I love LLMs, I hate hype

There's a strange split running through a lot of AI commentary lately, and one veteran hacker-turned-AI-tinkerer just laid it out plainly: love the technology, hate the propaganda around it. He spent 2007 to 2014 doing security and hacking work before pivoting his entire career to AI, and he's not shy about how much fun he's having with it. Last week he got a local GLM-5.2 model running through opencode on Linux and had it install tmux using his own personal configuration without a hitch — the kind of small, delightful win that makes the whole enterprise feel worth it.

What bothers him isn't the models. It's the narrative wrapped around them. He singles out two specific strains of hype: the doom-laced messaging about a closing window of opportunity, an emerging permanent underclass, or civilization hopelessly falling behind — and the logical leap from 'useful autocomplete tool' straight to 'entity that will control the future of the universe.' He calls the first strain a guilt trip designed to funnel people into San Francisco, and the second a strawman that skips every intermediate step of actual evidence. He's willing to bet everything he owns that no overnight flash-of-light singularity event is coming, and he points out that breathless superintelligence talk has circulated in slide decks since at least 2016, with the same basic movie-plot fears going back to 1991.

His deeper argument is about money, not metaphysics. He thinks frontier AI labs dress up their anti-open-source positions in safety language or national-security framing, but the real motive is fear of commodification. AI progress, in his view, is mostly riding the same wave as Moore's Law and decades of general computing advancement — not something any single lab invented or owns. That distinction matters enormously for how you value companies raising billions of dollars: the value created by AI could be enormous while the companies capturing headlines capture very little of it, because the underlying capability keeps getting cheaper and more replicable.

He also walks back some of his own past skepticism about AI and programming, admitting he was probably too harsh when he previously argued models can't really code. His current framing borrows from a Linus Torvalds line comparing coding agents to compilers — agents might make programmers ten times more productive, compilers a thousand times more, though he thinks both numbers are inflated. Still, he says he's noticeably better at using these tools now than he was, treating them as a genuine new skill rather than a gimmick. He's careful to flag the downsides too: heavier cognitive fatigue, and a flood of 'vibe coded' software that's mostly still sloppy. But he lands on a comparison that undercuts the hype without dismissing the tool — LLMs belong in the same category as find-and-replace, Stack Overflow, or regex, useful shortcuts that quietly raise the floor of what any one person can build.