Andrew Kelley, creator of Zig, and George Hotz (geohot) published two pieces critical of AI hype hours apart, collectively receiving 1653 points on Hacker News. Both say the same thing from different angles: LLMs work and are useful, but promises about closing windows and imminent singularity are harmful marketing.

Geohot says it explicitly: “I love the progress, I use a local agent to configure Linux, it actually works. What I hate is the negative hype (‘if you’re not in SF you’re done’) and the logical leap from ‘intelligent autocomplete’ to ‘flash of light in the sky and everything changes’”. Kelley criticizes Anthropic for inflating its models’ capabilities with claims that don’t hold up to scrutiny.

Why it matters to you. These aren’t isolated voices: they’re heavyweight technical figures (Kelley built a programming language, Hotz jailbroke iPhones and built autonomous vehicles) saying publicly trust your own judgment even when founders speak. If you’re deciding where to invest time or budget, their position is clear: evaluate on your use cases, not on promises. AI is a tool that gives real boosts on concrete tasks, not an oracle that solves everything nor a race where falling behind means losing forever.

The context. This convergence comes the same week Sam Altman reversed the narrative on jobs, shifting from “mass layoffs” to “net positive job creation”. It’s the same tension between marketing claims and verifiable reality.

In detail

Geohot’s piece

George Hotz starts from direct experience: “I configured a Linux box with opencode on GLM-5.2 locally last week, and it actually works: ‘install tmux with geohot config’ works, the Year of Desktop Linux is finally here”. Then he explains what bothers him:

1. Hype with negative valence: the constant talk about “closing windows”, “perpetually excluded class”, “hopelessly falling behind”. Hotz calls it “designed to make you feel bad and push you to San Francisco, where everything actually sucks as much as these people say”.

2. An unjustified logical leap: from the true observation that models are intelligent autocomplete, better compilers, more useful search engines, to “they’ll own the light cone, if you’re not at the right parties in SF one day you’ll see a flash in the sky and Everything Will Change”. Hotz bets everything he has that this won’t happen.

His thesis is that AI is a continuation of the computer revolution, driven by Moore’s Law and general computing progress, not something a specific lab is “making happen”. Labs have incentive to believe otherwise, because it justifies billions in funding. Hotz cites a 2016 presentation on superintelligence and a 1991 film about machines conquering the world: “A certain cult loves taking credit for things happening with or without them”.

On coding productivity, Hotz corrects his own past position: “Maybe I was a bit harsh about models not being able to program. What’s really happening is that programming is changing”. He cites Linus Torvalds: agents give 10x boost, compilers 1000x. Hotz estimates both numbers are exaggerated, but now he’s “confident enough I’m better at using them and getting some boost”. He warns though: “You have to be very careful, they can increase cognitive fatigue, and all the code written with AI in your place is still slop (where’s all this amazing new software that productivity improvements should imply?)”. Models are useful like find-and-replace, Stack Overflow, or all the regex he never learned to write and now never will.

Kelley’s piece

Andrew Kelley (the original post was redirected from a wrong URL, so complete content is missing from recovered sources) criticizes Anthropic for claims that “blow smoke” instead of calling things by their name. The original title was “Zig Creator Calls Spade a Spade, Anthropic Blows Smoke”.

The convergence

Both converge on three points:

1. Models actually work on concrete tasks (system configuration, code assistance, research).
2. Hype is harmful: it creates FOMO, pushes wrong decisions, and distracts from what actually works.
3. Evaluation should be on your own cases, not on founder promises or generic benchmarks.

The Hacker News reaction (1188 + 465 points, 896 total comments) suggests this position resonates with a significant part of the technical community: those using the tools daily see the gap between actual utility and marketing narrative.

Practical implications

If you’re evaluating where to invest time or money:

• Test on your own cases, not on demos. The benchmark from July 11 shows how everything changes between models on real tasks.
• Expect a boost, not replacement. Geohot makes it clear: it’s like moving from grep to a better search engine, not like replacing the programmer.
• Recognize negative hype. If a pitch makes you feel behind or excluded, it’s marketing, not analysis.

The real battle isn’t between those using AI and those not, but between those using it with judgment and those suffering it as a trend.