From my initial testings GPT-5 is as bad as GPT-4o on instruction following. This is a real bummer. No improvement at all – but I need more testing to validate this.
GENERATIVE AI
-

AI Video Editing: Consistent Modifications Without Unintended Side Effects
By
–
Using AI, you can modify parts of videos without the annoying side effect of subtly changing everything else. Lighting effects (as seen on the shirt) are very consistent. pic.twitter.com/RsROwR3agJ
— Varun Mayya (@waitin4agi_) 8 août 2025Using AI, you can modify parts of videos without the annoying side effect of subtly changing everything else. Lighting effects (as seen on the shirt) are very consistent.
-
GPT-5 hallucination claims questioned in system card
By
–
If every output is a hallucination, statements like "We’ve made significant advances in reducing hallucinations" from the GPT-5 system card would be nonsensical
-
Different GPT versions across devices and browsers
By
–
When I open ChatGPT in safari with my MacBook I have GPT-5.
When I open ChatGPT in brave on a Windows computer I have GPT-4o.
My own account is logged in.
In the iOS app again GPT-5. ?? -
Are LLM Factual Errors Hallucinations? A Definitional Question
By
–
Follow-up question: does "there are three Bs in blueberry" count as a hallucination? (I expect these poll numbers will be much closer to each other)
-

OpenAI Python Library v1.99.3 Release Notes
By
–
These release notes from them made me chuckle https://
github.com/openai/openai-
python/releases/tag/v1.99.3
… -
AI CLI Tools Effectively Navigate Large Codebases
By
–
I've been using it with the Codex CLI and it's extraordinarily effective at using grep and sed to navigate code a few hundred lines at a time, so in practice the overall context limit doesn't limit its ability to work with large codebases nearly as much as I expected
-
Mixed Reactions to AI Model: Routing Issues and Inconsistent Performance
By
–
At least on X and Reddit, the mood seems to be rather disgruntled and disappointed. Biggest problem: routing. It just works incredibly poorly. In my testing as well. You should manually enable reasoning as needed. Additionally: it's better than GPT-4o and o3 – but often not
-
GPT-5 Outperforms Claude Code in Early Codex CLI Tests
By
–
Premiers retours. GPT-5 est vraiment excellent pour le code. Le modèle est très performant. D’après les premiers tests, Codex CLI avec GPT-5 remplacera probablement Claude Code comme outil de codage quotidien.
-

Router overconfidence limits question routing to simpler models
By
–
El problema es que el enrutador sobre-confía en la dificultad de las preguntas, tirándolas al modelo más simple y fallando. En esta pregunta no ha querido razonar y GPT-5 se ha equivocado. Pero si se la mando directamente al GPT-5 Thinking, sólo pensando 10 segundo, acierta.