Geohot on LLMs: Love the Tech, Hate the Hype

George Hotz (geohot) published a piece titled "I love LLMs, I hate hype" that hit the Hacker News front page today with 233 points and 126 comments. His thesis draws a line between genuine technological progress worth celebrating and manipulative narratives designed to drive anxiety, investment, or relocation to San Francisco.

The Core Argument#

Geohot opens by establishing his credentials as an AI believer. He's devoted his entire post-2014 career to AI and finds current developments genuinely exciting. He cites practical advances: language models, autonomous driving, video generation, and coding assistants. These represent real productivity gains - not revolutionary consciousness, but meaningful incremental benefits comparable to other developer tools.

His analogy: "Compilers make programming 1000x more productive." LLMs offer similar incremental benefits. Useful extensions of human capability, not fundamentally different from other tools in your stack.

What He Rejects#

Geohot identifies two problematic hype categories:

Doom narratives. The constant messaging about "closing windows," "perpetual underclasses," and falling "hopelessly behind." He describes this as "negative valence hype" designed to make people anxious and relocate to expensive tech hubs. The implication: if you're not at the right parties in San Francisco, you're going to miss the rapture.

AGI inevitability. The logical leap from "sophisticated tools" to "unstoppable superintelligence." Geohot dismisses this as a strawman, arguing that fancy autocomplete or better search engines don't inherently lead to systems that "own the whole light cone."

The Underlying Thesis#

Geohot's more interesting claim involves incentives. He suggests frontier AI labs benefit from credit-claiming for progress that stems primarily from Moore's law and general computing advancement. The opposition to open-source development, he argues, masks a fear of commodification - which would eliminate competitive advantages and undermine valuations.

In other words: the hype serves financial interests. Both the doom narratives (creating urgency) and the AGI inevitability claims (justifying investment) align with what labs need people to believe to maintain their market position.

What HN Is Saying#

The Hacker News thread split into several camps, with both agreement and pushback.

The anti-hype sentiment resonated. Multiple commenters agreed with the core thesis. One wrote: "Honestly, who likes any hype in anything ever? Especially if you genuinely like and understand the thing being hyped." Another noted: "There's sort of this spiteful anti-hype here that I find very offputting, and ultimately I think it's because a lot of folks are going out and encountering opinions I never see."

San Francisco criticism drew mixed reactions. Geohot's characterization of SF as "shitty" where "everything really does suck" sparked debate. Some agreed: "The SF metro is possibly the worst in the entire world in terms of CoL vs QoL." Others pushed back: "Your SF hate isn't a good look... SF is more than Paul Graham worship parties."

Commenters questioned his position. Several noted that Geohot runs a company selling AI hardware (Tinybox), making him "one of the merchants" he's criticizing. One wrote: "Geohot is one of the (attempted) merchants, but maybe that is not going so well and he is changing his tune." This prompted responses noting that builders and merchants are different categories - building with available tools is different from marketing hype.

The cognitive impact debate emerged. An extended thread discussed whether LLMs are "poison for the brain." One commenter cited research from arxiv (2506.08872) and argued that most honest users admit the tools are "making them dumber or zombifying them." Counterarguments invoked Socrates on writing weakening memory and 1960s calculator protests - historically, similar claims have been made about every productivity tool.

Labor market anxiety surfaced. A commenter described hearing from "supposedly reputable publications" that AI will end knowledge work and take out "a large percentage of the world's labor force." They noted being told to pick up a trade because their career knowledge is now worthless. This matches what Geohot characterizes as anxiety-inducing doom hype.

AI-generated content debate. Geohot's claim that he could "never love any AI generated music, book or artwork" drew responses about the evolving quality. One noted: "It was only like 2 years ago that artists were arguing this on the basis that AI-gen images would consistently mangle hands. Now we're at a point where that never happens." The counterpoint: comparing to CGI, we'd be at the late 1970s in terms of nascency.

The Practical Subtext#

Buried in the thread is a discussion about code ownership and AI assistance. One commenter noted that "FOSS communities were never valuable because of the code. It was the shared written and oral traditions that make the software useful, usable, and updated." Another described building merge-conflict resolution into their workflow via Claude Code skills.

The implicit argument: LLMs don't replace the human context around code. They accelerate certain tasks while potentially creating new maintenance burdens (tracking upstream, managing AI-generated drift, reviewing security implications).

The Bigger Picture#

Geohot's piece arrives at an interesting moment in AI discourse. The industry has split between cautious optimists who see useful tools and vocal camps claiming either imminent doom or imminent transcendence.

His framing - love the technology, reject the narratives - offers a middle path. Use the tools. Acknowledge the productivity gains. But remain skeptical of messaging designed to create urgency, drive relocation, or justify particular investment theses.

The HN thread suggests this resonates with a segment of developers who feel caught between genuine enthusiasm for AI capabilities and exhaustion with the surrounding discourse. Whether Geohot's particular read on incentives and motivations is correct, the distinction between tool appreciation and hype resistance clearly struck a nerve.

Sources#

• George Hotz Blog Post - Original article
• Hacker News Discussion - 126 comments as of publication
• AI Cognitive Impact Research - Referenced study on LLM usage effects