Comment by andrewparker
2 days ago
OP point out that OpenAI used the "too dangerous to release" marketing ploy with GPT-2... Positioning this as "both sides" have played this card.
But at this time Dario was at OpenAI and was a co-author on the GPT-2 research paper announcing the model.
The "too dangerous to release" approach has been him the whole time, at both companies.
Maybe because Dario was actually reasoning through potential risks, rather than blindly thinking everything will be okay?
Thinking GPT2 was too dangerous is and was absurd.
It's dramatic for sure, but at the time it was genuinely alarming to think through the implications of a machine being able to generate plausibly human-authored text. I think many of the alarming implications have in fact come to pass, and the world is a more strange and dangerous place than it was before.
At that time, nobody believed a dead internet was technically feasible. Maybe this is hard to remember now.
The "danger" was in terms of spam / misinformation proliferation, not the same category of capabilities adjacent risks current discussed.
You can hold your own opinions on spam/misinformation as a problem, but to say there was no credibly anticipated outsized downside to a sudden jump in human-passing text generation feels pretty off to me.
5 replies →
All that happened was announcing said potential risks and then continuing to launch and push and release it anyways. Something that Anthropic continues to do under his leadership.
Google also analyzed for risks earlier in development of "Meena" and "Bard" and chose not to release, instead. And then got caught flatfooted when OpenAI did so anyways. (They also I think didn't really see a compelling business case for public propagation of it, either).
It starts to look very much like what is really happening with Anthropic is a lot of cynical attempts at regulatory capture. Make grand proclamations. Get your stuff out there first. Ask for regulation and then kick the ladder away from underneath you. But they did it clumsily and failed to grease the right palms.
I said this in the other thread, but they were proven right about their gpt2 worries, weren't they?
From the original 2019 release:
> We can also imagine the application of these models for malicious purposes , including the following (or other applications we can’t yet anticipate):
> These findings, combined with earlier results on synthetic imagery, audio, and video, imply that technologies are reducing the cost of generating fake content and waging disinformation campaigns. The public at large will need to become more skeptical of text they find online, just as the “deep fakes (opens in a new window)” phenomenon calls for more skepticism about images.
These worries are why they stated they were cautious in rolling it out
https://openai.com/index/better-language-models/
> The public at large will need to
Ah, yes. You see, it’s not them who are wrong for knowingly releasing something they knew to be harmful, it’s everyone else who needs to change. That seems reasonable. Humanity is famous for being able to rapidly adapt to fast changes as one voice. Oh, wait…
They are no different to the tobacco and oil companies. They know the harm they’re causing but care about personal profit about everything else.
I'm not an AI booster, but in this case I'd say that pausing the rollout for mitigations (such as public education) to be put in place was the responsible course of action.
With the benefit of hindsight, you can certainly argue that the pause wasn't long enough or that the mitigations weren't sufficient. But that wasn't a view held by many at the time - indeed, it was mocked as a marketing ploy (and still is; see gp's post as evidence).
2 replies →
Yes, it's not their fault, that people are using the tool they made in a malicious way.
I hate ClosedAI as much as the next guy, but this is an extremely illiberal take. It's not the kitchen knife manufacturer's fault that people are using their product for murder, it's not my fault that people are doing crimes over the Tor relay I run.
The Tobacco industry is evil because it misleads the public about its product being poisonous and bribes politicians through widespread corruption. Tobacco is also different because it is not a neutral tool that can be used for good and bad, but poisonous and will harm you no matter how you use it.
1 reply →
To be fair, generative AI is wrecking society in new and unexpected ways every week. From lies and misinformation to people choosing chatbots over therapy and relationships, there's a profound impact on society that will only get worse in the coming years. The look for junior programmers who are capable enough to get anything done when the AI is down has been depressing, and things are looking much worse for the years to come.
Important safety features ("do not generate child porn", "statements should be factual or backed by evidence") were simply not part of the design of these systems and have yet to truly solved to this day, but AI companies decided to release these technologies onto the general public regardless of their glaring flaws.
I like AI for its shitposting capabilities and its neat parlour tricks, but I also believe so far it has been a net negative for everyone but the richest minority of society who benefits from firing people and having computers do half their jobs badly. It's too late now, but in hindsight I do agree that these systems were too dangerous to release in this shape.
I feel like it could be a law that there is essentially no way to guarantee that AI is any more or less safe than humans. It kinda seems incompatible with what we understand to be "intelligence" which arguably requires a certain unpredictable freedom...Has a method of "baking in" such safety features even been conceptualized? Or is it just a matter of nurturing/raising/policing them after the fact and hoping for the best like with us?
Kind of taking this a step further, what is the worst thing that an AI run amok could do to existentially threaten the human race? I feel like almost anything could be short-circuited by some form of "pulling the plug" before it got too far. But, hypothetically, if it were possible to launch nukes without human intervention, or with maybe a small amount that could be socially engineered, that seems plausible (or releasing some kind of super-pathogen that is stored in a lab somewhere).
So, what if, along the lines of MAD doctrine and the plot of Battlestar Galactica, the best thing we could do for AI safety is just to engineer our other systems so that a hypothetical superhuman adversary could not use them against us? Which is just making our world safer all around rather than trying to kludge arbitrary limitations into an "intelligent" system.
(This doesn't really solve AI child porn and fake news but those things are mostly just imaginative reflections of the people using them and you can't really fix that any more than you can stop people from doing it themselves)
>Kind of taking this a step further, what is the worst thing that an AI run amok could do to existentially threaten the human race?
Probably convincing humans that untrue things are true. Think less "AI launches nukes" and more "AI convinces people in country A that country B is trying to subvert & destroy them, while it does the same for country B wrt A."
2 replies →
AI is often better than therapy as reported by users. Therapy has some inherent dark-patterns that AI doesn't have yet, like the therapist's financial incentive to trickle a solution to your problem to preserve their income.
With ai I can get government bodies to do their job. Escalate and make formal complaints until it’s done. Manually would take ages to find the right law and draft a proper complaint.
AI gives more than a billion people instant access to knowledge, it is starting to accelerate scientific research, it democratized software development, design, and illustration.
I strongly disagree with the opinion that it has been a net negative.
Lies and misinformation, or choosing chatbots over therapy and relationships might sound scary, but as of now I see practically negligible impact there. Even social media is still roughly as usable as before AI.
Dario's brain child