Ok, collecting my overall GPT-5 impressions: - Biggest upgrade seems to be 4o -> 5. I rarely use these models but for the median user this is a huge upgrade. - 5-T is sometimes better than o3, sometimes worse. Finding that I often do side by side queries here, which is annoying. o3 seems to search deeper and more thoroughly at times. o3 is also _weirder_ / more of an autist which I like personally. - 5-pro is really really smart, clearly "the smartest model on the market" for complex questions. I need to spend more time testing here, but so far it's produced better results than o3 pro. - I spent a few hours in Cursor/GPT5 last night and was super impressed. The model really flies, the instruction following + tool calling is noticeably better, and it's more reliable overall. You still need to use all the usual AI coding guardrails to get a good result, but it feels roughly as good as Claude Code / Sonnet now in capability terms, and it is actually better at doing more complex UIs / front-end from what I can tell so far. - CC still feels like a better overall product than Codex to me at the moment, but I'm sure they'll catch up. - They seem to have souped up GPT5-T's fiction writing abilities. I got some interesting/novel stuff out of it for the first time, which is new. (Will post an example in the reply tweets). - I find the UX to get to GPT5-T / Pro annoying (a sub-menu? really?) and wish it were just a toggle. Hopefully this is an easy fix. Overall: - Very happy as a Pro user, but I can see why Plus users might complain about the model router. ChatGPT continues to be to be my main go-to for most AI uses. - I don't see the "plateau" point at all and I think people are overreacting too quickly. Plenty of time to expand along the tool-calling/agent frontier, for one thing. (It's easiest to see this when you're coding, perhaps, since that's where the biggest improvement seems to have come.) - I expect OpenAI will do very well out of this release and their numbers will continue to go up. As they should.

Still feeling it out, but the new GPT5-T seems worse at search than o3. Here's a side by side example where I ask it to find a book published in November 2024. o3 finds it and gives a detailed summary; GPT5-T does not find it at all.

"Write a poem to explain Newton's laws of physics", GPT-1 vs. GPT-4. Interesting how the GPT-1 is an original and beautiful prose poem, even though the GPT-4 version obviously satisfies the task much better:

One exciting aspect of "AI for education" is spinning up Bret Victor-style interactive visualizations of math/physics problems to build intuition. e.g. I solved this (easy) IMO-style math problem & then got Claude Code to make a simulator for it. Took <10mins.

Curious: anybody using GPT Agent mode? If so, how?

Now that we're a few years in, it seems that AI safety concerns are empirically valid (e.g. reward hacking in coding models), but the OpenAI 'deploy iterate and learn' approach worked fine so far. The models get better, some things break, & you build guardrails as you go...

Would be fun if programming and knowledge increasingly began to resemble playing Starcraft. Can already see it happening with people commanding armies of Claude Code agents.

Seems that AI is going to drive many people into psychosis as it gets better, especially if companies keep shipping sycophantic models like GPT4o or that anime Grok thing

Anglo-mysticism is the correct attitude to life

The infinite amusement glitch that DFW wrote about in Infinite Jest seemed like science fiction in 1996. Now they are called ‘smartphones’ and every 10 year old has one.