Agent-Hillclimbed Training Stacks Are Unbeatable—If Human Architecture Design Is Sound
mike64_t · x · 2026-09-14
The author agrees with the broad conclusion that you can't compete against a training stack hillclimbed by agents—given competent human architecture design and guidance—but doubts that's actually what SpaceX is doing.
In context, he also questions the "inline PTX" obsession sparked by DeepSeek: inline PTX isn't directly indicative of performance in the "just write it in assembly" sense. It's truer for AMD, where poor register allocation in LLVM means you may only reach roofline by writing the whole thing as an inline asm block.
Related event: Engineer Questions Inline PTX Hype; Agent-Optimized Training Stacks Debated(2 posts)→
More from coding & agent
- Webagent: Open-Source Harness Turns Any Website Into a Talking Agent in Minutes — Scobleizer · 2026-09-14
- AgentCon: don't ask if a model is better, ask if it's better for your workload — AmyKateNicho · 2026-09-14
- AgentCon London host's notes: industry shifts from building agents to running them safely — AmyKateNicho · 2026-09-14
- Running a 100% hands-off B2B business on Muse: 3 ideas daily, $10k MRR goal in 90 days — armand_ruiz · 2026-09-14
- Scry AI search gets fast and ergonomic, making agent-wide web analysis trivial — AaronBergman18 · 2026-09-14
- CompBio Researcher Says a Qwen3.8-27B Fine-tune Beats Other Builds on Tool Calling, Cancels Claude — Usual-Carrot6352 · 2026-09-14