OpenAI, GoogleDeepMind, and Meta Get Bad Grades on AI Safety
spectrum.ieee.org
spectrum.ieee.org
It's very extensive, and looks at things from a lot of angles. Both sensible angles and... more hollywood-inspired.
They cover everything from jailbreaking, fine-tuning and respecting robots.txt, through board governance structures and SOC2 compliance, all the way to superintelligence and nuclear-weapon-related "model misuse".
I assume the authors of the report set themselves the goal of having at least some info on every possible "but what about" question, no matter how uninformed.
It's not that these guys would not have had bad takes before. Their job is to get as much as possible money allocated to AI safety companies, ie themselves, and the best way to do that is to sell fear. Humans have a natural fear of new things.
The key problem is this: while AI safety laws can mitigate the effects of existing AI, big tech has free reign to invent new models and new AI and introduce it that will propagate much faster than the law.
In my opinion, there can be no safe AI. We can only partially adapt to it and try not to be the ones crushed by it and live out our current lives in the hope that we can survive. The next generations, not so much. Therefore, I conclude that it was a mistake of humanity to ever invent AI, especially since don't have the wisdom to use it nor proper safeguards for it, especially because we can only operate in the realm of ruthless capitalism.
AI is almost a mirror image of this, in that AI is used because it is useful, not because it is ideal. Following the analogy, now that AI exists in increasingly useful forms, calls for AI safety echo those made by anti-nuclear advocates and Luddites centuries ago, but nuclear technology and AI both are simply too useful to be returned to that Pandora's box. So, we are seeing a relitigation of the same arguments used to curtail or prohibit technology in the same way we did in the crypto wars[0], only applied to AI instead of cryptography.
With MAD, the only winning move was not to play, and that thinking largely remains so to this day, as the nuclear stalemate has thankfully not been broken.
The delicate balance of terror[1][2] never ended, and may never end. The players and pieces may change along with the rules and the game board, but war will always be politics by other means.[3][4]
However, with AI perhaps the only winning move is to play early and often? The Balkanization of the AI landscape between China and the rest of the world[5] suggests that both sides view AI technology similarly to nuclear technology: a tool too useful to ignore, an idea whose time has come.
[0] https://en.wikipedia.org/wiki/Crypto_Wars
[1] https://en.wikipedia.org/wiki/Balance_of_terror
[2] https://www.rand.org/pubs/papers/P1472.html
https://web.archive.org/web/20241204201606/https://www.rand....
[3] https://en.wikipedia.org/wiki/On_War
[4] https://warroom.armywarcollege.edu/articles/grand-strategy-c...
https://web.archive.org/web/20241130183926/https://warroom.a...
[5] https://www.rand.org/pubs/perspectives/PEA3703-1.html
https://web.archive.org/web/20241214130855/https://www.rand....
A lot of money, time, and effort is being poured into AI, and putting anything but a trivial chunk of those resources into AI safety, or limiting further development of AI until you know it's safe, means that you'll lag behind your competition who aren't doing such pit stops.
If OpenAI, Google, or Meta decide that they won't release AI models or services into society before they know it won't do serious harm in some unexpected way, then the other two will speed ahead and capture the momentum and market share.
As usual, capitalism pushes us ever closer to the gaping maw by it's very design. If you aren't willing to entertain the destruction of society in your pursuit of new innovation and technology, you'll be out-competed by those who do.