Comment by otterley

> GPT and Gemini love to interrupt my novels and tell me certain behavior is illegal or immoral, and censor various anatomical words

Lol, are you using ai to create fan translations of エロ漫画 ?

It's also one of the few 'recommended' models in Kagi Assistant (multi-model ChatGPT basically, available on paid plans).

Really? They must've switched recently cause that was around before kimi came out

How much engineering do prompt engineers do? Is it engineering when you add "photorealistic. correct number of fingers and teeth. High quality." to the end of a prompt?

we should call them "prompt witch doctors" or maybe "prompt alchemists".

I think the selection of models is a bit off. Haiku instead of Sonnet for example. Kimi K2's capabilities are closer to Sonnet than to Haiku. GPT-5 might be in the non-reasoning mode, which routes to a smaller model.

Goes to show the "frontier" is not really one frontier. It's a social/mathematical construct that's useful for a broad comparison, but if you have a niche task, there's no substitute for trying the different models.

Just use something like DSPy/Ax and optimize your module for any given LLM (based on sample data and metrics) and you’re mostly good. No need to manually wordsmith prompts.

It's not fair to use prompts tailored to a particular model when doing comparisons like this - one shot results that generalize across a domain demonstrate solid knowledge of the domain. You can use prompting and context hacking to get any particular model to behave pseudo-competently in almost any domain, even the tiny <1B models, for some set of questions. You could include an entire framework and model for rendering clocks and times that allowed all 9 models to perform fairly well.

This experiment, however, clearly states the goal with this prompt: `Create HTML/CSS of an analog clock showing ${time}. Include numbers (or numerals) if you wish, and have a CSS animated second hand. Make it responsive and use a white background. Return ONLY the HTML/CSS code with no markdown formatting.`

An LLM should be able to interpret that, and should be able to perform a wide range of tasks in that same style - countdown timers, clocks, calendars, floating quote bubble cycling through list of 100 pithy quotations, etc. Individual, clearly defined elements should have complex representations in latent space that correspond to the human understanding of those elements. Tasks and operations and goals should likewise align with our understanding. Qwen 2.5 and some others clearly aren't modeling clocks very well, or maybe the html/css rendering latents are broken. If you pick a semantic axis(like analog clocks), you can run a suite of tests to demonstrate their understanding by using limited one-shot interactions.

Reasoning models can adapt on the fly, and are capable of cheating - one shots might have crappy representations for some contexts, but after a lot of repetition and refinement, as long as there's a stable, well represented proxy for quality somewhere in the semantics it understands, it can deconstruct a task to fundamentals and eventually reach high quality output.

These type of tests also allow us to identify mode collapses - you can use complex sophisticated prompting to get most image models to produce accurate analog clocks displaying any time, but in the simple one shot tests, the models tend to only be able to produce the time 10:10, and you'll get wild artifacts and distortions if you try to force any other configuration of hands.

Image models are so bad at hands that they couldn't even get clock hands right, until recently anyway. Nano banana and some other models are much better at avoiding mode collapses, and can traverse complex and sophisticated compositions smoothly. You want that same sort of semantic generalization in text generating models, so hopefully some of the techniques cross over to other modalities.

I keep hoping they'll be able to use SAE or some form of analysis on static weight distributions in order to uncover some sort of structural feature of mode collapse, with a taxonomy of different failure modes and causes, like limited data, or corrupt/poisoned data, and so on. Seems like if you had that, you could deliberately iterate on, correct issues, or generate supporting training material to offset big distortions in a model.