then it then went away and generated a load of confidently incorrect total bullshit

"phd level" my backside

[1] https://www.grc.nasa.gov/www/k-12/VirtualAero/BottleRocket/a...

While true, the person writing this article does not seem to understand the difference between flying inverted and flying with a negative angle of attack. These can happen at the same time, but not necessarily. If you're performing a loop or a barrel roll, you will be inverted, but the aircraft will be performing largely as it would be when you are straight and level, because you are still under positive g with a positive AOA on the aircraft. The lift vector will just be pointed someplace other than "up."

Naturally. This is how LLMs work. It regurgitates the data fed into it.

Interesting that will all the Ph.D. expert fine-tuning that GPT5 supposedly received, it still doesn't favor the more correct Newtonian explanation of airplane lift.

Especially the code generation part. It feels to me like Claude Web can do those illustration artifacts already for months equally well.

Also the example in Cursor just felt like a regular Claude Code session, just with different UI.

The only part I'm excited about is, that there is no distinction between reasoning and non-reasoning models anymore. I tend to default to reasoning models, because too often I feel like I need to switch mid-conversation to a reasoning model anyway. And reasoning models degraded the user experience drastically, because it often takes them quite some time to start responding.