← Wednesday, November 18 deck

71Opinion2 sources posting

Theo Retests GPT 6.2 Astra After OpenAI's Fixes and Says It Is Better but Still Misses the Point

I took this to heart and spent the last few hours doing another deep dive on GPT-6.2 Astra tl;dr - things are better than when I tested, but IMO still "directionally, very bad" I'm going to keep using the same example I've hyperfixated on since the start: "why is creating a worktree so slow?" I had a long thread trying to diagnose perf issues while using t3 code on a particular VPS. 2 of the 3 were fast, 1 was really slow. I asked astra outright "why is this so slow?", and it said a bunch of gibberish. I followed up with "I feel like I'm going insane, can you just tell me the cause of the slowness?" I'll give you the answer upfront: this VPS uses a virtual disk that is slow as shit, my other VPSs have physical NVME drives that are faster. Image 1 is from before the fixes Tibo mentioned. It shows an unnecessary graph, focuses too much on pnpm and file systems, bolds irrelevant text, and brings up ram just as much as it (subtly) mentions drive speed. Image 2 is of Opus asked same thing. It opened with the important part in bold: "Slow disk, made worse by VDO". Ideal! Image 3 is 6.2 Astra post-fixes Tibo mentioned. Better? Absolutely. Much more concise. Sadly, it still focuses way too much on second-order problems that are a small % of the cause. "XFS over a compression and deduplication layer on a virtual disk." - so close, but still a miss. I think this perfectly encapsulates the problem I have with 6.2 Astra. It doesn't miss anything. It's SO thorough. But in it's thoroughness, I find it misses the most important thing: the actual point. It wants to know everything, and then it wants to share everything. It doesn't understand what is and isn't important. Did it say anything wrong in my screenshots? Not really. But IMO, it still entirely failed to miss the point of my question, and gave an answer that was too focused on accuracy and not enough on clarity. In the wise words of Michael Seibel: "Most aim for 100% accuracy, and they end up with 50% clarity. Aim for 100% clarity, 80% accuracy". Silly way of putting it, but I genuinely wish OpenAI models would lie to me a LITTLE more so I could better understand the vibe, instead of having my time wasted by all these useless details.

Context

Why it matters

A developer who criticized an early build says OpenAI's late fixes helped, a useful data point in the split reaction to GPT 6.2.

OpenAI launched GPT 6.2 Astra and Sol on Nov 17. OpenAI's Tibo Sottiaux said some early reviewers had tested older builds before a round of late changes. Theo is a developer who posts as t3.gg and had criticized the model after testing an earlier build.

Deep dive

Theo posted that he spent several hours on another deep dive of GPT 6.2 Astra after Tibo Sottiaux of OpenAI said early reviewers had tested older builds. His summary was that things are better than when he first tested, but in his view still directionally very bad.

He used one example: a long thread diagnosing why creating a worktree was slow on one of three servers. He wrote that the real cause was a slow virtual disk, while his other servers use faster physical NVMe drives. He shared three screenshots. The first, from before the fixes, showed Astra producing an unnecessary graph and focusing on pnpm and file systems. The second showed Opus answering the same question by leading with the slow disk. The third showed Astra after the fixes, which he called much more concise but still too focused on second order issues.

He argued that Astra does not miss anything and is very thorough, but in that thoroughness misses the actual point. He said it did not say anything wrong in his screenshots, yet favored accuracy over clarity, and quoted Michael Seibel on aiming for clarity over accuracy.

This is one developer's hands on opinion based on one example, not a benchmark.

What to expect next

Watch for more retests from early reviewers now that OpenAI says the shipped builds differ, and whether broader user reports agree with Theo's view on answer focus. OpenAI may ship updates aimed at taste and ease of use, which Sottiaux said the team wants to improve. Side by side comparisons with Opus on everyday debugging tasks are likely to keep shaping how developers choose between the two.