Hardly Promethean
3 hours ago
- José Valim questions why, with powerful tools like LLMs, developers cannot find ways to significantly improve their tech stacks.
- Dan Luu argues that LLMs have drastically reduced the cost of specialized performance work, making it accessible to many, but notes limitations like overfitting benchmarks and poor experimental design.
- Varun Gandhi counters that while LLMs lower costs, the dominant costs are still shipping, maintenance, and avoiding regressions, not just writing code.
- DHH's keynote generalizes his own successful use of LLMs (e.g., rewriting Hey Next) to all programmers, ignoring that he works in a unique, unmanaged environment.
- Gandhi points out that LLMs shift bottlenecks from writing code to post-development tasks, and DHH's solution is to skip these steps entirely.
- DHH's increased tolerance for messy agent-driven code doesn't apply to others who care about code quality and maintainability.
- The real scarcity is not technical or budgetary, but a lack of new ideas; DHH's projects with LLMs (e.g., calculator, video editor) are not novel.
- LLMs excel at replicating existing products, not generating original concepts like Rails was in 2004, highlighting a deficit in innovation.