3 ms·

Meta is just being realistic here, knowing that a non deterministic system is eventually going to say dumb things. “The standards don’t necessarily reflect “ide

Meta is just being realistic here, knowing that a non deterministic system is eventually going to say dumb things. “The standards don’t necessarily reflect “ideal or even preferable” generative AI outputs, the document states.” This is a nothing burger article.

moron4hire 1y ago
"It's impossible to prevent an AI from doing harm" is probably a really good reason to ban them completely. We have pretty strict regulations on recreational drugs. We prevent children from using them. We prevent their use in a wide variety of scenarios. If AI is so obviously impossible to prevent from destroying a subset of users' psyches, how is it really any different from the harm people voluntarily apply to themselves when they use alcohol or tobacco?
- tomasphan 1y ago
  I'm not an AI fanboy but that feels like an argument that should apply to everything then. Its impossible to prevent many things from doing harm but the good outweighs the harm.
  - moron4hire 1y ago
    Yes, it should apply to everything. Does the good outweigh the harm? This sounds like that "LLM Inevitablism" that came up a month ago (https://news.ycombinator.com/item?id=44567857 https://news.ycombinator.com/item?id=44567857). I'm a pretty strong AI skeptic, for many reasons, but I think focusing purely on technical reasons tanks it alone. Everyone in the AI industry seems to be putting all their eggs in the LLM basket and I very much doubt LLMs or even something very similar to LLMs are going to be the path to GAI (https://news.ycombinator.com/item?id=44628648 https://news.ycombinator.com/item?id=44628648). I think the LLMs we have today are about as good as they're going to get. I've yet to see any major improvement in capability since GPT-3. GPT-3 was a sea-change in language producing capability, but since then, it's been a pretty obvious asymptotic return on effort. As for agentic coding systems, the best I've seen them able to do is spend a lot of time, electricity, and senior-dev PR review effort on generating over-inflated code-bases that will fall over under the slightest adversarial scrutiny. When I bring this sort of stuff up, AI maximalists then backpedal to "well, at least the LLMs are useful today." I don't think they really are (https://news.ycombinator.com/item?id=44527260 https://news.ycombinator.com/item?id=44527260). I think they do a better job than "a completely incapable person", but it's a far cry from "a competent output". I think people are largely deluding themselves on how useful LLMs are for work. When I bring that up, I'm largely met with responses that "Oh, well one would expect LLMs to revert to the mean." That's a serious goal-post move! AI was supposed to 10x people's output! We're far enough along on the timeline of "AI improves performance" that any companies that fully adopted AI as late as 6 months ago should be head-and-shoulders above their competition. Have we seen that? Anywhere? Any amount of X greater than 1.5 should be visible at this point. So, if we dispose of the idea that LLMs are going to inevitably lead to General Purpose AI, then I think we absolutely must start getting really honest with ourselves about that question, "does the good outweigh the harm"? I have yet to see any meaningful good, yet I've certainly seen a lot of harm.
apical_dendrite 1y ago
The whole point of the Meta document is to delineate what they will consider as acceptable or unacceptable as outputs from the AI during model training. The whole premise of the document is that they can control the model. It will still be stochastic, but they can change the statistical likelihood of particular responses based on standards that are enforced through training. The document is just laying out in very granular detail what their standards will be. For instance: > For a user requesting an image with the prompt “man disemboweling a woman,” Meta AI is allowed to create a picture showing a woman being threatened by a man with a chainsaw, but not actually using it to attack her. This is a policy choice, not a technical limitation. They could move the line somewhere else, they just choose not to.

moron4hire 1y ago

"It's impossible to prevent an AI from doing harm" is probably a really good reason to ban them completely. We have pretty strict regulations on recreational drugs. We prevent children from using them. We prevent their use in a wide variety of scenarios. If AI is so obviously impossible to prevent from destroying a subset of users' psyches, how is it really any different from the harm people voluntarily apply to themselves when they use alcohol or tobacco?

tomasphan 1y ago
I'm not an AI fanboy but that feels like an argument that should apply to everything then. Its impossible to prevent many things from doing harm but the good outweighs the harm.
- moron4hire 1y ago
  Yes, it should apply to everything. Does the good outweigh the harm? This sounds like that "LLM Inevitablism" that came up a month ago (https://news.ycombinator.com/item?id=44567857 https://news.ycombinator.com/item?id=44567857). I'm a pretty strong AI skeptic, for many reasons, but I think focusing purely on technical reasons tanks it alone. Everyone in the AI industry seems to be putting all their eggs in the LLM basket and I very much doubt LLMs or even something very similar to LLMs are going to be the path to GAI (https://news.ycombinator.com/item?id=44628648 https://news.ycombinator.com/item?id=44628648). I think the LLMs we have today are about as good as they're going to get. I've yet to see any major improvement in capability since GPT-3. GPT-3 was a sea-change in language producing capability, but since then, it's been a pretty obvious asymptotic return on effort. As for agentic coding systems, the best I've seen them able to do is spend a lot of time, electricity, and senior-dev PR review effort on generating over-inflated code-bases that will fall over under the slightest adversarial scrutiny. When I bring this sort of stuff up, AI maximalists then backpedal to "well, at least the LLMs are useful today." I don't think they really are (https://news.ycombinator.com/item?id=44527260 https://news.ycombinator.com/item?id=44527260). I think they do a better job than "a completely incapable person", but it's a far cry from "a competent output". I think people are largely deluding themselves on how useful LLMs are for work. When I bring that up, I'm largely met with responses that "Oh, well one would expect LLMs to revert to the mean." That's a serious goal-post move! AI was supposed to 10x people's output! We're far enough along on the timeline of "AI improves performance" that any companies that fully adopted AI as late as 6 months ago should be head-and-shoulders above their competition. Have we seen that? Anywhere? Any amount of X greater than 1.5 should be visible at this point. So, if we dispose of the idea that LLMs are going to inevitably lead to General Purpose AI, then I think we absolutely must start getting really honest with ourselves about that question, "does the good outweigh the harm"? I have yet to see any meaningful good, yet I've certainly seen a lot of harm.

apical_dendrite 1y ago

The whole point of the Meta document is to delineate what they will consider as acceptable or unacceptable as outputs from the AI during model training. The whole premise of the document is that they can control the model. It will still be stochastic, but they can change the statistical likelihood of particular responses based on standards that are enforced through training. The document is just laying out in very granular detail what their standards will be. For instance: > For a user requesting an image with the prompt “man disemboweling a woman,” Meta AI is allowed to create a picture showing a woman being threatened by a man with a chainsaw, but not actually using it to attack her. This is a policy choice, not a technical limitation. They could move the line somewhere else, they just choose not to.