LLMs: Love the Tech, Skip the Singularity Hype

George Hotz on LLMs and Hype

George Hotz published the post on his personal site on July 12, 2026. He states direct enthusiasm for current LLM capabilities, self-driving systems, video models, and coding agents. He also rejects two common narratives: claims of an imminent window closing that leaves most people behind, and abrupt jumps from tool-level improvements to total control over future outcomes. Hotz attributes progress mainly to hardware scaling rather than any single lab's unique contributions.

Local Setups and Coding Agents

Hotz described running a local GLM-5.2 instance with opencode on a Linux machine. Commands such as installing tmux under a specific configuration succeeded without manual intervention. He notes this matches earlier statements from Linus Torvalds that agents can deliver roughly 10x gains while compilers historically produced closer to 1000x improvements in output per developer hour.

Current agent workflows still require careful oversight. Generated code often passes initial tests yet introduces maintenance debt once integrated into larger codebases. Hotz points out that simple operations like find-and-replace or Stack Overflow lookups already provide measurable speedups; LLMs add another layer on top of those established patterns rather than replacing them outright.

Valuation Arguments and Open Source

Hotz argues that frontier labs will capture less value than their current valuations imply. He cites a 2016 superintelligence presentation and a 1991 film about machine takeover as evidence that the core ideas predate recent funding rounds. The push against open releases, he claims, stems from concerns over rapid commoditization once weights and training techniques spread.

This view aligns with observed behavior in the ecosystem. Once a model family reaches usable quality, forks and fine-tunes appear quickly on public repositories. Companies that rely on closed checkpoints face pressure to keep releasing paid updates to justify continued investment, even as base capabilities diffuse.

Measured Productivity Trade-offs

Hotz revised an earlier claim that models cannot program, now framing the change as a shift in the programming task itself. The practical limit appears to be cognitive load: reviewing and correcting agent output can offset some of the raw generation speed. Projects built entirely through prompt iteration frequently require later rewrites once edge cases surface in production.

Developers report the largest gains when models handle repetitive scaffolding or boilerplate while the human maintains the system architecture. This division keeps the output closer to the quality level already achievable with conventional compilers and static analysis tools.

FAQs

Does Hotz believe LLMs will stay at the level of autocomplete? No. He treats them as a useful incremental tool alongside existing developer practices and expects continued capability growth from hardware improvements.

What specific local setup does the post mention? A Linux machine running GLM-5.2 through opencode that handled package installation and configuration tasks via natural language commands.

Why does Hotz criticize frontier lab valuations? He states that the underlying advances follow from general compute progress and will commoditize regardless of closed development, reducing the labs' ability to capture outsized returns.

---

📖 Related articles

• Agentic Coding: Una Trappola per lo Sviluppo Software?
• File agents.md: utili per gli agenti di coding?
• Lean-ctx: Ottimizzatore Ibrido Riduce Consumo Token LLM del 89-99%