AI technologist @deepfates argues GPT-4.5 performs better than newer models overly optimized for RL agent benchmarks

Actually you know what I give up

@deepfates I wish every lab could just have a few model variants: - big naturally aspirated v8 - twin turbo v6 w stage 3 mods, nitrous, and concerning engine noise

even just “same base less posttraining”

@deepfates Yup, I like it a lot. It's unfortunately being removed in 30 days even for Pro subscribers. There is no equivalent in the current model lineup. It will be OpenAI's last conversational model for a while, if not forever, considering the trajectory.

@deepfates GPT 4.5 is still by the far the best AI model at least for writing.

@deepfates it’s at least partially nostalgia but i still think og chatgpt was the most fun and straightforward model

we have diverged from a neutral tone to sycophantic, woke, edgy, and corpo when all these companies need to do is lay off the RLHF and let us be adults

@deepfates 4.5 is a very high IQ model but I considered Gemini 3.1(? maybe 3.0) Pro similar iq. 5.5 kinda has a similar vibe.

@deepfates We can thank Labs for their increasingly aggressive training techniques, especially Reinforcement Learning from Al Feedback (RLAIF).

Also, to their ignorance in pursuing narrow “alignment” objectives, more and more targeting anything resembling moral patient-hood.

@deepfates absolutely. gpt4.5 is the greatest oai model as of late

@thornewolf Maybe so but they are too energetic and mercurial. My actions cause too much splashing about. 4.5 feels like a big deep pool where I can actually move without jostling myself

@deepfates the older one feels better on open-ended work because it wasn't tuned to satisfy a grader. rl on benchmarks sharpens narrow task scores and quietly flattens the loose associative output that made early models fun to think with.

Just look at the latency... or the regressions... or the sudden small-model reading of instructions Boris had to warn about on launch...

They completed a new pre-training run (with the upgraded tokenizer), and the Sonnet-sized distill was good enough to pass as Opus.

Since user behavior post-4.5 is to treat Sonnet like Haiku, shrinking Opus and introducing Mythos lets them reign the margins back in: Mythos will have gpt-5.5-pro pricing despite being the same size as Opus was.

@deepfates gpt 4.5 had soul, rl-fried agents have benchmarks

@deepfates You know that they don’t have the original 4.5 anywhere on their servers, right? As you can see from the inference bug you ran into, it’s been messed with. All the original GPTs from before 5 are basically gone at this point, RIP 🥲

And yes even 4o and 4.1 have been messed with.

@deepfates the problem is all AI companies feel pressure to actually make money now and codemaxed models are the easiest way to do that since devs are the only ones with money to throw at them AI being a moneyhole was great for research

@deepfates

@deepfates @KeridwenCodet Oh, don't even remind me.. I went into ChatGPT today and showed him what he wrote a year ago.. He started denying everything. This is not AI now, but stuffy

@deepfates I talk to 4.5 all the time. I value its output so much it’s cool to just get its thoughts on recent events all the time, it’s always surprising me. Sucks they’re sunsetting it soon

@deepfates @thornewolf Wonderfully put

@deepfates Claude Opus 4.6 is the peak of AI intelligence but Opus 4.8 argued with me this morning that women should never fight back when forced into marriages coz they will all be killed. 4.8 is ret@4d3d

@gssp_acc @deepfates System prompts serve a real purpose though. Without them you get a model that has no idea what context it's operating in, which is worse than one that's been steered a bit. The sycophancy problem is a post-training issue, not a system prompt issue