OpenAI's purpose is to build "safe" (i.e., controllable) artificial general intelligence. It's motivated by two factors: (a) optimism about how intelligent AI can become, and (b) fear about what said AI actually
does when it gains enough intelligence. This latter clause is responsible for a lot of dystopian sci-fi literature about AI killing humanity (see the Terminator series as an example of AI maliciously destroying humanity, and the
Paperclip Maximizer scenario as an example of AI innocently destroying humanity). That's probably isn't going to happen...but then again, you don't know
what's going to happen...and "fear of the unknown" is the most primal fear of all.
That's probably the main reason why this debate over 'harmful content' is so interesting...if OpenAI can't even stop people from producing content they don't like, then how could OpenAI stop people from doing more serious stuff with this technology? How can it stop the AI itself from doing harm to other intelligent forms, without any human prompting?
OpenAI, for its part, has released a blog post discussing
their experiments in self-censoring their AI by fine-tuning it on a curated dataset. All well and good (unless for some reason, the user doesn't trust the watchmen). This suggest that it could be possible to reduce harmful content without engaging in privacy violations...
but mere reduction might not be enough to satisfy the "powers that be".
The irony here is that OpenAI doesn't have a monopoly on text generation. GPT-Neo is in development, of course. And it turns out that
Google and the
Beijing Academy of Artificial Intelligence has both independently developed models that contain over 1 trillion parameters. The largest GPT-3 model, by contrast, has a mere 175 billion parameters.
So even if OpenAI build the most safe, controllable AGI possible, that won't stop
others from building uncontrollable AGI and messing everything up. And if the uncontrollable AGI is more powerful than the controllable AGI (Google, Beijing Academy)...or the uncontrollable AGI offers more freedom than the controllable AGI (GPT-Neo), then why would anyone use OpenAI's model?
In addition, I can see GPT-Neo, Google, and the "Beijing Academy" agree to do self-censorship...but they might implement self-censorship in different ways, or even in completely opposite ways (OpenAI wants their AI to produce certain rhetoric about democracy and human rights...do you think the Beijing Academy is going to let their AI generate exactly the same rhetoric?). Would OpenAI be satisfied with that? Would OpenAI be satisfied at this proliferation of AIs, which is starting to look more like the proliferation of nuclear weaponry? Will this lead to a dangerous
AI arms race, where "safety" takes a backseat to "building the best AGI model"?
In
a 2018 blog post, OpenAI announced:
We are concerned about late-stage AGI development becoming a competitive race without time for adequate safety precautions. Therefore, if a value-aligned, safety-conscious project comes close to building AGI before we do, we commit to stop competing with and start assisting this project. We will work out specifics in case-by-case agreements, but a typical triggering condition might be “a better-than-even chance of success in the next two years.”
I thought that they would actually do it, but that was before OpenAI decided to turn from being non-profit to
capped-profit in March 2019, and before
accepted Microsoft investment in July 2019. Will they actually give up? Is it too late to give up, now that Google and the Beijing Academy have also stepped into the fray?
I don't necessarily
want OpenAI to have a monopoly on text generation - centralizing power in the hands of a single entity is bound to lead to problems. But at the same time, I do share OpenAI's concerns about artificial intelligence...and am somewhat concerned about this technology being abused. So I have very mixed feelings about what has happened.
In the past, I thought OpenAI had a chance (however, small) of producing safe artificial general intelligence and reducing the probability of sci-fi nightmare scenarios. Now I see that OpenAI was doomed from the start.