George Hotz on LLMs: Progress vs. Hype

AI Progress is a Continuation of the Computer Revolution

Large Language Models (LLMs) and AI agents are not magical anomalies but the continuation of the computer revolution, driven primarily by Moore's Law and general progress in computing. The utility of LLMs lies in their role as highly efficient search techniques, smart compilers, and advanced autocomplete tools that increase developer productivity, rather than as precursors to a sudden, world-altering singularity.

The Utility of Coding Agents

Coding agents are becoming practical tools for professional software craftspeople. For example, using a local GLM-5.2 model with OpenCode allows for complex system configurations (e.g., "install tmux with the geohot configuration") to be executed successfully. While these tools provide a productivity boost, they are comparable to previous advancements like find-and-replace, Stack Overflow, or regular expressions.

However, this productivity gain comes with caveats:

Cognitive Fatigue: Using AI at high speeds can increase mental exhaustion.
The "Slop" Problem: Much of the software produced via "vibe coding" remains low-quality or "slop."
Programming Evolution: Programming is not being replaced but is changing. While agents may provide a 10x boost, the fundamental impact of compilers—which can be viewed as a form of automation—has historically provided a 1000x boost.

Deconstructing "Negative Valence Hype"

Much of the current AI discourse is characterized by "negative valence hype"—narratives designed to create anxiety and urgency. This includes claims about a "closing window" of opportunity, the creation of a "perpetual underclass," or the idea that those outside specific geographic hubs like San Francisco are falling hopelessly behind.

These narratives often serve two purposes:

Psychological Pressure: Making individuals feel inadequate to drive them toward specific industry hubs.
CapEx Driver: Using FOMO (Fear Of Missing Out) and singularity myths to justify massive capital expenditures in LLM development.

The Valuation Crisis of Frontier Labs

Frontier AI labs are currently valued based on the premise that they can capture the vast majority of the value created by AI. However, there is a strong argument that they will fail to capture this value due to the inevitable commodification of the technology.

The Fear of Commodification

Arguments against open-source AI are often framed as concerns over "safety" or geopolitical competition (e.g., with China). In reality, these arguments are frequently a cover for the fear of commodification. Because AI progress is largely a result of general computing trends, the "secret sauce" of intelligence is less a proprietary discovery and more a result of scaling compute and data.

The Economic Moat Problem

Community discussion highlights that the current business models of frontier labs are precarious. If users are pushed from flat-rate subscriptions to usage-based token billing, the lack of a "moat" makes it easy for users to switch to competitive open-source models.

"The frontier labs need everyone to answer 'yes' to spending 100x what they currently spend to justify the valuations, and it's just not going to happen as long as everyone knows how to make these models."

Shifting Paradigms in Software Development

As AI tools become more accessible, the nature of software creation is shifting toward highly specific, one-off tools and private deployments.

The "Have It Your Way" Era: There is a growing trend of building stripped-down, one-off software for highly specific use-cases in private homelabs. This reduces the incentive to upstream improvements to open-source projects, as forking and maintaining a custom version is now significantly easier with AI assistance.
The "Sausage" Risk: Without a clear architectural vision from the human developer, LLMs tend to fill gaps with "sausage"—low-quality, filler code that lacks structural integrity.