#LLMs

George Hotz published a blog post arguing that while LLMs are transformative, frontier AI labs' valuations are inflated because the value they create will be captured by open-source commoditization, not proprietary models. This challenges the dominant narrative of AI mega-cap valuations and highlights the tension between proprietary and open-source AI, which could reshape investment strategies and industry dynamics. Hotz argues that productivity improvements from LLMs have not yet led to visible new software products, suggesting value is being captured privately (e.g., in homelabs) rather than by frontier labs.

The article draws a parallel between the film industry's shift to CGI and the software industry's adoption of LLMs, warning that over-reliance on LLMs could devalue core coding skills. This analogy highlights a potential long-term risk for software engineers: while LLMs boost productivity, they may erode fundamental skills and craftsmanship, much like CGI diminished practical effects artistry. The article notes that writing every line by hand is no longer the norm, but reading and understanding architecture remains crucial. It also mentions that tests are more important than ever since large refactors are common.

Several Asian AI startups have released models comparable to Anthropic's Mythos, such as Sakana AI's Fugu Ultra, a multi-agent orchestration system, while Anthropic's export restrictions on Mythos remain in place. This development signals a shift in AI leadership, as Asian startups begin to compete with Western frontier models, potentially reshaping global AI supply chains and prompting regulatory responses. Fugu Ultra is not a single model but a learned multi-agent orchestration system that routes tasks across a pool of underlying models, as described by OpenRouter. Early user reports indicate it can be slower and more costly than Anthropic's Opus.