The core problem with building AGI/ASI is we want the good things1 without the bad things2, and the common assumption is that the good and the bad are two sides of the same coin; any attempt at building ASI without also enabling the bad things is deemed impossible.
I claim that systems which are super-human in narrow domains but intentionally sub-human in most domains (NASI3) will provide most of the benefits of ASI without the dangers of ASI. I will argue that:
- NASI can solve many existing problems in the world
- NASI makes deception significantly more difficult
- NASI makes regulation of dangerous capabilities easier
- NASI makes it easy to share non-dangerous capabilities with the entirety of humanity while incurring very little risk
Why might NASI be safer than ASI?
Lots of the dangers of ASI come from a general-purpose intellect being able to act on the world via text4, as opposed to something more constrained. For example, it is far easier to imagine ChatGPT convincing a human to do arbitrary things than it is to imagine StockFish, AlphaZero, AlphaGo, or AlphaFold5 convincing a human to do arbitrary things. This is due to the medium of chess moves/go moves/protein structures6 being inherently less convincing than regular text.
There are two parts at play here, and they work together:
- The artificial narrow intelligence behind AlphaFold is (probably) not capable of general cognition and reasoning.
- Even if AlphaFold were capable of general cognition, the output of AlphaFold is severely constrained (i.e. atom locations with confidences).
While ASI is, by definition, better at humans at all tasks, even an ASI will find manipulating someone to be trickier if the only means of communication the ASI has with the human is atom locations. Manipulating someone is far easier via text (or voice, or video) because people are taught from a young age to give text write-access into the brain, so that thoughts can be shared efficiently. This write-access is incredibly useful, but also dangerous.
It is entirely possible that ASI could still manipulate humanity via atom locations alone, but all means of manipulation would require some level of indirection. For example, one method might be to make protein predictions that are incorrect in subtle ways that either slow down progress in key areas, or — more nefariously — make protein predictions that result in unexpected and unwanted interactions between AI-designed proteins, eventually leading to behaviour that is advantageous to the ASI.
Language is dangerous because language directly inserts ideas into your brain. Ancient Greek is not dangerous to me because I do not speak ancient Greek, and so before any ancient Grecian wishes to manipulate me, they must figure out how to communicate with me.
The ability of a given communication medium to be used for manipulation of other players could be tested by designing a strategy game where AI players could gain significant reward from communication, and then restricting the only means of communication between the players to the communication medium under question. It seems clear that protein predictions map less intuitively onto ideas such as “I’ll pay you 100BTC to synthesize this molecule” than literal text “I’ll pay you 100BTC to synthesize this molecule”. Chess moves might be easier to communicate this information, since the trade could be expressed as a sacrifice, although conveying the mapping of “you capture my queen” to “100BTC” and “synthesizing this molecule” as “I capture your bishop” might be difficult.
Limited communication mediums are our friends, and we should embrace them if we want to reduce the risk of manipulation. NASI deals exclusively with hyper-specific outputs, thereby reducing the risk of manipulation.
NASI can solve many existing problems in the world
What’s the purpose of ASI anyway? Many will say it’s for various humanitarian or world-improving side effects, problems which people struggle to solve but a smarter intelligence might be able to solve faster. Such problems are cancer, global warming, political coordination and unrest.
NASI makes deception significantly more difficult
NASI makes regulation of dangerous capabilities easier
NASI makes benefits-sharing nearly risk-free
Furthermore, it’s not clear what sort of general intelligence can arise if the output is protein predictions. It seems likely that the model would gain some sort of model for atom-level physical interactions, but could such a model learn manipulative debate? Investigating this question would require training a narrow ASI and then attempting to put it into situations where it would be able to make use of any latent knowledge it gained that might be applicable to “the bad stuff”.
One other advantage of NASI is that it would be significantly easier to prevent misuse. Since any single intelligence would be super-human in only one domain, we would need multiple NASIs in order to achieve the goals of current ASI-builders (curing cancer, world hunger, etc). Each of these NASIs would, by definition, be specialised in one particular area of expertise and sub-human in other areas. NASIs specialised in high-risk areas of expertise (biology, AI research, chemistry, negotiation, politics) and the datasets usable to train them could be specifically targeted for regulation while NASIs in low-risk areas (logistics, supply-chains) and the respective datasets could be put under lesser regulations or none at all.
General purpose datasets which would be applicable to AGI/ASI could be more thoroughly protected, given the known dangers of AGI/ASI.
Moreover, concentrating on NASI is a more incentive-compatible solution than stopping or pausing AI research until alignment research can catch up. Focussed efforts on NASI have already been shown via AlphaFold, GenCast, GraphCast to provide benefits
NASI would be an explicitly worse intelligence when compared to ASI, but it ~solves the alignment problem while still giving us 90% of the benefit that we are hoping to get from ASI. At the very least it buys us time to actually solve super-alignment, while being immensely easier to regulate. We could straightforwardly outlaw general intelligences, in favour of narrow intelligences. Narrow intelligences designed for certain benevolent purposes would be generally available, narrow intelligences designed for possibly-dangerous purposes (biology, tactics, politics, business) could be more heavily regulated.
It’s instructive to imagine the space of ideas that AlphaFold could think compared to the space of ideas that ChatGPT can think. ChatGPT (probably) has a larger volume of space, although AlphaFold almost certainly covers regions of thought-space related to protein folding that are not reachable by ChatGPT.
mathematics, joining different domains required to solve problems, narrowness won’t be enough
Even if 1% of probelms are solvable by narrow AI, it’s fine, those problems can be solved
There’s economic arguments as well, we can have 100s of thousands
Soliving world hunder, requires AGI, individual problems (farming, logistics) can probably be solvable by ANI
Footnotes
-
(cure cancer, feed everyone, end poverty) ↩
-
(bioweapons, misinformation, military super weapons) ↩
-
Yes, narrow AI is an existing term that maybe could have worked here instead of defining a new term (NASI), but given the lack of a standard definition for “AI”, I felt it important to distinguish NASI from any program that solves MNIST. ↩
-
Also via photo/video or other multi-media, but in general we give ASI the ability to communicate in a general purpose way. ↩
-
I didn’t intend this to be a DeepMind shill-piece, but they’ve done a lot of very impressive and obviously-impactful narrow ASI research. I’m less inclined to include research on video-game AIs because it’s a bit clearer how that could lead to strategy, deception, and manipulative behaviour being learnt and how these AIs could learn to influence human behaviour by being very good at these video-games. ↩
-
For the rest of this essay I’ll use AlphaFold and protein folding as an example NASI because it was largely unexpected and is clearly super human, but NASI could equally apply to AlphaGo, StockFish, or other models where the output is sufficiently constrained. ↩