Post

Wow perplexity fails with scientifically researched health information. ;)

I think the most basic fact to understand about LLMs is the one response they can’t give is “I don’t know”.

Clearly, you should carry your car when you walk to the car wash.

We can't feed that data into an LLM, not because we can't collate it (although we have little idea what all of it is) but because that's not how LLMs work. They are the wrong solution to the problem - a dead end.

Thanks. This is a most excellent thread. My own little associated observation is that AI can never doubt, and without doubt neither intelligence nor understanding are possible. The people whose ideas I trust least are the ones who never express doubt.

"ask a stupid question, get a stupid answer" doesn't mean that you're being deliberately judged on the intelligence of your question. it's just a description of cause and effect.

A few days after Blue Prince came out, I fed one of them some randomly chosen examples of the harder three-box parlour puzzles (blue-prince.fandom.com/wiki/Parlor_...). I got a pretty high rate of getting the right answer. Any thoughts on what was going on there? I don't think it was regurgitation.

It clearly can't infer from a simple one sentence question that I might need my car to wash my car, because it isn't intelligent, doesn't possess logic, and can't reason. When you ask it anything complex, it's coming across 10s, 100s, or 1000s of points where it has to infer context. 5/?

There's also the problem that if someone does ask a stupid question like that in the training data, it's in the context of a joke, where the answer is deliberately wrong but the LLM cannot distinguish that.

"Since you are getting your car washed, driving is the practical choice so you can actually bring the vehicle to the service area. Additionally, the nearest car wash is about 2.7 miles away, which would be a fairly long walk." Seems like you're wrong

Many people too young to know about early days of Alta Vista, Hotbot, Lycos where the operators in your query were make-or-break for useful results. Prompt construction is very similar in this regard...except 5x the work.

The “AI” response had gone straight the most frequent match in its corpus — people trying to reduce the size of video files — and had entirely ignored the relevant info about the source being a slide deck. Not surprising; as you say, this is what they do. A bit disappointed in my colleague, though.

bsky.app/profile/arav...

Your original post describes GPT5. It actually does try and figure out how small of a model it can get away with (how complex is the query?) and then routes you to the cheapest one. It wasn't designed to give a stupid response, but a lower effort one.