Nice balanced perspective there at the end:

"as such [LLM-coded interactive] supplements are not mission-critical to the core of the paper, I again feel that the downside risk of using guided interaction with LLM agents to generate such visualizations is acceptable."

It's a tool. Good for some things but not others and generally not to be trusted.

▲

dahart 2 hours ago | parent | next [-]

> It’s a tool. Good for some things but not for others and generally not to be trusted.

I agree completely you always need to check the work of LLM agents, but it does strike me as a tiny bit funny to anthropomorphize AI by using ‘trust’ while warning against anthropomorphizing the AI by using unchecked output. ;) Generally speaking, “trust” in AI has been going up very quickly as the models & harnesses improve, and as people figure out effective workflows.

I trust my hammer with nails but not screws… does that mean the hammer should generally not be trusted? The problem with AI is we don’t know the difference between nails and screws. (This may be where my analogy breaks down. :P) But I feel like saying don’t trust it isn’t as helpful as saying something like you should expect to spend more time planning and iterating than before, and you should expect tot spend more time reviewing and checking output than before, and learn how to use skills and context and subagents, and learn to use AI on some non-production low-consequence projects first. Saying ‘generally not to be trusted’ implicitly suggests not using AI, and doesn’t leave the reader with how to use AI. The goal is to build trust by building good workflows and by understanding what works well and what doesn’t, right?

	▲	lukan an hour ago \| parent [-]
		"I trust my hammer with nails but not screws… does that mean the hammer should generally not be trusted?" I trust a hammer to be able to hit a nail, without breaking. But if the hammer is old and the wood brittle, I don't trust it anymore. Using it for anything else (screws) has nothing to do with trust, but using the wrong tool.

▲

falcor84 2 hours ago | parent | prev | next [-]

I don't understand what trust means in this context. Even if I were able to hire Donald Knuth to write all my code, I wouldn't "trust" it to be bug-free, let alone to be the right fit for my needs.

	▲	dhosek an hour ago \| parent [-]
		You could trust it to be probably correct but he wouldn’t have tried compiling it.

▲

an0malous 4 hours ago | parent | prev | next [-]

> and generally not to be trusted

There are many AI bulls who adamantly disagree and cite Tao’s statements about LLMs for mathematical proofs as an example of how advanced and autonomous these systems already are

▲

mxkopy 3 hours ago | parent | next [-]

I mean just from the above quote it’s clear he doesn’t trust them for “mission-critical” tasks. And I doubt LLms have evolved significantly from their stochastic parrot nature over the last few years

▲

irishcoffee 3 hours ago | parent | prev [-]

Statistical gradient descent token vomiter. We can all say it together. Nothing about this is advanced or autonomous.

▲

CuriouslyC 2 hours ago | parent | next [-]

This is like saying humans are a self contained electron transport system, nothing special or advanced about that, just a scaled up nematode.

▲

abecedarius 2 hours ago | parent | prev | next [-]

The same AIs are doing math research now, you know. At what point do you stop explaining it all away?

	▲	satvikpendem 44 minutes ago \| parent [-]
		They never will, because it seems to be a psychological effect among humans via the AI effect (see my sibling comment).

▲

satvikpendem an hour ago | parent | prev [-]

[dead]

▲

andrepd an hour ago | parent | prev [-]

Indeed. LLMs produce truly atrocious code, unmaintainable and unreliable. If you're vibecoding a toy to amuse yourself or something similar low-stakes, that's perfectly fine! For higher-stakes code, it's definitely not.