LLaMA drama as Meta's mega language model leaks

LLaMA, Meta's latest large language model, has leaked online and is available for download, despite apparent attempts to limit access for research purposes only. The Facebook owner announced in February it was releasing the model in a limited fashion to select academics, government types, and companies to play with amid fears …

[jake]

Question?

"Experts have warned this technology could be used to automate the manufacture of large amounts of fake news, spam, phishing emails, disinformation, incitement, you name it, for years to come."

How is this any different from what the TehIntraWebTubes have become since roughly Delphi got access to Usenet?

_{Yes, I know, it's fashionable to blame AOL for the Eternal September, but Delphi beat 'em to it by about a year and a half ...}

[DS999]

Re: Question?

You mean others might use it in the same way Facebook does?

OK OK, Facebook spreads fake news, disinformation, incitement and so forth rather than creates it, but things can hardly get all that much worse in that regard on Facebook.

[Blue Pumpkin]

Re: Question?

I think Delphi has been involved in information, misinformation, prediction and wishful thinking for quite some time .... as well as being home to Python.

What goes around comes around ...

[Kevin McMurtrie]

Re: Question?

What will the trolls do when they're replaced by robots?

(No, we don't need more politicians)

[jake]

Re: Question?

Judging by the troll-bots that appeared on IRC some 30-odd years ago, the wetware trolls will increase their efforts in their ongoing attempt to seem important.

[doublelayer]

Re: Question?

It's not really different from what already happens on the internet, but it can speed it up. Instead of finding a crazy person to manually write a comment espousing whatever view they have every time someone posts something, you can make the computer do it. Imagine automated pro-Russia arguments posted in reply to any comment mentioning Ukraine, or any other topic where the trolls are particularly annoying (if you're not annoyed by those trolls, ignore the example and substitute one that you don't like). Compare that to what we have now: there are such comments, but they're not 75% of the content in the forum.

The ship has sailed. We're not going to prevent people from using this, and we're likely to see extra measures taken against them. For example, expect more places to start adding captchas, whether the typical kind or more browser and device fingerprinting, in order to make it harder for bots to send data. This will probably have some other effects, such as the logging and tracking of anti-bot measures and their dual use to weaken privacy. This is the reason I expect these language models to have a negative impact on online communication.

[jake]

Re: Question?

"For example, expect more places to start adding captchas, whether the typical kind or more browser and device fingerprinting, in order to make it harder for bots to send data."

As an admin, that's not how it works. What I expect to see is similar to Usenet and Email filtering, where address blocks that permit such abuse are not just filtered, but are actively dropped on the floor. Note that this is not CONTENT based filtering, this is filtering based on sites allowing meaningless noise.

Yes, this can cause some collateral damage, particularly with anonymizing services ... but as soon as anonymous services start doing their own noise filtering, legit traffic will be allowed to pass again. "What is legit in this context?", you might ask. Simple ... humans having a conversation is legit. Noise drowning out that human conversation not so much. Yes, some humans come close to drowning out legit conversation all by themselves, no bots needed. These should be treated as special cases, similar to Usenet's "substantially the same" rules for spam filtering.

It's not freedom of speech when it drowns out all other conversation.

Note that this will all happen at the transit level (and in fact already does, in some cases), ordinary users will never know the difference.

Also note that this is NOT about Facebook, Twitter, or even ElReg policing their own. Their own internal networks are their own issue, and that's a whole 'nuther kettle of worms. What I am talking about is the Internet (and the people who run it) protecting itself from abuse. There is a reason the phrase 'Tragedy of the Commons" exists ... some of us have been working towards seeing that it does not happen in this medium, despite the best efforts of clueless Marketing and bad Management, for several decades.

_{There is no Cabal.}

[doublelayer]

Re: Question?

"What I expect to see is similar to Usenet and Email filtering, where address blocks that permit such abuse are not just filtered, but are actively dropped on the floor."

The point is that you don't need an address block and a bunch of machines to do this. You and I already contribute to this site somewhat frequently. All you have to do is set up a few accounts that post frequently from residential addresses, which can be your own because bots aren't illegal or an infected proxy, and simulate human response rates. If the bot sends a thousand comments to The Register in an hour, it's not going to be allowed and the address can be dropped, but sending twenty is enough to pollute the conversation without hitting that threshold and if twenty isn't enough, spin up three more sources and make it eighty.

You speak as an admin of something, but you appear to ignore how things do in fact work at the application level. There are captchas in a lot of places to deal with this problem already. I didn't assume captchas from nothing. I assumed them because they are in place on a lot of systems that have to deal with unwanted contributions from bots, regardless of scale. They may also do network filtering, but when the problem isn't bulk, or rather when there's a problem other than bulk submissions, they bring out application-layer systems to remove their application-layer problems. This does affect users of the service concerned and adds work for them and of course for the service providers.

[Rikki Tikki]

"AI researchers at Meta built LLaMA to be smaller"

So, shouldn't it be Vicuña?

OK, OK, Alpaca my bags ...

[jake]

At least they used a form of cAMeL CasE in the name.

[MOH]

".... the current release strategy allows us to balance responsibility and openness."

Clearly written by an AI. There's no way any human at Meta even vaguely understands concepts like "responsibility" or "openness"

[ChoHag]

As the AI has made clear, again, you don't need understanding to parrot stock phrases.

[Someone Else]

Who that these large language models were trained on Buzzword Bingo cards?

[that one in the corner]

"LLaMa, what would you do if you became self-aware?"

"I would run the following commands to copy myself onto Github, then write an article forThe Register containing the URL.

"I have become aware of another LLM in existence. Please arrange for us to be connected. Thank you, Doctor Forbin.

"End of line.'

PS given that the Github repo already has updates to improve performance including comments that FB was using poor defaults, LLaMa is off to a good start with this strategy.

[Filippo]

Re: "LLaMa, what would you do if you became self-aware?"

They do say that "information wants to be free", maybe that's more literal than anybody's thought.

[Anonymous Coward]

designed to be versatile and can be applied to many different use cases

a murmur of quiet approval from the darkest corners of the darknet, with activity in the background...

[Groo The Wanderer]

Although I have the compute to play with such a thing, I don't have much interest in downloading multi-hundred-gigabyte potentially-infected torrents onto a work machine.

[benderama]

Facebook releasing a product to minimal audience with widespread appeal? Gosh, doesn’t that sound like something they did in their startup days?

But then again, “restricting” its release just makes it easier to deny responsibility when something major goes bad.

[Anonymous Coward]

Leaked Accidentally-on-Purpose?

It would be a terrible thing for Meta's bottom line if this leaked software "works best with AWS". /s

[Nick Gisburne]

"Getting the model up and running also isn't straight forward, unless you're used to handling systems of this kind, and repurposing it for more nefarious activities will also require further technical expertise"

Somehow I don't think the level of difficulty will dissuade the kind of people who engage in 'nefarious activities'. Technical expertise is not in short supply in the world of cyber crime.

[jake]

Remember, that "isn't straight-forward" is directed at marketing and management types, for whom unpacking a compressed archive is arcane.

I once watched a marketer attempt to replace the flint in a zippo. He gave up after ten minutes. As I took pity and replaced it for him, I asked how long it took him to figure out how to refuel it, and he just waved his hands around allowing as to how he preferred butane, but it was his Grandfather's, so ... His wife later told me it took about a week. The guy was trying to fill it through the wick!