Custom Midtrain Plus Own RL Matches Astra Max at Half the Inference Price
hsu_byron · x · 2026-09-16
A debate sparked by andrewho03: is it "bad" to spend enormous effort collecting your own data, running a custom midtrain, and doing your own RL just to match Astra max at half the inference price?
madiator reads it the opposite way — replicating frontier performance with your own pipeline at half the cost is remarkable, and worth imagining what comes next. The exchange captures the build-vs-frontier debate: copying frontier capability is getting dramatically cheaper.
More from Infra
- Chipotle partners with Palantir on Foundry-based food safety risk platform — eliano · 2026-09-17
- Emerald AI, Google and NVIDIA launch alliance for flexible AI data centers — ArtificialOther · 2026-09-17
- Vercel makes Secure Compute and Static IP builds 64% faster with prewarmed containers — cramforce · 2026-09-17
- Why specialized inference engines are multiplying: generality vs. specialization in vLLM/SGLang era — sh_reya · 2026-09-17
- LLM Inference After Training: a walkthrough of prefill, decode and KV cache — kmeanskaran · 2026-09-17
- Anthropic engineer's plea: 'I just want to serve five petaflops' — charles_irl · 2026-09-17