Frontier models may build the assets, but traditional software captures most operating gains
curious_vii · x · 2026-07-25
The author argues that most operating-expenditure gains will come from traditional software and data work, not from self-hosted weights. In their view, the big wins come from placing products where none existed before, rather than trying to force “agents in production” everywhere.
The quoted thread adds a broader theory: frontier models are best at producing capital assets, while open-weights models are better at operating them. The implication is that both will have a place, but the economic sweet spot often favors paying for frontier models when they retain even a small lead.
More from AGI Musings
- Gary Marcus says AGI still means doing all expert-level human tasks — GaryMarcus · 2026-07-25
- Substack essay links the Jacobian conjecture to AI’s changing role in math careers — kylekabasares · 2026-07-25
- AI productivity gains may take 10 years to become macro gains, Garry Tan says — garrytan · 2026-07-25
- The Ultimate Form of AGI: Beyond a Hyper-Honest Machine — DionysianAgent · 2026-07-25
- A repost mocks a 2022 AGI definition as poor intuition and hindsight — soumitrashukla9 · 2026-07-25
- Guardrails for closed models do not map cleanly onto open-source AI — evijit · 2026-07-25