site banner

Small-Scale Question Sunday for October 12, 2025

Do you have a dumb question that you're kind of embarrassed to ask in the main thread? Is there something you're just not sure about?

This is your opportunity to ask questions. No question too simple or too silly.

Culture war topics are accepted, and proposals for a better intro post are appreciated.

1
Jump in the discussion.

No email address required.

No, it isn't. I'm not talking about remembering a bunch of explicit instructions or preferences. I'm talking about learning in the way a competent person goes from a newbie to a domain expert. That is completely missing in LLMs. No matter how much I guide an LLM, that doesn't help it generalize that guidance because LLMs are static snapshots.

If you want truly online learning, you're in for an indefinite wait. Fortunately, most people get a great deal of mundane utility out of even static LLMs, and I'm not sure what you need that precludes this.

And if your answer is "but GPT-6 will totally have been trained better", then why on earth would I waste any time whatsoever with GPT-5?.

Because... it's the model we have? Can't have tomorrow's pie today, even if we're confident it's going to be tastier. Why buy an RTX 5090 when Nvidia will inevitably launch a better model after a few years? Why buy a car in the dealership today when you can wait for teleportation with complimentary blowjobs?

If you're trying to force everyone to use the solution you like, you better be damn sure your solution actually works for them instead of constantly resorting to "no, you're just using it wrong".

Hold your horses buddy. When have I forced anyone to do anything? @SubstantialFrivolity has clearly articulated his concerns about the weaknesses of LLMs as of Today AD. I invite you to tell me which of his concerns online learning is strictly needed to address? As far as I can tell, I have emphasized that his boss has a point, or is directionally correct, and that he could benefit from using LLMs more. I hope you've noticed multiple caveats and warnings attached.

If you are so convinced that even the best LLMs today are a waste of your precious time, then good luck with whatever you're using as an alternative. It's not like they're so entrenched that you can't lead a productive human life without one. They also happen to be very helpful for most people.

Patiently waiting for Scott's next prediction project, "teleportation with complimentary blowjobs 2027"

Pretty excited, should we start a Metaculus prediction market?

If you want truly online learning, you're in for an indefinite wait.

This is why I keep blackpilling on AGI. I have zero expectation of AGI without a system that can learn on its own.