Reflections on AI
Three Approaches to AI Safety
Apr 17, 2023
A lot of people know that there are seismic shifts happening right now with artificial intelligence.
A lot of people may not know that many of those driving much of the progress here, in both research and industry, are very concerned about the dangers of AI. This includes Elon Musk, Steve Wozniak, Andrew Yang, and many other researchers, executives, and other prominent figures who published an open letter in March imploring all AI labs to pause AI experiments for 6 months. It also includes the researcher Eliezer Yudkowsky who, along with many others, believes AI will inevitably lead to human extinction.
To truly appreciate the depth and urgency of this concern, consider that the majority of AI researchers polled in a 2022 survey estimated somewhere between a 5-10% probability of AI “causing human extinction or similarly permanent and severe disempowerment of the human species”. (There were two similarly worded questions to this effect, with median responses of 5% and 10%.)
Given this state of affairs it should not be surprising that quite a lot of thought (although most of the individuals above would argue not enough thought!) has been put into the topic of AI safety: how best to responsibly proceed with artificial intelligence without leading to the downfall of civilization. Broadly I am aware of two approaches that are being explicitly championed by thought leaders in this space; and there is a third approach that I think is pretty obvious, and which I think it is safe to assume is being implicitly pursued by at least a percentage of those in positions of power who care about this topic.
These approaches are:
- Terminate (shut it off)
- Inoculate (let it out)
- Consolidate (lock it down)
1. Terminate (shut it off)
The most dramatic proposal on the table that I’m aware of is to completely shut down AI research. This is what Eliezer Yudkowsky is advocating. In his article published in Time magazine (linked above), he writes:
Frame nothing as a conflict between national interests, have it clear that anyone talking of arms races is a fool. That we all live or die as one, in this, is not a policy but a fact of nature. Make it explicit in international diplomacy that preventing AI extinction scenarios is considered a priority above preventing a full nuclear exchange, and that allied nuclear countries are willing to run some risk of nuclear exchange if that’s what it takes to reduce the risk of large AI training runs.
Here is my oversimplified version of the argument for this:
- Superintelligent AGI, when achieved, will be unstoppable by humans. (I think of this as AI achieving “escape velocity”, at which point human intelligence, like gravity, will no longer be able to contain it.)
- Superintelligent AGI is likely to result in human extinction, because it will be capable of exterminating us, and we don’t know how to instill it with values or design it in such a way to make sure it doesn’t do that.
- We don’t get multiple tries to solve this problem. Unlike every other line of scientific research up to this point in human history, we don’t have the luxury of learning from failed experiments and forming new hypotheses if we get this one wrong on the first attempt. The “failed experiment” would be that AI kills everyone.
- This isn’t just a theoretical problem anymore; with the surprising rate of progress in LLMs like GPT-4 we are racing towards superintelligent AGI faster than many researchers anticipated. Therefore we are already too close for comfort to point 1 above, and the only sane thing to do is just stop.
Personally, the part of this argument I have the hardest time subscribing to is the 2nd point. It’s possible I simply haven’t come across the version of this point that resonates for me. But my best attempt to explain it, for now, is that we don’t have a robust argument for why a superintelligent AI would not wipe us out. I’m genuinely not sure if this is a logical argument or a pragmatic one; i.e. “If we aren’t confident it won’t happen, we ought to assume it will happen.”
Silly reference: this is sort of the same point Bruce Wayne makes in the film Batman v Superman when explaining why he believes he needs to fight Superman.
If we believe there’s even a one percent chance that he is our enemy we have to take it as an absolute certainty.
Note: I am not saying this is the same as Yudkowsky’s argument. Rather, it is the same as my probably-wrong attempt to represent Yudkowsky’s argument, given my very shallow understanding of it.
Is this even possible?
It isn’t obvious to me this is possible. But I am currently listening to Lex Fridman’s interview with Max Tegmark, who replied to this question in the affirmative and gave the example of human cloning.
According to Tegmark, we gained the technological capability of cloning humans decades ago; but thanks to public pressure we collectively avoided ever actually doing it. (He shared that there is one known historical case of a Chinese scientist doing it, but he was subsequently imprisoned by the Chinese government.)
And then what?
Something I’m not currently sure about is whether Yudkowsky and others arguing for the “Terminate” approach believe we should ever resume AI research after shutting it down. Should we just discard everything we know so far? Lock it up and throw away the key?
Or are there some thresholds we need to reach, e.g. in the areas of alignment and interpretability, after which it will be safe to pick up where we left off? (Is there any chance we could ever reach these thresholds without continuing research into more advanced AI? Sam Altman doesn’t think so, according to my memory of his conversation with Lex Fridman.)
1.1 …or we could just slow it down (decelerate)
A less dramatic proposal than shutting off AI research is to hit the pause button for a little while and give AI safety researchers some time to catch up. This is effectively what the open letter that Musk et al signed is suggesting. The letter proposes a 6-month pause on any AI experiments using systems as powerful as GPT-4 or greater.
During this time, the letter advocates for AI labs to collectively focus on safety, not building more powerful systems; and it also encourages AI developers and policymakers to collaborate on instituting thoughtful regulation to direct the safe development of AI systems moving forward.
2. Inoculate (let it out)
If the first approach to safety, i.e. “Shut it off”, is analogous to a quarantine, then the second approach, “Let it out”, would be more like a an inoculation campaign.
Think of the development and spread of AI like a contagion. Preventing the spread of an infectious disease via quarantine can be very effective, but it requires significant coordination. This is achievable to some extent in states with powerful centralized control (e.g. China). It turns out to be much harder in parts of the world with far greater individual freedom (e.g. the United States).
However, even in the parts of the world that are more centrally controlled, globalization makes it difficult if not impossible to truly contain a highly infectious contagion, when people are still passing in and out of the country’s borders.
An alternative approach is to willingly allow the contagion to spread, to give the population some opportunity to develop natural immunity to it. (You could also try to develop a vaccine, which would have a similar effect. Interestingly, if you think about it, the way a vaccine works is more or less a microscopic version of the same phenomenon that happens with a population developing natural immunity through infection. It’s just on the scale of a single human body versus an entire population.)
This is the approach that OpenAI assert they are taking: exposing the world to AI capabilities as they’re developed so that we can build up natural defenses against them.
Is this even possible?
It does seem possible to me. As I pondered yesterday, to be successful with this strategy inoculating for any threat seems to require at least two things:
- You need to be able to present a non-lethal dose of the threat. If the smallest amount we can produce is already lethal, then we’ll die. With AI, part of the risk we’re facing is that we may not recognize when we’ve already exceeded this dosage. (Again, this is Yudkowsky’s fear.)
- The recipient needs to be capable of developing effective antibodies. A vaccine doesn’t do you any good if it just makes you sick, and your body doesn’t develop any new defenses as a result. Thus far it’s difficult to say whether we are hardening ourselves against more powerful AI in any meaningful way. We probably are in some ways, e.g. the open letter raising awareness of the dangers of AI and inspiring some smart scientists to get more involved. On the other hand, people are already trying to get AI to destroy the world. So it’s a mixed bag.
And then what?
I think the hope with this approach is that eventually, by the time we do inevitably achieve superintelligent AGI, human society will have developed a robust set of defenses (effective strategies for achieving alignment and ensuring AI safety, international agreements enforcing safe deployment, education programs, etc.) to protect ourselves.
Let’s hope so!
3. Consolidate (lock it down)
I haven’t heard anyone explicitly advocating for this approach. In fact, what I’m about to describe as an approach to AI safety is more commonly identified as a concern that many people justifiably have about the future of AI.
Nonetheless, while this would be bad in many ways, another path towards making the development of AI more safe, I suspect, would be to restrict access to a much smaller group of humans. Realistically this would be AI labs at major corporations, and world governments. Public APIs like ChatGPT, DALL-E 2, Midjourney, and others would be taken offline. Open source projects providing reusable building blocks for building AI systems would be removed from public repositories and made illegal to use with steep penalties.
I realize this all feels a bit dystopian, and it probably would be. It would also put us in that really unfortunate position of immense power being consolidated under a tiny percentage of human beings on the planet, and large corporations such as Microsoft and Google exerting more control than ever before. (Alternately the government could claim all of the power, which probably wouldn’t be much better.)
How is this an approach to safety at all? It would significantly reduce the surface area of the existential risk we face. It would make AI research much more like the Manhattan Project, where a small team of scientists with top security clearance worked on a very dangerous technology that was not shared with the general public. Yes, this put tremendous power into the hands of very few; and yes, it still led to the creation of some incredibly destructive technology, resulting in the loss of millions of lives.
But what is happening today is a bit like if all of the research conducted for the Manhattan Project were published in real time, daily, in every newspaper in the country. And if the materials to make nuclear weapons at home were readily available to everyone. (Well, maybe not quite like that last part. I don’t have a great feel yet for the amount of computing power you actually need to create something like GPT-4. But given the sheer number of personal computers connected to the Internet and the fact that peer-to-peer technology is an established thing, it’s difficult for me to imagine that orchestrating that much computing power would truly be prohibitive.)
I will say this. At a high level, it seems like a good thing that the number of people with the power to launch nuclear weapons is very small. I would prefer for it to be zero; but if it’s going to be a positive number, I’m glad it’s in the single digits and not in the millions.
Is this even possible?
Once again, I’m not sure. If the U.S. government passed a law tomorrow making AI illegal everywhere… would the development stop? The consequences would have to be absolutely massive, and there would need to be unmistakable early demonstrations of the government’s commitment to enforcing the law. Like “Sam Altman sentenced to life in prison” or “Google fined $100 billion” or something like that.
It’s really hard for me to imagine that happening, but what do I know?
And then what?
Again, good question. The future where humanity survives but we’re all powerless to resist the permanent control of a monopolistic organization, or an iron-fisted government, doesn’t exactly feel like one to aspire to. I suppose I would probably still pick it over extinction.
Ben Franklin would not approve:
Those who would give up essential liberty to purchase a little temporary safety, deserve neither liberty nor safety.
On the other hand, I’m also not sure how much thought Ben Franklin put into AI.