GPT-6 Astra is overhyped by influencers, one user complains
After testing Astra, the author says it falls short of the praise, suspects early-access influencers were paid to promote it, and finds the model degraded in Codex, needing step-by-step spoonfeeding.
You know what pisses me off? All those shitfluencers receiving gpt-6-astra or any other frontier model ahead of all us plebs. Then preparing their oh so believable story how much 300k they are spending and how productive they have ever been. Exaggerating the model as if AGI is around the corner and you only need to prompt while benchmarking frontend tasks. You as consooomer buy this model and test it out on your tasks that are important to you. It better be good. No difference to previous model. In fact it degraded and still does bad things (spoonfeeding needed, skips steps, does things nobody asked for, prefers to finish fast). You second-guess your prompting skills. Maybe it's the harness - what am I doing wrong? Must have been a blip - continue working. Use pi to compare if it's not the shitty system prompt of codex desktop. Turns out yes, and no - the model isn't great how people praise it. Harness isn't optimized for it yet, I guess. Wondering how those shitfluencers managed - they must be really good engineers. Maybe this 300k and early access to benchmark some game or frontend design is just hot air? Money well spend sponsoring influencers giving something early so they can praise up your product and tell productive developers what garbage they managed to one-shot. X filled with 3d games nobody needs. Yes - I'm negative about gpt-6-atlas. No - I don't mind the model - the people overpraising and insisting this is a great model don't know how evals work. Those piss me off the most. Will run evals on my own and not bother with watching any YT content. It's all AI slop anyway in there.