OpenAI chart reveals unreleased GPT-6 Astra models; Domingos questions billions for <0.3 bits
pmddomingos · x · 2026-11-23
Pedro Domingos shared an OpenAI base-model log-perplexity chart and questioned whether tens of billions of dollars are worth less than 0.3 bits of improvement.
The chart plots GPT-3.5 through GPT-4.5 against capability thresholds (fluent conversation → thinking → useful coder → cybersecurity expert → solving a Millennium Prize problem) and names unreleased models: GPT-5/5.1, 5.2, 5.3-Codex/5.4, 5.5, 5.6 Sol, GPT-6 Astra, GPT-6.2 Astra, plus an even lower "Internal model", extending past 2026 — a notable leak of OpenAI's roadmap and diminishing returns at the scaling frontier.
More from Infra
- KV cache offload to RAM runs Qwen 3.8 with 8x200k streams on 4x3090 — Mr_Moonsilver · 2026-11-23
- wllama author: RPC-linked wasm instances show little performance loss, credits GGML RPC backend — ngxson · 2026-11-23
- layerpool open-sources layer-split fine-tuning across machines with unchanged HF Trainer — Historical_Syrup_461 · 2026-11-23
- wllama v4.0.0 ships: in-browser llama.cpp up to 2.8x faster with WebGPU hybrid — ngxson · 2026-11-23
- French dev laments missed chance to build green data center ecosystem on low-carbon grid — Dimillian · 2026-11-23
- Fchollet: deployment friction may be the best thing for AI infra, stretching the buildout timeline — npinto · 2026-11-23