News for hackers. Stuff that matters.
Re:Run 100B+ language models at home, BitTorrent‑style by planb on Tuesday March 21, 2023 @02:56AM (#35243233)
I don't understand this argument. Have you tried running a website with an open comment section in the last 10 years? Every corner of the internet is already stuffed with low-qualtity spam. Does it really matter if the spam quality gets better? Search for any combination of 2 words that are not related to each other on Google and you find some bullshit site that just lists random words. Arguably, wouldn't it be better if there actually was AI generated content that combines the 2 words in some meaningful way and maybe, maybe, presents something useful? It's also not like every information on the internet - even if generated by humans - is correct and fact checked, so you need to do the critical thinking yourself anyway.
Re:Run 100B+ language models at home, BitTorrent‑style by visarga on Tuesday March 21, 2023 @03:00AM (#35243257)
Yes, this is a growing stage. In one or two years LLMs will have Wikipedia quality or even research paper quality. The spam they will produce might be better than most human written stuff.
Re:Run 100B+ language models at home, BitTorrent‑style by ithkuil on Tuesday March 21, 2023 @03:31AM (#35243417)
At which point does high quality spam cease to be spam?
Re:Run 100B+ language models at home, BitTorrent‑style by amelius on Tuesday March 21, 2023 @04:50AM (#35243883)
The point where it is just misinformation?
Re:Run 100B+ language models at home, BitTorrent‑style by Yajirobe on Tuesday March 21, 2023 @05:42AM (#35244198)
Misinformation is false information. Spam can be facts.
Re:Run 100B+ language models at home, BitTorrent‑style by amelius on Tuesday March 21, 2023 @06:30AM (#35244462)
Theoretically, yes. But better treat it as misinformation.
Re:Run 100B+ language models at home, BitTorrent‑style by zamnos on Tuesday March 21, 2023 @04:57AM (#35243929)
Might refer you to XKCD 810.

https://xkcd.com/810/

Re:Run 100B+ language models at home, BitTorrent‑style by ithkuil on Tuesday March 21, 2023 @09:52AM (#35246088)
There is a XKCD for everything.

hmm, is there XKCD for "might refer you to XKCD $number" ?

Re:Run 100B+ language models at home, BitTorrent‑style by bugcheck7b on Tuesday March 21, 2023 @03:31AM (#35243416)
If the spam is better quality than the human written stuff, who's to say we aren't better off?
Re:Run 100B+ language models at home, BitTorrent‑style by lanternfish on Tuesday March 21, 2023 @03:40AM (#35243457)
Quality in this case doesn't necessarily mean ground truth accuracy - it just means ability to look accurate to humans.
Re:Run 100B+ language models at home, BitTorrent‑style by fauigerzigerk on Tuesday March 21, 2023 @05:04AM (#35243975)
I agree, that's the problem, but I think it's still somewhat complicated.

Imagine someone posting an extremely well written and insightful postmortem of an outage. It would show advanced and accurate usage of all kinds of tools to get to the bottom of the outage. It would be extremely useful reading for anyone investigating a similar outage, but the outage never actually occurred.

Now you have both ground truth accuracy and misleading fiction at the same time. Whether or not that makes the post useful depends entirely on the conclusions you're drawing from it.

Re:Run 100B+ language models at home, BitTorrent‑style by jonathanstrange on Tuesday March 21, 2023 @05:52AM (#35244243)
> Does it really matter if the spam quality gets better?

It matters a lot. Spam is easy to recognize and e.g. my current email client filters out dozens to hundreds of spam mails per day without any false positives. If you cannot distinguish spam from normal posts, this could even cause democracy to break. Unfortunately, there are strong anti-democratic forces in the world who want this to happen. In my humble opinion, this is the biggest threat to humanity right now because (unlike other threats) it's not hypothetical, it's going to happen.

Re:Run 100B+ language models at home, BitTorrent‑style by emporas on Tuesday March 21, 2023 @09:22AM (#35245744)
>If you cannot distinguish spam from normal posts, this could even cause democracy to break.

You can distinguish however online accounts from real people and bots. That's easy and so cheap, i consider it it, essentially free. Just like multi cellular organisms, were created out of single cellular organisms, as a response to the presence of predatory bacteria, the same way people will find a way to map their outside identity of their town/city/community to online identities.

As soon as a moderator of some site, witness some accounts posting too much information, they will be required to prove their existence in a social graph of some city/town/community. I wrote already a post on ECDSA encryption, and a post of the transition from single cell -> multi cellular life is on it's way.

Re:Run 100B+ language models at home, BitTorrent‑style by unity1001 on Tuesday March 21, 2023 @04:09PM (#35251827)
> democracy to break

As if there is any democracy in the countries that claim to have democracy. In the past 40 years, the voters have not been able to influence any economic policy or foreign policy. 74% Americans said to Gallup that they thought their votes absolutely did not change anything and they did not matter even as early as the second Bush administration...

Re:Run 100B+ language models at home, BitTorrent‑style by oefrha on Tuesday March 21, 2023 @06:20AM (#35244409)
Aside from a few skids spamming for fun, the dominant forms of online spam by far are (1) content mills farming impressions for AdSense $$$; (2) user-generated content on third party platforms pushing something for economic or, to a lesser extent, political gain, whether it's SEO backlinks, fake product reviews, crypto shilling, or whatever.

(1) getting better is bad because you can enter the two words into Bing Chat or whatever to generate the same shit yourself, so you won't need them anyway, they only get in the way when you want to look for actual human-generated/curated content.

(2) getting better is obviously bad. Imagine most user-generated content turning into Quora-style ads or Amazon fake reviews, except with eloquence and bullshit knobs turned to 120%. Everything you read is coherent, convincing prose, you just don't know whether they're 100% false.

Re:Run 100B+ language models at home, BitTorrent‑style by mgoetzke on Tuesday March 21, 2023 @06:51AM (#35244574)
Without user verification it will now be quite easy to e.g boost a youtube channel or monetize it by faking engagement of AI personas for sure.
Re:Run 100B+ language models at home, BitTorrent‑style by DeathArrow on Tuesday March 21, 2023 @07:05AM (#35244669)
This is worse because you can filter out low quality spam much faster than LLM generated spam.