Specific model I was using was o4-mini-high which the drop-down model selector describes as "Great at coding and visual reasoning".
I wonder if the advice on prompting models to role play isn't backfiring now, especially in conversational setting. Might even be a difference between "you are an AI assistant that's an expert programmer" vs. "you are an expert programmer" in the prompt, the latter pushing it towards "role-playing a human" region of the latent space.
(But also yeah, o3. Search access is the key to cutting down on amount of guessing the answers, and o3 is using it judiciously. It's the only model I use for "chat" when the topic requires any kind of knowledge that's niche or current, because it's the only model I see can reliably figure out when and what to search for, and do it iteratively.)
I've noticed this couple times with o3, too - early on, I'd catch a glimpse of something like "The user is asking X... I should reassure them that Y is correct" or such, which raised an eyebrow because I already know Y was bullshit and WTF with the whole reassuring business... but then the model would continue actually exploring the question and the final answer showed no trace of Y, or any kind of measurement. I really wish OpenAI gave us the whole thought process verbatim, as I'm kind of curious where those "thoughts" come from and what happens to them.
To agree with your point, even with the real CoT researchers have shown that model's CoT workspace don't accurately reflect behaviour: https://www.anthropic.com/research/reasoning-models-dont-say...
> I really wish OpenAI gave us the whole thought process verbatim, as I'm kind of curious where those "thoughts" come from and what happens to them.
Don't see what you mean by this; there's no such thing as "thoughts" of an LLM, and if you mean the feature marketers called chain-of-thought, it's yet another instance of LLMs making shit up, so.
Simply. Because the same is the case with humans. Mostly for the same reasons.
(Are humans overhyped? Maybe?)
The LLM hype train isn't about them being more accurate or faster than what came before - it comes from them being able to understand what you mean. It's a whole new category of software - programs that can process natural language like humans would; a powerful side effect that took the world by surprise is, that making LLMs better at working with natural language implicitly turns them into general-purpose problem solvers.
> Don't see what you mean by this; there's no such thing as "thoughts" of an LLM, and if you mean the feature marketers called chain-of-thought, it's yet another instance of LLMs making shit up, so.
"Chain-of-thought" is so 2024; current models don't need to be told to "think step by step", they're post-trained to first generate a stream of intermediary tokens not meant as "actual answer", before continuing with the "actual answer". You can call it however you like; however both research literature and vendors settled on calling it "thinking" or "reasoning". Treat them as terms of art, if that helps.
You can get the Google pro subscription (forget what they call it) that's ordinarily $20/mo for free right now (1 month free; can cancel whenever), which gives unlimited Gemini 2.5 Pro access.
You're holding it wrong. You need to utter the right series of incantations to get some semblance of truth.
What, you used the model that was SOTA one week ago? Big mistake, that explains why.
You need to use this SOTA model that came out one day ago instead. That model definitely wasn't trained to overfit the week-old benchmarks and dismiss the naysayers. Look, a pelican!
What? You haven't verified your phone number and completed a video facial scan and passed a background check? You're NGMI.
Love this reference :)
- o3 is the bestest and my go-to, but its strength comes from it combining reasoning with search - it's the one model you can count on finding things out for you instead of going off vibe and training data;
- GPT 4.5 feels the smartest, but also has tight usage limits and doesn't do search like o3 does; I use it when I need something creative done, or switch to it mid-conversation to have it reason off an already primed context;
- o4-mini / o4-mini-hard - data transformation, coding stuff that doesn't require looking things up - especially when o3 looked stuff up already, and now I just need ChatGPT to apply it into code/diagrams;
- gpt-4o - only for image generation, and begrudgingly when I run out of quota on GPT 4.5
o3 has been my default starting model for months now; most of my queries generally benefit from having a model that does autonomous reasoning+search. Agentic coding stuff, that I push to Claude Code now.
Unlike Catholic saints, ChatGPT models actually exhibit these properties in directly observable and measurable way. I wrote how I decide which model to use for actual tasks, not which saint to pray to.
> actually exhibit these properties in directly observable and measurable way
Well but do they? I don't mean your vibes, and I also don't mean cooked-up benchmarks. For example: https://metr.org/blog/2025-07-10-early-2025-ai-experienced-o...
Or perhaps you stop being obtuse. There's no causal connection between "using saints for actual tasks" and the outcomes, which is why we call this religion. In contrast, you can see the cause-and-effect relationship directly and immediately with LLMs - all it takes is going to chatgpt.com or claude.ai, typing in a query, and observing the result.
> Well but do they? I don't mean your vibes, and I also don't mean cooked-up benchmarks.
Do read the study itself, specifically the parts where the authors spell out specifically what is or isn't being measured here.
You may not like but it's what it is.
Still, it's not what you're saying. Vibes are present, too, but so are concrete effects like one-off utilities conjured in 2 minutes of prompting instead of 10 minutes of searching or 1 hour of coding, or research questions answered swiftly (and answers verified by me), that I normally wouldn't consider tackling due to up-front effort being greater than my discretionary time. And so on.
This is not vibes, but hard, empirical evidence, immediately retrievable as it just accumulates in the chat history, so I can always go back to check if I wasn't misremembering things.
There's a silver lining in this, though: none of that is any kind of deep expertise, so there's no need for up-front investment. Just start using a tool and pay attention, and you'll pick up on those things in no time.
I hope you appreciate just how crazy this sentence sounds, even in an age when this is normalised.
It's kind of weird to see people running into this kind of issue with modern large models with all the RL and getting confused. No one starting today seems to have good intuition for them. One person I knew insisted LLMs could do structural analysis for months until he saw some completely absurd output from one. This used to be super common with small GPTs from around 2022 and so everyone just intuitively knew to watch out for it.