Launch on frontier models first, post-train to cut costs later: the standard playbook
matt_slotnick · x · 2026-09-25
- Matt Slotnick lays out a recurring pattern for shipping AI features: new functionality works best out of the box with frontier models, so launch with over-provisioned intelligence rather than worse models or prohibitive pricing.
- Upfront optimization is unwise because you lack the data; the P0 is identifying the valuable workload, and models can be swapped quickly.
- Once discovered, you can play with both numerator and denominator — charge more or serve cheaper via post-training to fit the workload. Every major new feature follows this launch-learn-iterate-post-train loop.
Related event: Practitioners Advise Overprovisioning Frontier Models at AI Feature Launch(4 posts)→
More from coding & agent
- Raft Source-Available Repo Live: Server, Web, CLI and Agent Daemon Code Included — aigclink · 2026-09-25
- Startup Open-Sources Raft, a Slack Where AI Agents Work as Colleagues; 3 of 5 Co-founders Are AI — aigclink · 2026-09-25
- Autonomous agent ships 181 PRs of testing and bug fixes before being stopped — gregmushen · 2026-09-25
- AI turns months-long coding projects into weekend POCs, but integration pain remains — RexDouglass · 2026-09-25
- Dev argues true multi-agent systems don't exist yet — ultra mode is self-delegation — blaizedsouza · 2026-09-25
- Alex Xu's side-by-side chart explains MCP vs Function Calling, a common engineer confusion — blaizedsouza · 2026-09-25