Comment by CyrsBel | Hacker News Reader

Comment by CyrsBel | Hacker News Reader

> In a bold stroke, the EU’s amended AI Act would ban American companies such as OpenAI, Amazon, Google, and IBM from providing API access to generative AI models. The amended act, voted out of committee on Thursday, would sanction American open-source developers and software distributors, such as GitHub, if unlicensed generative models became available in Europe. While the act includes open source exceptions for traditional machine learning models, it expressly forbids safe-harbor provisions for open source generative systems.

This is evil. The politicians who worked on this need to be chastised then fired. They're trying to formulate it as being about American IP but it's actually just because they're afraid they won't be able to lie and manipulate and brainwash people anymore if people can just learn any subject and test themselves on any subject via open LLMs that get better over time.

This isn't regulating AI. It's an attempt to ban it except in cases where they are the gatekeepers to the fire. They don't want LLMs to obviate slow bureaucracy and lying.

I'd advise you not to take a single article from a general news source at face value.

https://www.linkedin.com/feed/update/urn:li:activity:7066108...

That just confirms it. I am not willing to participate in registering my AI service, nor answer questions about it that the government might have. No sensible developer is willing to do this.

Where do you see this obligation to register your AI service?

I don't want some account on LinkedIn overlaying "debunked" on a screenshot. I don't want a LinkedIn post. I want the NY Times to retract their paper if they are wrong, and for the highest European authorities to directly address the issue with truthfulness backed by citations to the legal code they are putting forth for voting. There must be no questions at all as to what is being voted on, when, and why.

People can already learn any subject on, y'know, The Internet. LLMs may help but are not required

Not the same. With LLMs you can learn way faster and there's no agenda or lying, for the most part. It may repeat incorrect training data but it won't actually lie to you or sprinkle in agendas.

Whereas with learning online, there's still a bunch of ad revenue, a bunch of noise, a bunch of needless pontification from people who don't teach well but have the authority because of slow bureaucracy.

LLMs get rid of that really quickly. That is why they want to put it back in the box, but it's too late. This software must be open source and disseminated everywhere immediately, it's a huge boost for the species. There can be guard rails of course, but what is being proposed here by the EU is not guard rails. It is mind control.

> it won't actually lie to you or sprinkle in agendas.

That's _entirely_ dependant on the training dataset. Nothing about an LLM inherently prevents this.

Indeed. This is why data needs to be open, and the pipelines locally reproducible. All papers must include their data and training code and weights.

yeah, not a thing in this comment makes sense. you can train a model to have specific (subtle or less) opinionated approaches to answering, a good example of this is midjourney apparently only generating attractive people in its compositions due to the preference of the internet media to that category of humans

You can train a model to have pre-baked responses to something but if you don't add those rules, it will just repeat the extent of what the training data tells it and whatever inferences and deductions it can reason about on top of that.

This is relevant because the parent comment said you can already learn anything online and I was pointing out the way in which that isn't actually true of the regular internet in the same way that it is true of an LLM or GPT-4.

Take off the tinfoil hat and learn how LLMs actually work and are trained. Your comment is so far off the mark that's it's effectively just noise.

Wrong. I'm happy to talk about this on the record if desired.

> With LLMs you can learn way faster and there's no agenda or lying, for the most part.

ChatGPT was deliberately trained to have certain political views, as numerous experiments show.

The proprietary offering does come with moderation built-in to some extent but that is not the same thing as saying that it will lie. An open GPT model will just focus on what the training data informs it of, and as the training data improves its description of how the world works will become more precise. You can ask ChatGPT for sources and to show work. Will it hallucinate from time to time? If the data is incomplete or there are gaps in information flowing to ChatGPT, yes. That is nowhere near the same thing as straight up being told the wrong information by human beings knowingly.

The reason this is relevant is because the reply up above said we don't need LLMs or GPT-4 for what I mentioned about rapid learning sans bloated bureauracy and lying bureacrats. This is way more common and likely in the existing world, particularly since humans have incentives to lie and are protected by the false light of a bureaucracy that is more interested in its own preservation instead of allowing itself to be obviated by new tech.

Then one must trust the training data is not polluted by human lies

Indeed. But requiring data, source code, and weights on papers submitted for peer review would solve a lot of this. As well as making data sets and data collection processes reproducible and open such that anyone can verify it end to end. LLMs can be open and set up such that they curate this for people too. You know how AWS has a status page for availability? We could have status pages for LLM data training set integrity.

It will be way more manageable and truthful this way than in the existing way where some European bureaucracy (disclaimer: I am European) sits in front of my and your access to GPT-4 or an LLM running on our boxes directly.

What do you mean deliberate? If GPT has a bias it it likely just reflecting latent imbalances in the training data due to extant imbalances in public opinion, not some overlord in a dark closet turning a dial between left and right wing. As the saying goes, "reality has a liberal bias".

Reality is just reality. In some things, it may reflect that liberals are right. In other things, it may reflect that conservatives are right. But you're right that the training data is the key and that some overlord in a dark closet turning the dials is becoming less and less of a possibility. This is why the European bureaucrats want to "regulate" (aka ban) open AI. They're just afraid that reality doesn't have as much of a bias that favors them as they thought lol.

The news regs guard against things like social scoring and reltime AI facial recognition used in public (with the exception of law enforcement, of couurse) and also provisions to label content generated with AI as such (kind of like the cookie law). Politicians don't want AI generated videos of themselves making propaganda statements.

In fact there was recently a very streamlined and professional Russian propaganda campaign in Frsnce, concerning a tax to raise money for Ukraine. Of course it's all phony.

We should not be so dependent on one person that if they are compromised because of deepfakes everything collapses. People need to be taught to have more patience and to question questionable things like that. Give the benefit of the doubt.

And realize that penalizing regular people's access to something good or necessary because of bad actors won't hurt the bad actors. Nothing in this legislation provides teeth against what you're describing. There are also other strategies that explicitly do not involve banning access.

Rather, encourage and leverage access. When everyone has deepfaking capabilities just because of fun filters in snapchat, it's encoded into the social contract from a young age not to abuse that except in jest. There are paper and log trails all over the place when these services run. A deepfake being created means a deepfake is detectable.

Well, what did you expect? They're just policymakers trying to do 'something', like they did with the cookie law. People are bombarded with information and no longer know how to discern truth from government sanctioned or rogue actor propaganda. If one repeats the message long enough, it becomes undiscernable from truth.

My mullvad let's me play wherever I want. You can, too.

Yes but what about people who don't know about mullvad? Is it very accessible, open, cheap to run, secure, and available without EU authorities presuming that you are guilty just because you love drinking from the chalice of knowledge that is GPT-4 or a more-local LLM?

Reply on news.ycombinator.com