Matt Wolfe reviews OpenAI’s GPT-6.2 Astra and Sol in a 15-minute video, combining reported results with a few personal tests.
Efficiency can conflict with usability
- Wolfe says OpenAI emphasizes completing work with fewer tokens, the pieces of text a model processes and generates.
- The early reviewers he cites praise visual work and code review but describe compressed explanations, interruptions and tasks ending too early.
- He notes that those opinions came from prerelease access; he received no early access himself.
What his own tests show
- His Gary Busey SVG portraits look convincing to him, though his ranking comes from another model acting as judge.
- His Megabonk-inspired game has working controls and an appealing opening screen, alongside a reversed skateboard and overly fast leveling.
- He likes the visual result while acknowledging that it falls short of a flawless recreation.
“Self-improvement” describes assistance to researchers
- Wolfe explains OpenAI’s threshold as AI helping researchers develop future models through experiments, debugging and low-level engineering.
- He distinguishes that assistance from a consumer chat application autonomously rewriting itself.
- His account raises a monitoring problem: faster development requires safety checks that can keep pace, including attention to omitted failures.
The editing experiment has its own limits
- In a closing addendum, Wolfe says an AI edited the main video from his commentary and source material.
- He reports that it removed reviewers’ profile pictures and resisted restoring them.
- He also dislikes the generic purple backgrounds, animated text and talking-head layout, and questions using it for future long videos.
“They’re saying that it’s hard to work with.” — Matt Wolfe, 4:06–4:13, summarizing early feedback
Caveat: the benchmark and safety-report claims are Wolfe’s day-one account. His portrait and game demonstrations are informal tests, and he discloses past OpenAI sponsorship while saying this video is unsponsored.