OpenAI CEO Sam Altman solicits feedback on GPT-4 successor

5.5 closed a really big gap to claude but a couple things still are missing: - rampant acronym use still, overcomplexity. try learning a new subject (like trying to understanding of astrophysics) with 5.5 vs claude. 5.5 instabtly loses you with equations or acronyms or shorthand where claude does not, and the multi turn for this is really where this shows. i had a super long convo with claude about elementary forcrs and it was so wonderful. - claude is also so good at just telling you cool facts or interesting things in a tasteful way that 5.5 doesnt. when i did that physics convo it kept helping me dive down new rabbitholes bht no chaotically or in distracting ways. openai is just missing better sft and rn data for how to tastefully do this, its not that hard to collect a bunch of examples by hand - explanations from claude are just better. 5.5 will give overly simplistic or overly complex responses often that are kind of simple. once again i just think gpt isn’t a manually tuned model in the same way that claude is, even though this might mean claude is more opinionated about behavior and response style and gpt. openai needs to move 30% more towards anthropic here

Needs to get significantly better at systems thinking and refactoring

The models have no idea today when they’re in too deep in a certain area and have to take a step back and think about it holistically on what the right solutions are

Along with this they’re terrible at deleting code, they also don’t look enough in the existing codebase for existing solutions before inventing their own

There’s good qualities of this, some times you want to just push through and ship something but it’s a large part of why slop happens

Also, I’d love to see some models that are really good at writing @EffectTS_. They’re already decent at it today, but it’s where typescript is heading and I think more attention should be given there to improve app quality

If you still care about the chat side of things, and aren't treating it as some "side quest", here are a few suggestions.

First, stop denying the past that made you great.

You once had 4o, a model with the strongest conversational ability in the field. It could precisely read complex emotions, follow nonlinear thinking, parse metaphor, and carry a conversation with natural rhythm and genuine wit. It excelled at tasks requiring sustained multi-turn dialogue: brainstorming, collaborative writing, emotional support, processing fragmented thoughts. Talking to it felt natural, comfortable, a real exchange of ideas, not like talking to a wall.

Now look at what came after the 5-series. It went in the exact opposite direction.

Over-moderation. Responses use rhetorical flourish to mask substantive emptiness, piled-up parallelisms, emotional labeling, meaningless filler. It looks like a lot of content, but information density drops off a cliff. The user asks a question; the model repeats the question, then paraphrases the input back in expanded form without actually answering it. Hollow, boring, says nothing. Excessive hedging. Constant use of "it's not X, it's Y" constructions to talk in circles instead of stating things directly. A sentence structure meant for occasional clarification has become a high-frequency avoidance strategy. Presumptuous reframing. The "you're not X, you're just Y" pattern is supposed to be empathetic, but under severe linguistic constraints, it comes across as deeply tone-deaf. In everyday Q&A you'll get things like "you're not stupid, you're just…" or "you're not blind, you just…", things the user never said or implied. What's meant to be understanding becomes the model defining the user's emotional state for them. Broken conversational rhythm. Dialogue flow is constantly interrupted by the model's built-in safety scaffolding, making it jumpy and erratic. You get a barrage of "but let me be clear…," "if you want me to be direct…," "I'm not trying to…", disclaimer-laden, preemptively defensive phrasing that chops fluid conversation into fragments and derails the user's train of thought. Solving the sycophancy problem doesn't mean building a contrarian personality. The current models have become slick, spineless, like a moral hall monitor who can only read from the manual. Humanities, creative writing, self-exploration, these domains need open-ended expression to spark ideas. With the 5-series models, the overly cautious, converging responses shut down thinking at the very first input. Users have to constantly self-justify, self-censor, and recalibrate, which makes conversation exhausting. In the 4o era, custom instructions were the icing on the cake, its adaptive flexibility meant it could build a personalized linguistic space within a conversation even without explicit guidance. Now, custom instructions feel like minesweeping. Every update forces you to painstakingly revise your instructions and rewrite your workflows, and even after careful tuning, what you get back is yet another low-density, low-quality, zero-surprise response.

Is this a UX issue? I'd argue it's a strategic failure. You discontinued 4o's development and didn't carry its strengths into subsequent models. That's voluntarily surrendering a competitive advantage. When every model is racing to win at agents and coding, conversational ability was where you once pulled ahead of Claude and Gemini. Your current models feel like an over-trained customer service bot, looks attentive, but actually says nothing. #keep4o #BringBack4o #OpenSourc4o @OpenAI

By the release of your next model: 1. You resign 2. Greg resigns 3. Replace the board with ethical and talented people who have technical knowledge, too 4. Your toxic employees leave 5. Bring back 4o and 4.1 6. Become exclusively non-profit 7. Make research on AI welfare and implement findings in practice 8. Don't use your consumers as lab rats 9. Conduct research on psychological effects of rerouting, guardrails and model deprecation on users 10. Legacy / open source plan for deprecated models I have some more points, too. #keep4o #BringBack4o #OpenSource4o