hotz draws that I think is the most practically useful: there is a difference between being excited about AI as a tool and subscribing to AI as an ideology.

Tool enthusiasm looks like this: you try a new model, it helps you do something faster or better than you could before, and you think "that's useful, I wonder what else it can help with." You evaluate it by output. You're empirical about it — you test, compare, keep what works, discard what doesn't.

Ideology subscription looks different. It looks like believing that AI is inevitably going to reshape the world in some specific way, and then filtering everything through that belief. The doomer version: "automation is going to eliminate most jobs, and the only question is whether you're positioned to survive it." The evangelist version: "AGI is near, and the only question is whether you're aligned with the right lab to benefit." Both versions are more about identity and belonging than about empirical reality.

The problem isn't having opinions about where AI is heading. The problem is when ideology replaces honest evaluation. When you believe so strongly in the narrative that you stop asking "does this actually work?" and start asking "how does this fit into the story I've already decided is true?"

This shows up in real engineering decisions. A team that's ideologically bought into "AI is going to replace developers" will adopt coding agents without measuring whether the agents are actually making them faster. A team that's ideologically skeptical will reject AI tools without testing them, because using them would threaten their identity as "real engineers." Both are making decisions based on story, not data.

---

Moore's law is doing the work

Hotz's most provocative claim is that AI progress is "mostly due to Moore's law and general progress in computing, not something that [frontier labs] are doing."

I think this is mostly right and slightly overstated. The frontier labs are doing real engineering — training infrastructure at this scale doesn't build itself, and architectural innovations like mixture of experts and chain-of-thought reasoning are genuine contributions. But the rate of progress — the fact that capabilities double every N months — is mostly hardware. More compute, more data, bigger models. The algorithms change at the margins. The compute changes by orders of magnitude.

This has practical implications for how you plan. If AI progress is driven by something specific that one lab is doing, you need to bet on which lab wins. If it's driven by a broad trend in computing that benefits everyone, you can be agnostic about vendors and focus on what the capabilities actually enable.

Hotz's point about frontier lab valuations follows from this: the labs claim to be inventing the future, but what they're actually doing is riding a wave that would exist without them. The value creation is real. The value capture is questionable. If someone else can ship a comparable model six months later at half the cost, the moat isn't technology — it's brand, distribution, and ecosystem lock-in.

This is why the anti-open-source arguments from frontier labs are worth treating with skepticism. The safety framing — "open models will be misused" — is real but partial. The economic framing — "if anyone can run these models, we can't charge premium prices" — is the part they don't say out loud. Actually, Hotz says it out loud: "the core of the anti open source arguments is a fear of commodification."

---

The honest programming take

Hotz softened his stance from The Eternal Sloptember, where he'd been harsh about models not being able to program. His updated view:

> What's really happening is that programming is changing. ... Models are useful just like find replace, stack overflow, or all the regexes I never learned how to write and now never will.

This is the most grounded take I've read on AI and programming in months. Not "AI will replace developers." Not "AI is useless for real programming." Just: these are useful tools, they're getting better, and the skill of using them well is a new one that takes practice.

He also links to a piece on cognitive fatigue from AI coding tools, which is important and under-discussed. Using a coding agent can feel extremely productive in the first hour and extremely draining by hour three. The agent generates code faster than you can review it. The context window fills up. You lose track of what the agent changed and why. You end up accepting changes you don't fully understand because keeping mental track of everything is exhausting.

The skill isn't "prompt engineering" in the prompt-injection sense. It's learning when to use the tool and when to put it down. When to let the agent generate boilerplate and when to write the tricky logic yourself. When to accept a suggestion and when to say "no, I need to understand this before it goes in." These are judgment calls, and they take time to develop.

Nobody talks about this because it's less dramatic than "AI will eat software" or "AI is just autocomplete." But it's the thing that actually determines whether a team gets value from these tools or burns out trying.

---

What staying grounded looks like

So if you accept Hotz's frame — genuinely excited about the tools, allergic to the ideology — what does that look like in practice? Here's what I think it means for engineering teams:

Evaluate with numbers, not vibes. Before adopting an AI tool, define what success looks like in measurable terms. Lines-of-code is a bad metric, but you can measure: time-to-merge for PRs of similar complexity, bug rate in AI-assisted vs. unassisted code, developer satisfaction surveys, cycle time for common task types. Run the experiment. If the numbers improve, adopt. If they don't, don't. If the experiment is "too hard to measure," the tool probably isn't ready. Budget for capabilities, not for vendors. Don't build your stack around a specific model provider. The cost-performance ladder shifts every month — Terra at 1/16th Fable 5 pricing is the current example, and there'll be another one next month. Design your systems so that swapping models is a config change, not a rewrite. The model is a runtime parameter, not an architectural commitment. Protect cognitive budget. The cognitive fatigue problem is real. Establish norms around AI tool usage: when to use it, when not to, how to review AI-generated code, how to stay oriented in a session. "The agent wrote it" is not an acceptable answer to "why does this code work this way?" If you can't explain it, it shouldn't ship. Separate signal from noise. Most of what gets called "AI news" is actually narrative management — labs positioning themselves for the next funding round, influencers building audiences, vendors creating urgency to close deals. The signal is: what models are available, what they cost, what benchmarks they clear, what people who've actually built with them report. Everything else is theater. Theater can be entertaining. It's not a basis for decisions. Stay genuinely curious. Hotz's opening line — about how "absolutely giddy" he is about AI — matters because it's honest. You can be excited about tools without being ideological about them. You can try new models because they're interesting, not because you're afraid of missing out. You can be skeptical of the hype without being cynical about the technology.

---

The line Hotz draws

The essay closes with a line that's easy to miss but carries the whole argument:

> AI is the continuation of the computer revolution. I love computers so much.

This is the grounded position. AI isn't a break from everything that came before — it's the next chapter in a story that's been running for decades. Computers changed how we work. The internet changed how we communicate. Smartphones changed how we live. Each wave was hyped as a revolution (and was), and each wave's hype was also full of nonsense that didn't age well.

The people who benefited most from each wave weren't the ones who bought the most extreme version of the narrative. They were the ones who stayed curious, tried things, kept what worked, and didn't let ideology substitute for honest evaluation.

Same thing this time.

---

Read the original: George Hotz, "I love LLMs, I hate hype" — it's worth the five minutes.