[hospitalJail]

> "expectation that companies like yours must make sure their products are safe before making them available to the public."

Lets make a guess, they are going to say its dangerous and we need regulation to prevent competitio---terorrism.

Here is what you need to do instead, get some smart people from various agencies that are trustworthy, have them use the openAI playground and see what can be accomplished. Then show them that you can torrent facebooks LLM right now, and that its already on computers worldwide. The cat is out of the bag.

Then let them make policy decisions.

Hard to imagine this is anything other than a ploy for regulations and lobbying.

[JumpCrisscross]

> get some smart people from various agencies that are trustworthy, have them use the openAI playground and see what can be accomplished

This is a punt to committee. Likely what this meeting will result in. It’s as performative as it’s useless.

Suggestions of pauses have always been a farce. But I’m struggling to see solutions from experts, apart from constant predictions of generic doom. (I’m in favour of a domestic registry, so we know who’s training what on which data. Maybe a copyright safe harbour in exchange for registration?)

The other side is competitiveness: what can the federal government do to make America the best place to build AI? (I'm continually drawn to the Heavy Press Program [1], the era's massive forging presses being loosely analogous to modern training costs.)

[1] https://en.wikipedia.org/wiki/Heavy_Press_Program

[kalkin]

LLama is meaningfully behind the state of the art AFAIK, so the cat isn't fully out of the bag in that sense. Whether a GPT4 or better model has public weights in 2-3 years may in fact depend on government regulation.

[AmericanChopper]

1A makes regulating software basically impossible. I can’t imagine what additionally regulation they think they’re going to implement that will survive the judiciary. The only legal barrier I can see that could influence these AI services is that that their output is obviously not shielded from liability by Section 230. But that will play out in the civil courts, if at all.

I’d say the more likely outcome is the far more subversive scenario where the government simply pressures private organisations into doing its unconstitutional bidding

[snapcaster]

Sorry but this comment is just totally wrong. Look into weapons exports that applied to cryptography, google providing android to Huawei, etc. etc. etc. etc. i mean jeez you don't have to look far for counterpoints to what you're saying

[generalizations]

Yeah, they tried to limit crypto exports. Remember that munitions t-shirt? https://en.wikipedia.org/wiki/Crypto_Wars#/media/File:Muniti... You can put any regulations you want on the books - the point is that enforcing them can become laughably infeasible.

[AmericanChopper]

How can people (especially on this forum) know that crypto export controls existed, but not know that they were struck down for violating 1A?…

https://en.m.wikipedia.org/wiki/Bernstein_v._United_States

[snapcaster]

I guess we maybe differ here on what would constitute success for suppression. I see it as a "success" because the law stood for so long despite being pretty apparently unconstitutional. It is fair to see it as a failure since it eventually was overturned. Maybe the crux here is I probably agree with you that permanent suppression is impossible but temporary is super doable and temporary can be a fairly large fraction of a human lifetime so I still count it

[AmericanChopper]

I guess that’s an interesting way to backpedal without acknowledging how outright wrong your highly dogmatic comment was. There was a time in the past when there could be some merit to where you’ve chosen to move the goalposts. But there is now, and has been for some time, a higher court precedent that invalidates this position.

[snapcaster]

What do you mean? they were banned, the company i was working for almost went under because we were banned from providing any technology to a chinese tech company. That did get reversed but only after almost destroying the company. The android OS was caught up in that as well. How are these not suppressions?

[MilStdJunkie]

I'm pretty sure that's not accurate, at least for cryptography, GPS, and some industrial control software (a few special LabView plugins, from my own personal experience). I mean, the actual enforcement, I agree with you, it's theatre. But if you get caught getting on a plane with it without your paperwork. . hoo boy. That ain't good. Bring extra kneepads and learn to relax your throat muscles.

[pc86]

Software is not speech.

[AmericanChopper]

The judiciary disagrees.

> One of EFF's first major legal victories was Bernstein v. Department of Justice, a landmark case that resulted in establishing code as speech

https://www.eff.org/deeplinks/2015/04/remembering-case-estab...

[aftbit]

I disagree. See the history of DeCSS for more information.

[ajmurmann]

It also depends on what "safe" means. Before I read your comment I assumed this was about not accidentally building skynet, rather than making it easy to break copyright etc. I hope it's about AI Safety given Anthropic is there and not Midjourney.

[pers0n]

Fundraising time for the 2024 elections

[Mistletoe]

Why are regulations bad? When there are no regulations the results are usually not pretty. Current big tech dystopia, stock market crash of 1929, UK textile mills worked by children in the Industrial Revolution...

Why should we expect AI to not repeat the same abuses and errors?

[blihp]

Knee-jerk regulations rooted in emotion and ignorance are bad. When there are bad regulations the results are usually not pretty (i.e. they tend to fail to achieve their stated objectives) and all they end up doing is protect the established players from competition as only they can afford the cost of compliance.

[visarga]

> Why are regulations bad?

Depends, do you mind not being allowed to have AI without surveillance?

[pc86]

Regulations are good when they're made by smart people who are experts in what they're regulating, with an eye toward creating an environment of fair competition, and perhaps protecting US interests for things like export controls, etc.

Regulations are bad when they are made from a position of emotion, partisan hackery, protectionism of either incumbents as we are likely to see here or specific industries/sectors, or written by out of touch, mentally declining octogenarians (I don't mean Biden, I mean half of Congress, of both parties).

Are there any executive agencies or legislative committees qualified to regulate development of AI? I just mean on basis of knowledge and education, not authority.

[rqtwteye]

"Are there any executive agencies or legislative committees qualified to regulate development of AI? I just mean on basis of knowledge and education, not authority."

Agencies and committees are full of very smart and conscientious people who are able to understand an issue and create reasonable regulations. Problem is that political leadership usually messes up things by going with ideology or lobbyist interest.

[jonhohle]

Unintended consequences. Power surrendered by the people is not easily recovered.

[surgical_fire]

This presumes that without regulations the people have the power.

In truth, without regulation, the power is at the hands of the ultra wealthy corporations and their overlords. Petty tyrants in the making, all of them.

[Kbelicius]

How is someone on HN against innovation.... blows my mind.

There are unintended consequences with innovation too.

[dimitrios1]

Regulations carry a cost of compliance. When regulations are being clamored for by the big players, I can pretty much guarantee you it is not out of altruism.

[gordian-mind]

Companies are not the only ones who have something to gain from that exchange. Democratization of AI will also be used to push harsher legislation on the internet, in the guise of fighting "misinformation" (now possibly AI-generated!).

[jjoonathan]

Genuine question: how should we fight LLM astroturfing?

[JieJie]

In my humble opinion, LLM astroturfing is only a productivity increase over what is currently already a very automated process (social media bot farms, etc).

While this seems like it would exacerbate an already existing problem, it may not have so profound an effect. You see, while LLMs may be able to increase the amount and quality of fake information, fake information isn't currently in short supply, so increasing the amount and quality of it may not have that strong an effect.

In short, we already have 24-hour fake news cable channels and infinite doom scrolls. The bottleneck is there, not in the quality or quantity of fake news.

Now, if they invented a LLM that doomscrolled Twitter and voted based on generated summaries, we would have much greater grounds for concern.

[edit: I hope this doesn't sound too snarky. What I mean to say is that we should fight it by creating less gullible consumers of information, a project in which AI may be uniquely qualified to assist us.]

[jjoonathan]

Human attention doesn't scale. LLMs do. "Git gud" is a losing strategy in aggregate.

[JieJie]

We're talking past each other.

The whole point of what I was saying is that yes, human attention doesn't scale and that is what is going to save us from a deluge of LLM spam.

A billion pieces of fake news or ten, it makes no difference. Humans can only look at one at a time.

[MichaelZuo]

Eventually they'll spend 99% of their time on LLM generated stuff, who cares if it's only 1 at a time?

[DethNinja]

How do you propose we fight government/corporate astroturfing?

At least LLMs will democratize the astroturfing.

[hospitalJail]

From what I've seen on reddit/here, companies buy old accounts. This costs too much for the average person.

[SuoDuanDao]

LLMs can usually tell when they wrote something, so we can at least recognize it. But really, LLMs could be used by genuine grassroots campaigns as well - possibly make grassroots campaigns easier, since the skills to write persuasively aren't always available to the groups that most need them.

I suspect the observation that 95% of anything is crap will hold true, and simply have to filter out more crap now that it's easier to produce it. There'll also be more gems, so it's hardly all bad.

[visarga]

Use LLM anti-astroturfing to fix it. Fight fire with fire.

[permo-w]

stop using the internet

[jonhohle]

With the exception (currently) of in person conversation, AI can be used basically every other form of communication (phone calls, written communication, TV, radio, etc). It’s not just an Internet issue.

[permo-w]

I was mostly joking, but even so most of this is a stretch as of now

I could be wrong, but I’m fairly sure we’re not yet at the point of convincing, real-time voice gen, nor any kind of decent quality TV (sadly), although printed text and (non-live) radio are certainly viable right now

[trafficante]

Real-time convincing-sounding TTS is extremely close - I’d be surprised if we don’t get an elevenlabs equivalent to gpt-turbo before the end of summer.

I spent an hour or three over the weekend messing around in Skyrim VR with a mod that does player speech recognition, pipes it out to GPT with identifier tags to give scene context, sends GPT output to elevenlabs (optional), and then the mod integrates it into Skyrim mouth rigging etc.

Yes, there’s an extremely obvious lag before you get a response, but it’s on the order of seconds even though this is an early Skyrim mod with a ton of moving parts interacting.

And the result is…astounding. As someone said in a comment on the below linked video: this is the biggest leap for video games since Half Life 1.

https://youtu.be/d6sVWEu9HWU