YourNetworkIsHaunted

This is a good example of something that I feel like I need to drill at a bit more. I'm pretty sure that this isn't an unexpected behavior or an overfitting of the training data. Rather, given the niche question of "what time zone does this tiny community use?" one relatively successful article in a satirical paper should have an outsized impact on the statistical patterns surrounding those words, and since as far as the model is concerned there is no referent to check against this kind of thing should be expected to keep coming up when specific topics or phrases come up near each other in relatively novel ways. The smaller number of examples gives each one a larger impact on the overall pattern, so it should be entirely unsurprising that one satirical example "poisons" the output this cleanly.

Assuming this is the case, I wonder if it's possible to weaponize it by identifying tokens with low overall reference counts that could be expanded with minimal investment of time. Sort of like Google bombing.

It's a solid intro CS puzzle for teaching recursion. I think the original story they invented to go with it also had 64 disks in a temple in, well, Hanoi. Once the priests finished it the world was supposed to end or something.

Can confirm that about Zitron's writing. He even leaves you with a sense of righteous fury instead of smug self-satisfaction.

And I think that the whole bullshit "foom" argument is part of the problem. For the most prominent "thinkers" in related or overlapping spaces with where these LLM products are coming from the narrative was never about whether or not these models were actually capable of what they were being advertised for. Even the stochastic parrot arguments, arguably the strongest and most well-formulated anti-AI argument when the actual data was arguably still coming in, was dismissed basically out of hand. "Something something emergent something." Meanwhile they just keep throwing more money and energy into this goddamn pit and the real material harms keep stacking up.

So I don't know if it's a strong link, but I definitely learned to solve the Towers when playing through KotOR, then had it come up again in Mass Effect, and Jade Empire, both of which I played at around the same time. From a quick "am I making this up?" search, it's also used in a raid in SW:TOR, and gets referenced throughout the dragon age and mass effect franchises even if not actually deployed.

What does the “better” version of ChatGPT look like, exactly? What’s cool about ChatGPT? [...] Because the actual answer is “a ChatGPT that actually works.” [...] A better ChatGPT would quite literally be a different product.

This is the heart of recognizing so much of the bullshit in the tech field. I also want to make sure that our friends in the Ratsphere get theirs for their role in enabling everyone to pretend there's a coherent path between the current state of LLMs and that hypothetical future where they can actually do things.

That's probably true, but it also speaks to Ed Zitron's latest piece about the rise of the Business Idiot. You can explain why Wikipedia disrupted previous encyclopedia providers in very specific terms: crowdsourced production to volunteer editors cuts costs massively and allows the product to be delivered free (which also increases the pool of possible editors and improves quality), and the strict* adherence to community standards and sourcing guidelines prevents the worse loss of truth and credibility that you may expect.

But there is no such story that I can find for how Wikipedia gets disrupted by Gen AI. At worst it becomes a tool in the editor's belt, but the fundamental economics and structure just aren't impacted. But if you're a business idiot then you can't actually explain it either way and so of course it seems plausible

At least the lone crypto bro is getting appropriately roasted. They're capable of learning.

As the bioware nerd I am it makes my heart glad to see the Towers of Hanoi doing their part in this fight. And it seems like the published paper undersells how significant this problem is for the promptfondlers' preferred narratives. Given how simple it is to scale the problem complexity for these scenarios, it seems likely that there isn't a viable scaling-based solution here. No matter how big you make the context windows and how many steps the system is able to process it's going to get out scaled by simply increasing some Ns in the puzzle itself.

Diz and others with a better understanding of what's actually under the hood have frequently referenced how bad Transformer models are at recursion and this seems like a pretty straightforward way to demonstrate that and one that I would expect to be pretty consistent.

That would be the best way to actively catch the cheating happening here, given that the training datasets remain confidential. But I also don't know that it would be conclusive or convincing unless you could be certain that the problems in the private set were similar to the public set.

In any case either you're doubledipping for credit in multiple places or you absolutely should get more credit for the scoop here.

The thing that galls me here even more than other slop is that there isn't even some kind of horrible capitalist logic underneath it. Like, what value is this supposed to create? Replacing the leads written by actual editors, who work for free? You already have free labor doing a better job than this, why would you compromise the product for the opportunity to spend money on compute for these LLM not-even-actually-summaries? Pure brainrot.

New rule: you're not allowed to tell people to shut up and look at the numbers unless you're actually good at math.

As usual the tech media fails to consider the possibility that part of the reason for Anthropic poaching people with promises of more money and huffable farts is to get this exact headline to try and get another round of funding from the VCs.