Meanwhile, there are deep problems with the “let’s build transformative AI in order to make sure it’s safe” strategy. In particular, OpenAI and DeepMind both express that they want to race to generate highly transformative intelligent systems. The goal they both profess is to be the first to develop them so that they can exercise responsible stewardship and ensure that it is as aligned and beneficial as possible. This is a benevolent form of what Nick Bostrom refers to in Superintelligence as gaining a “decisive strategic advantage” which may make the first developer of particularly transformative AI too powerful to compete with. There are many problems with this strategy including: (1) It is entirely based on racing to develop transformative AI, and faster timelines exacerbate AI risks. This is especially perverse if multiple actors are competitively racing to do so. (2) Nobody should trust a small set of people like Sam Altman and Demis Hassabis to unilaterally exercise benevolent stewardship over transformative AI. Arguably, under any tenable framework for AI ethics, a regime in which a small technocratic set of people unilaterally controlled transformative AI would be inherently unethical. Meaningful democratization is needed. (3) OpenAI’s approach to DALLE-2 should further erode confidence in them in particular. Their overly-convenient technical report on risks that failed to make any mention of copycatting combined with how quickly they worked to profit off of DALLE-2 are worrying signs. (4) Copycatting makes racing to build transformative AI strictly more risky. Even if one fully-trusted a single actor like OpenAI or DeepMind to exercise perfect stewardship over transformative AI if they monopolized it, how quickly DALLE-2 was copycatted multiple times suggests that copycatting may undermine attempts at benevolent strategic dominance. Copycatting would most likely serve to broaden the set of technocrats who control transformative AI but still fail to democratize it. So if a company like OpenAI or DeepMind races to build transformative AI, and if it is still copycatted anyway, we get the worst of all worlds: unsecure, non-democratized, transformative AI on a faster timeline. If a similar story plays out with powerful, highly transformative AI as has with DALLE-2, humanity may be in trouble.
Let’s be honest, a lot of the claims by OpenAI and Deepmind shows bad signs of having motivated reasoning. This is equivalent to a tobacco creating company claiming that their research helps make tobacco safe.
No, it only benefits the company in the form of profits.
Let’s be honest, a lot of the claims by OpenAI and Deepmind shows bad signs of having motivated reasoning. This is equivalent to a tobacco creating company claiming that their research helps make tobacco safe.
No, it only benefits the company in the form of profits.