1) There was a non-zero chance that their releasing it would do more harm than good by making things like fake news, spam, and sock puppeting cheaper and more scalable, and
2) It's very likely that eventually they or others will develop more unambiguously dangerous models, and it's valuable to begin experimenting with responsible disclosure policies now so we're better prepared.
Perhaps some here would disagree with the premise of (2), that it's possible for a model to be dangerous. I can understand that.
Working in the field, and having been exposed to many of the same arguments that are likely informing OpenAI's concern, I think it's valid. Valid enough at least that I'm prepared to believe that OpenAI is acting in good faith, and that the disclosure policy is not "just marketing".
A lot of what AI can and cannot do depends on cost and computational power as much as anything else. If I understand correctly, there's a whole bunch of "you could but it's prohibitively expensive" stuff going around.
I'm convinced that the people behind OpenAI understand this nuance, even if most non geeks struggle with the difference between possible and feasible. This means that OpenAI mixed the two up on purpose.
If the people warning us of the robot apocalypse can't be trusted to communicate openly and honestly about odds, then what do we do? They clearly have no qualms about misinforming the public to serve a hidden agenda. It might be pretty harmless in this case, but it's indicative of a cultural pattern.
Basically I just hope (and kind of believe) that we'll never achieve AGI because that would make all this talk unimportant.
You should probably actually read what OpenAI have to say before making claims about what they have to say.
“We are aware that some researchers have the technical capacity to reproduce and open source our results. We believe our release strategy limits the initial set of organizations who may choose to do this, and gives the AI community more time to have a discussion about the implications of such systems.”
Musk said in 2015 that the point of OpenAI was "to empower as many people as possible to have AI. If everyone has AI powers, then there's not any one person or a small set of individuals who can have AI superpower." Sikka argued that "openness" was the fundamental reason he supported the project. OpenAI was heavily criticized by other AI risk researchers for its public approach, and argued strongly that it was doing the right thing. It specifically invoked the threat of AI tools being abused by small groups of people in secret as more pressing than the threats from self-directed or public-use AI.
But the closest the statement comes to addressing that is describing the restrictions as "an experiment", and mentioning that in 2018 OpenAI added a charter with a caveat saying it might publish less stuff. That caveat and the decision to invoke it here are essentially unexplained; since the charter almost certainly postdates some GPT-2 work it seems to be a circular argument.
Frankly, GPT-2 seems like exactly the sort of work OpenAI was worried about in the first place. Ten thousand people can troll and spam ideas without any AI support, but with a GPT-2 equivalent it becomes possible to a government, company, or intelligence group to massively scale its impact on public debate. If OpenAI thinks state actors can't replicate their work, the fears invoked at its founding are called into question. If it instead thinks public access to such tools is dangerous and they need to be kept in select hands, the project's entire rationale is invalidated. If they simply hope to spark discussion and give people time to prepare (mentally or technologically) for the idea of compelling auto-generated text, that really ought to be said more clearly.
> Sikka argued that "openness" was the fundamental reason he supported the project.
I believe the relevant quote is this one
> Sam asked me if I would be ok with the fact that such an endeavor would be untethered and would produce results generally in the greater interests of humanity, and he was somewhat surprised by my reaction, that indeed I would only support this venture if such an openness was a fundamental requirement!
https://web.archive.org/web/20151222094518/http://www.infosy...
I'll leave this open to interpretation; I think there are multiple ways of taking it and I'm not convinced which are accurate.
> OpenAI was heavily criticized by other AI risk researchers for its public approach, and argued strongly that it was doing the right thing. It specifically invoked the threat of AI tools being abused by small groups of people in secret as more pressing than the threats from self-directed or public-use AI.
I'm not sure who you are referring to here. The only critics I've heard claiming OpenAI are too open are those who believe in Bostrom-style AGI risk, which is the idea that (far-term) AI is intrinsically dangerous, rather than being dangerous predominantly because of malicious use.
> If OpenAI thinks state actors can't replicate their work,
I don't believe OpenAI believes this.
They have been pretty clear about this [1]:
"We believe our release strategy limits the initial set of organizations who may choose to do this, and gives the AI community more time to have a discussion about the implications of such systems."
I think this is reflective of a substantial disconnect and polarisation in society, because OpenAI's position here was driven by a very fundamental intuition about human nature which is not universally shared.
Namely, their rationale was something like this: the world is dominated by people whose opinions are fundamentally shaped by what others are saying, and moreover, by how frequently other people seem to be saying it. They don't really think about things or rely on their own experience: they're just mimics.
In such a world being able to auto-generate fake messages of support for some political position or another at scale would give you immense power, because the population would automatically swing behind you based merely on the perception that everyone was swinging behind you.
So that's their fear. But is it realistic?
Well, this is where we get into the polarisation. "You aren't smart enough to have an opinion" is the sort of viewpoint that leads to an elitist vs populist conflict. There have been reams of analysis about this and it's not really AI specific - e.g. did people vote for Brexit because of Twitter, or because of things they saw in the press, or did they vote based on their own experiences, or what their close friends/family thought, or what mix of "all of the above" is the truest mix?
When they refused to release GPT-2 the OpenAI researchers took an extreme on that spectrum, asserting that essentially they had built a mind control device. But of course which didn't work on themselves, only on lesser minds. Not surprisingly this was very controversial, albeit I think the Valley/Hacker News set would be surprised at just how controversial it would have been if anyone outside the AI world had really noticed. You can't tell most of the world their opinions aren't really their own and not expect pushback.
Personally I don't share those intuitions at all. The controversy over DeepFakes is similar. Photoshop has existed for decades and hasn't led to a dystopia, but suddenly DeepFakes is going to create fundamental social change? I don't think so.
AI researchers, especially in academia, need big narratives to keep up the funding and suggest they're on a mission to save the world vs e.g. making slightly better playlist recommendations. GPT-2 seems like a classic case of this being taken to absurdity.
Having read everything I could find from OpenAI on this topic, I have no recollection at all of such claims by OpenAI. Sources would be appropriate.
Here it is again.
https://openai.com/blog/better-language-models/
They said
"Due to concerns about large language models being used to generate deceptive, biased, or abusive language at scale, we are only releasing a much smaller version of GPT-2 along with sampling code."
And in case it wasn't clear what they mean by "abusive language", they clarified:
"Today, malicious actors—some of which are political in nature—have already begun to target the shared online commons, using things like “robotic tools, fake accounts and dedicated teams to troll individuals with hateful commentary or smears that make them afraid to speak, or difficult to be heard or believed”"
"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."
Or summarised:
• We think people believe things they read online today.
• We think there are political "disinformation campaigns" being waged, by the sort of people who would use AI to wage them more effectively.
• We think GPT-2 would let them create "deceptive, biased or abusive language" automatically, and this is too dangerous to release because people aren't skeptical enough of what they read online.
• We think the entire population needs to change its behaviour for GPT-2 to be safe to release.
I think that directly supports my arguments. Their whole worldview assumes the population is naive, easily manipulated by text and more importantly the volume of it ... they also draw a bogus analogy to DeepFakes. It's bogus because nothing stops anyone writing words today that OpenAI staff might find "biased" or "deceptive", literally anyone can do that. But substituting a celebrities face onto porn would require very advanced video editing skills most people don't have: DeepFakes is a truly new capability, whereas GPT-2 is merely a capability of scale.
To believe GPT-2 changes anything you must believe the majority of people make up their minds on political issues based purely on volumes of anonymous online comments. Hence my argument.
More like asserting that they had built a nuclear reactor, which someone else could easily turn into a nuclear bomb. They haven't turned it into a nuclear bomb themselves, so the fact that they haven't been blown up by it doesn't require any special justification.
If they just said "we are developing AI because AI is awesome" I would have a lot more respect for them than I currently do. I wish they would at least change their name.
As it stands, it's hard for me to see their release process as truly an experiment and not as just a clever marketing ploy - because people are talking about them a lot more now than before.
People are products of their experiences, sure. That experience is much wider than "possibly bot authored comments on social media", but OpenAI is acting as if it isn't.
Most people won't agree with their stance. It does however seem to have some weight in certain social circles, the sort of circles that assume most people aren't educated enough to make good decisions.
The idea of a good text-generating AI possibly being able to flood forums is a little scary, and the idea that even the little bit of caution OpenAI has done is elitist is ridiculous.
Even if it is entirely safe, it's still good to withhold it because it helps create a culture where people think about the safety of their projects before releasing them
OpenAI's raison d'etre was democratizing access to AI tools, and preventing them from being abused by concentrated powers like governments. If replicating GPT-2 is trivial for state actors but prohibitive for hobbyists and other private citizens, it's creating the same issue Musk described OpenAI as setting out to oppose. Even the general idea of treating AI tools as hazardous-by-default goes a long way to validating the project's original critics.
I don't think it was a moral question. I think the lawyers got involved and were uncomfortable with potential litigation exposure.
MSFT has sort of invested $1 billion into OpenAI so I guess it worked!