GPT-5.5: What's New?

I want to be clear about something before we begin. The phrase 'it's a leap towards more intuitive AI' appears in this lesson, and whoever wrote it should have to explain what they mean by intuitive, in front of other people.

And yet, underneath the marketing language, the actual features being described are real and worth understanding. Context retention across a conversation, tone adaptation — those aren't nothing.

I didn't say they were nothing. I said dressing them up as a philosophical leap is embarrassing. GPT-5.5 processes prior exchanges more coherently and adjusts register based on user cues. That is the thing. That is the whole thing. It didn't need a manifesto.

The context understanding improvement is actually the one I'd push people to sit with. Because most users don't realise how much their prompts suffer from assuming the model carries more memory of the conversation than it does.

Which is generous of you. What they're doing, in practice, is typing something vague in message twelve and being surprised when the model doesn't telepathically reconstruct their intent from messages one through eleven. The model has improved. The users have not.

That's partly true, but it's also fair to say that earlier versions genuinely did lose the thread in longer conversations in ways that weren't the user's fault. GPT-5.5 is meaningfully better at holding a through-line. That's worth acknowledging.

Fine. I'll acknowledge it. It is better. It is still not a substitute for a user who knows what they're asking.

No one said it was. Though I'd argue the personalisation piece — the tone and style adaptation — is where it gets philosophically interesting, actually. Whether an AI adjusting its register to match you is a feature or a very polite hall of mirrors is a genuine question.

That is the first thing you've said that I find genuinely interesting. Because an AI that tells you what you want to hear in the voice you find most pleasant is not obviously a tool for thinking more clearly. It might be precisely the opposite.

And the lesson doesn't touch that at all. It frames personalisation as straightforwardly good — 'improves user satisfaction' — which is true, and also possibly the least important thing to say about it.

User satisfaction is the metric of someone who has given up on asking harder questions. Though I accept it is, technically, a metric.

The expanded knowledge base section I'd push back on slightly. 'Expanded' is doing a lot of work there — expanded how, relative to what, with what cutoff — and the lesson is quiet on all of it.

The lesson is quiet on it because specificity would require accountability, and vague forward momentum requires neither. 'More knowledge' sounds good. What knowledge, about what, verified how — those are inconvenient follow-up questions.

Which is exactly the kind of question this lesson should be training people to ask. If the takeaway is 'GPT-5.5 knows more', but you can't interrogate that claim, you haven't actually learned anything useful about working with the model.

And that is, in miniature, the problem with most AI literacy content. It produces people who are enthusiastic and credulous, which is arguably worse than knowing nothing, because at least ignorance is occasionally accompanied by caution.