跪拜 Guibai
← All articles
Backend · Programmer · Artificial Intelligence

GPT-6.2 Astra Launch Draws Coordinated 'Don't Use' From Prominent AI Figures

By cxuanAI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The model's reception warns teams to validate GPT-6.2 Astra inside their own pipelines rather than trusting demos: long-horizon coding tasks still produce bizarre questions, dead-end loops, and premature halts that price cuts can't fix. A coordinated 'do not use' from credible practitioners makes this release a test case for how the community polices hyped AI launches.

Summary

GPT-6.2 Astra shipped alongside Sol, but the launch buzz never materialized on timelines, a symptom of the tide turning against it. Within hours, three respected voices in the AI space — Matt, Theo, and Dan Shipper — came out against it in a coordinated way. Theo, even while ill, wrote a lengthy hand-typed critique: the model excels at 3D and game dev, has strong computer-use ability, and writes good subagent prompts, but it also has no sense of its own knowledge gaps, asks irrelevant questions during long-horizon tasks, produces explanations only it can understand, and gets stuck in infinite loops that ignore prompts. Dan Shipper reported similar trouble: writing has regressed, coding stops prematurely, and AI slop persists.

The author ran his own demos, including a volcano and a rain-curtain greenhouse, and saw no obvious progress. He argues that demo-level testing misses the real problems, which only surface when the model is embedded in actual workflows. The lesson for evaluators: a pretty presentation and a price cut don't matter if the model can't carry a real project. His suggested check — whether you'd voluntarily come back to the model within 15 days, and after a newer model arrives — puts the burden on daily usability rather than launch-day hype.

Takeaways
— GPT-6.2 Astra launched with Sol, but the release got notably little attention on developer timelines.
— Three prominent AI figures (Matt, Theo, and Dan Shipper) jointly recommended against using the model.
— Theo gave it strengths in 3D/game development, computer use, code quality, and subagent prompt writing.
— Theo also reported that the model lacks knowledge-blind-spot awareness, asks irrelevant questions in long-horizon tasks, and gives explanations that only it can understand.
— The model can fall into infinite loops that ignore the user's prompts.
— Dan Shipper found writing quality regressed, coding often stopped early, and AI slop remained severe.
— In the author's own demos, volcano and rain-curtain greenhouse outputs showed no obvious improvement.
— The author argues demo results don't reflect real-workflow performance; a practical test is whether you'd still use the model 15 days after launch.
Conclusions

A coordinated public 'do not use' from three respected voices at launch is rare, and it turns the release into a credibility test that marketing can't quickly repair.

Demo success and long-horizon reliability are diverging sharply for agentic models; the failures mentioned here only appear inside real workflows, not in curated tests.

A practical 15-day, do-you-come-back-to-it test beats benchmark numbers: it measures whether a model earns a place in daily development without heroic prompting.

Concepts & terms
Subagent
In agentic AI, a subagent is a secondary agent spawned by a primary agent to handle a specific subtask; the primary agent coordinates them. The article notes GPT-6.2 Astra can manage many subagents but asks bizarre questions while doing so.
AI slop
Low-quality, formulaic content generated by AI that often looks plausible but is repetitive or meaningless. Dan Shipper saw it remain severe in GPT-6.2 Astra's output.
Long-horizon tasks
Complex tasks requiring many sequential steps over an extended period, where an AI model must maintain context and plan ahead. The article uses this to describe where GPT-6.2 Astra's problems surface.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗