It is interesting that Meta would have to develop a social or political content identification system in order to either label, limit targeting, or ban/censor it. It seems the hard part of complying was developed, but only to thwart the law, rather than comply. That might be a bigger FU to the EU than ignoring the law.

Your text may well have been in the training corpus, but not searchable with whatever terms and search engine that LLM used after you prompted it. It doesn’t have recall of sources of documents comprising its training corpus, unless the source is widely cited in other of its training documents. Sourcing and provenance aren’t currently an intentional part of LLM training.

That may already have been obvious, and central to the complaint, but I was just knee-jerking to the anthropomorphic language.

I would assume the parent understands that. The criticism is about the LLM behavior this results in. Being able to explain a behavior doesn’t necessarily excuse it. By some moral standards, you wouldn’t have trained an LLM that way, or wouldn’t have made it available, given the predictable outcome.

It also ostensibly replaced a search system that would likely have been capable of finding the original document. It's damning that the new one can't.

>Sourcing and provenance aren’t currently an intentional part of LLM training.

Congrats, brother. That's the issue TFA and GP are bringing up. And I do agree it is unfair and equivalent to (intellectual) theft.

Do you mind excerpting the valid two paragraphs?

That was a figure of speech. It boils down to the fact that, by asking for laws mandating AI research to slow down, the incumbent Big AI firms get regulatory capture and cartel power, which might be the real motivation underlying the apocalyptic language (follow the money and all that).

Well, in fairness, both sides were bad, and still are -- although, now, one side is spectacularly worse. Dems can't ever agree on a candidate or position or message, and intentionally run candidates the otherside would abhore, and Republicans would happily vote in unison for an old shoe as long as it was sufficiently dog-whistley.

One "side" is actively and intentionally destroying the country, the other is letting them. It's not that both sides are bad, it's that both sides are the same side, working for the same people.

You'll find people on both sides of the aisle who agree with this sentiment, they just disagree with what side is destroying the country and what side is letting it happen

It may be turtles all the way down, but we don't want to need to visit each tier, wading through obvious slop engineered for engagement as we go. The sloppiness reveals the low effort machinery meant to sap our vital attention and time. It's probably best to build a growing resistance to it.

… but, a power power user would know that it was transparently saved every several keystrokes, even while named Untitled and ostensibly unsaved, ready for immediate transparent recovery on relaunch, after an app or system crash -- so they might again be okay with it.

Like an idiot, I once wrote a simple, internal, in memory chat room via web page at work. Because it let you pick a nickname, co-workers immediately gave themselves names of other employees and started sexually harassing each other and being asses. I pulled it down immediately before they got me fired. Anonymity does something to some people. They lurk among us biding their time, looking for any opportunity to be reputationally unrestrained.

At that rate you could use an image model, which was designed for the task. Thats the absurdity of this test. It’s often a text only generation model that has never seen a pelican, coerced into creating an xml graphical representation that a human might recognize. That it does anything passable is already astounding.

I agree, it's hard! That's why it's still a good benchmark.

It probably has a few million inputs on how a pelican and bicycle looks, not to mention the amount of data on how to create SVG’s. Ask it to create a relaxing spa website and it will, even though it has never seen a spa.

Most of the models are multi-modal and trained on images no? That's what they claim at least.

You’re right. Modern frontier models are now multimodal. I used often as weak a hedge, because I know at least his gpt3.5 turbo and llama3.1 generated pelicans were from text only models without image training. The chinese models are interesting, because before their vision models existed they may have been distilling text only models from text output of American vision models, so they could have benefited from the teacher model’s vision capability without being vision models themselves.

I think the evil part is the [at any cost] part, the incentives are irrelevant. An AI that only wants to make paperclips isn't evil because of the paperclips.

But, they were wrong ... in assuming what they were saying was at all relevant. And they were so sure they knew what they were talking about and that it applied that they dismissed all the ample evidence to the contrary and also dismissed the, they assumed, delusional people pointing out the contradictions. Their initial misunderstanding was understandable given the overloaded terms but their confidence was poorly calibrated to their weakness detecting the misunderstanding.

I’ve seen younger engineers (still very senior and capable but with quick tempers) make this mistake time and time again.

The funny part, to me, is that it all could have been avoided by simply saying something to the effect of “the APIs they call in UTM are Apple’s and there are no hacks or workarounds they can do at the UTM level. Any differences between the hardware are just that.”

The actual thread was wildly inefficient but if it hadn’t gone that way no one would have learned anything, including us!