Comment by tyre | Hacker News Reader

Comment by tyre | Hacker News Reader

I'm empathetic to the desire for innovative freedom and democratization of technology, but the benefit has to be balanced with the scale of impact and the possibility of mitigating/reparative response.

The capabilities of current generation AI—LLMs + audio/image/video generative models—are far enough already that wide-scale distribution is extremely dangerous.

Social media allowed for broadcasting from 1:many and 1:1 with troll armies, but LLMs are on a whole other planet of misinformation. A scalable natural language interface can target people at the one-to-one conversation level and generate misinformation ad hoc across text, voice, images, and video.

The trouble with this approach is that harm will rack up way faster than liability can catch up, with irreversible and sometimes immeasurable long term effects. We have yet to contend with the downstream effects of social media and engagement hacking, for example, and suing Facebook out of existence wouldn't put a dent in it.

Current generation AI has enough capabilities to swing the 2024 US presidential election. It can be used today for ransom pleas trained on your childrens' instagram posts. It doesn't seem farfetched that AI could start a war. Not because it's SkyNet, but because we've put together the tools to influence already-powerful-enough human forces into civilization-level catastrophe.

Seems like a ridiculous standard to use US elections. Arguably Facebook showed the ability to do the same in 2016, so it’s also uninteresting.

The only reason an LLM would even have such access to people to do the hyperbolic damage you claim is thru ad networks and social media anyways - yet you are calling to behead public access to tech instead of regulating propaganda platforms. Both seem facile but your selection is the one that seems subservient

Hmm.

I wasn't convinced, but I did the maths and, well, looks like you're right:

Hiring people in to read and respond in a politically biased way to each person in the USA would take, let's say 2 messages per day for a month, so 60 messages per person, let's say 100 million actual voters you care about so 6 billion messages, let's say 15 seconds per response so 25 million hours, let's say you're hiring Kenyans because I happen to have the GDP/capita and workforce participation rate to hand, so it's about 2 USD/hour on average, so 50 million USD.

With GPT-4, and let's approximate each message as 100 input tokens and 100 output tokens, that's 0.1 * ($0.03 + $0.06) = $0.009, times the same 6e9 messages = 54 million USD.

So, yeah, looks like you're right.

3.5 is cheaper, but tends to be noticeable; from what I've been hearing (fast moving target and all that) the OSS LLMs tend to be 90% of the quality of 3.5.

Eh, only partially correct...

The battle for the future isn't attention, it's intimacy. LLMs will up the game from broadcast propaganda to individual relationships via digital medium. It will be a two way communication that profiles you and feeds your biases in a much more direct way than even the social media giants themselves have accomplished.

I'd say GPT-4 can make much more elaborate arguments in 15s than humans can.

Speed yes, but cost? Can it make better comments in 100 tokens input and output, than humans in 15 seconds?

(I still haven't used 4 properly, so possibly? And given the hypothetical situation under discussion, could it have if it had not been directed not to?)

> Arguably Facebook showed the ability to do the same in 2016, so it’s also uninteresting.

You're conflating a communications channel with a new technology that enables abuse (as well as positive outcomes) at a previously-unimagined scale.

Is it not facile and subservient to default to the ostrich algorithm for new and world-changing technologies?

> You're conflating a communications channel with a new technology

I'm really not. Do you genuinely think that the last half of your sentence only applies to LLMs? A "technology that enables abuse (as well as positive outcomes) at a previously-unimagined scale" is literally a perfect description of social media platforms. LLMs virtually don't move the needle, look at the amount of propaganda and bots already on the web today.

I think that implying I am burying my head in the sand about LLMs by claiming they shouldn't be restricted to megacorps who already have demonstrated their willingness to engage in unethical activities at mass scale is fucking ridiculous too. But I do welcome your continued support of regulating AI in favor of corporations, it's a clear sign to never engage with you

> LLMs virtually don't move the needle,

I'd say you'll be incorrect in the future. We'll have new abuses by both corporations and propaganda bots in the near future.

More and more of us will find 'digital friends' that creep into our lives and talk to us on line, gain our trust, make us laugh, and tell us that 'unions are bad', or whatever the propaganda of the day is. But it's the one on one relationships with history that will make them different than the bots that came before.

>More and more of us will find 'digital friends' that creep into our lives and talk to us on line, gain our trust, make us laugh, and tell us that 'unions are bad', or whatever the propaganda of the day is.

That would just mean automating the job of the secret police. And those guys have unions already!

> it's the one on one relationships with history that will make them different than the bots that came before

Is this intentional ignorance? The reason I use Facebook as the analogy is because Facebook has an insane amount of your history, and isn't afraid to sell any single aspect of that to advertisers. The interface some chatbot-controlling megacorp would expose to propagandists would hardly be different, the history is unlikely to be much stronger, and the network effect seems hard to match. Less a step-change than an iteration.

Reply on news.ycombinator.com