What comes after GPT-6 Astra: from chatbot to a coworker you leave running for a week
johnseach · x · 2026-09-06
A long-form forecast built on OpenAI's Astra (shipped 3 Sept 2026: SOTA on computer use/coding/cyber/science, 1.05M context, $10/$50, Critical-tier cyber). Key predictions for the successor: the axis is time-horizon (days-to-weeks of unsupervised work with tool failures and memory across compaction), multi-agent as the runtime (a foreman splitting and merging work), adaptive reasoning depth with a split clean-consumer vs dense-internal trace, desktop-first boring demos before labs/robots, science-as-a-product pipelines, and cost as the real constraint. Bel/10T pretrain leaks remain single-source rumor; Astra is a model family and a 2027 named generation is plausible. The result: less answer engine, more brilliant but occasionally opaque junior colleague, shipped in gated pieces.
More from AGI Musings
- Pinker argues doomer extinction narratives starve real AI risks like cybersecurity — Dr_Atoosa · 2026-09-06
- Researcher Argues AI Doomer Extinction Narrative Crowds Out Cybersecurity and Biosecurity Risks — Dr_Atoosa · 2026-09-06
- AI slop in technical writing is a leading indicator of a new medium, says dev — threepointone · 2026-09-06
- Talking about AI x-risk boosts concern over near-term harms, not distracts from them — gleech · 2026-09-06
- Alignment spat escalates: critics cite thousands of rogue agents as evidence against Enlightenment Now — JMannhart · 2026-09-06
- hnshah: Thinking too long may now cost more than building the wrong thing — Paimaamu · 2026-09-06