7 ms·
The risks of "AI harm" seem overblown to me. The negative examples in the news (eg Sydney) look more like PR problems than societal risks. While improved LLMs w
by verteu 4y ago
The risks of "AI harm" seem overblown to me. The negative examples in the news (eg Sydney) look more like PR problems than societal risks. While improved LLMs will aid bad actors in spam/astroturfing, there's almost nothing we can do about it -- gatekeeping the models helps a little, but they seem to be reverse-engineered in <1 year regardless.
I can't help but wonder who's pushing this "AI safety" narrative, and why.
- jerpint 4y agoIt’s PR on top of PR. OpenAI is “dedicated to making ai open TM” but also to making AI safe, so for “our safety”, opts not to release models. Not for business reasons /s
- ShredKazoo 4y agoThe AI safety narrative is being pushed by people who think that if we can't control today's AI, we won't be able to control tomorrow's AI either. Yep, current AI harms seem fairly limited. And when gorillas first saw humans on the savanna, the human harms to their habitat were also quite limited. We may only have a single chance to get smarter-than-human AI right.
- RandomLensman 4y agoThe AI safety discussions mask that we are nowhere near high powered AGI. Maybe we'll never get there, but the discussions make AI look more powerful than it is, i.e., it's a marketing angle.
- HDThoreaun 4y agoIf there's a 1% chance of a fast takeoff within the next decade is that not something we should take seriously? I really don't think AGI within the next decade is completely off the table.
- RandomLensman 4y agoWhat if the chance is only 0.001%? How would we know?
- HDThoreaun 4y agoI mean yea, that's the real question. Not where AI is now, but where it will be a decade or two from now. Talk about AI not currently being dangerous is just ignoring the point AI safety people are making.
- RandomLensman 4y agoShould we worry about other non-existing technology the same way? We had the nano technology people pushing all sorts of scary stuff a few decades ago and nothing came of it, for example.
- HDThoreaun 4y agoMaybe. The thing about preparing for unlikely outcomes is that when they don't happen it's easy for detractors to say "see how stupid we were to worry about that" even though everyone agreed that nothing happening was the most likely outcome all along.
- ShredKazoo 4y agoAgreed, and the pandemic seems like a good recent example of this.
- RandomLensman 4y agoSo how worried should we be about the impact of as yet to be invented anti gravity technology, for example? My point is: doomsday scenarios are often used to gain influence/funding/etc., which makes me somewhat skeptical of the motives of people pushing them.
- 4y ago
- ShredKazoo 4y agoIt wasn't even 2 weeks ago that HN users were arguing we are "beyond the point of no return" https://news.ycombinator.com/item?id=34831728 https://news.ycombinator.com/item?id=34831728 and "Pandora's box is open, the genie's bottle is uncorked" https://news.ycombinator.com/item?id=34809607 https://news.ycombinator.com/item?id=34809607 I sit somewhere between the despondent nihilism and the blithe optimism. We owe it to everyone, present and future, to get it right.
- ryanblakeley 4y agoIt's easier to not understand the risks. It takes extra effort to understand AI safety problems and more effort again to solve them. People who blow off AI safety are being blithe and not engaging the arguments.
- verteu 4y agoAs a former Sr ML researcher at Google, I do have a cursory understanding of AI and its risks. But as I mentioned in another comment, the current discussion feels like we're blaming a slot machine's bytecode instead of the casino owners.
- johnsimer 4y agoRight but the eventual discussion is that we need to be able to understand and control the generated bytecode, if possible at all. Eventually smarter than human AIs will exist and we currently do not have real interperability into dumber than human AIs
- pixl97 4y agoWhich then you should realize is the crux of the alignment problem. Humans generally have the same alignment constraints, we need food, shelter, air, water, etc. We can all agree on that. But if you give 10 humans $1 and 1 human $1 billion dollars the human with a billion dollars can quickly ignore and forget about the needs of those with $1. The human alignment problem is not solved! With the potential for all knowing information ingestors with a realtime input system that far outstrips human capabilities, the delta of the problem just becomes worse, particularly because the AI goals are much more likely to be aligned with the billionaires ideas.
- mistrial9 4y agosome serious professionals in tech and elsewhere are driven by internal or external forces to do things professionally that may not align well with larger considerations?
- arrosenberg 4y ago> I can't help but wonder who's pushing this "AI safety" narrative, and why. After watching Facebook evolve the last ~20 years, I think a bit of skepticism is warranted.
- verteu 4y agoIMO, Facebook's problems (body image issues, fake news/clickbait) aren't caused by a genuine inability to control their "AI" algorithms[1]. The root cause is they prioritize profits over user well-being. Framing the issue as "AI safety" feels like blaming a slot machine's bytecode instead of the casino owners. But I appreciate the reply. I'm heartened to learn that people (besides the NYT/WSJ) are genuinely concerned. [1] I'd expect the algorithms are pretty similar to 'show me things like the ones my friends interacted with.' And if they really cared about addiction, they could just disable infinite scroll.
- arrosenberg 4y ago> The root cause is they prioritize profits over user well-being. Framing the issue as "AI safety" feels like blaming a slot machine's bytecode instead of the casino owners. AI will have the same problem, except instead of sociopaths at the helm, it will be an actual computer.
- CatWChainsaw 4y agoSociopaths will still be in charge of the computers though.