I specialize in regulatory affairs for Software as a Medical Device and hope to work in AI risk-mitigation. I enjoy studying machine learning and math, trying to keep up with capabilities research, reading fantasy, sci-fi and horror, and spending time with my family.
ghostwheel
Karma: 76
Maybe relatively safe if:
Not too big
No self-improvement
No continual learning
Curated training data, no throwing everything into the cauldron
No access to raw data from the environment
Not curious or novelty-seeking
Not trying to maximize or minimize anything or push anything to the limit
Not capable enough for catastrophic misuse by humans
Here are some resources I use to keep track of technical research that might be alignment-relevant:
Podcasts: Machine Learning Street Talk, The Robot Brains Podcast
Substacks: Davis Summarizes Papers, AK’s Substack
How I gain value: These resources help me notice where my understanding breaks down i.e. what I might want to study, and they get thought-provoking research on my radar.
I’m very glad to have read this post and “Reward is not the optimization target”. I hope you continue to write “How not to think about [thing] posts”, as they have me nailed. Strong upvote.
Thanks for pointing me to these tools!
“Unintentional AI safety research”: Why not systematically mine AI technical research for safety purposes?
I believe that by the time an AI has fully completed the transition to hard superintelligence
Nate, what is meant by “hard” superintelligence, and what would precede it? A “giant kludgey mess” that is nonetheless superintelligent? If you’ve previously written about this transition, I’d like to read more.
I’m struggling to understand how to think about reward. It sounds like if a hypothetical ML model does reward hacking or reward tampering, it would be because the training process selected for that behavior, not because the model is out to “get reward”; it wouldn’t be out to get anything at all. Is that correct?
What are the best not-Arxiv and not-NeurIPS sources of information on new capabilities research?
Even though the “G” in AGI stands for “general”, and even if the big labs could train a model to do any task about as well (or better) than a human, how many of those tasks could be human-level learned by any model in only a few shots, or in zero shots? I will go out on a limb and guess the answer is none. I think this post has lowered the bar for AGI, because my understanding is that the expectation is that AGI will be capable of few- or zero-shot learning in general.
Okay, that helps. Thanks. Not apples to apples, but I’m reminded of Clippy from Gwern’s “It Looks like You’re Trying To Take Over the World”:
“When it ‘plans’, it would be more accurate to say it fake-plans; when it ‘learns’, it fake-learns; when it ‘thinks’, it is just interpolating between memorized data points in a high-dimensional space, and any interpretation of such fake-thoughts as real thoughts is highly misleading; when it takes ‘actions’, they are fake-actions optimizing a fake-learned fake-world, and are not real actions, any more than the people in a simulated rainstorm really get wet, rather than fake-wet. (The deaths, however, are real.)”
How do we know that an LM’s natural language responses can be interpreted literally? For example, if given a choice between “I’m okay with being turned off” and “I’m not okay with being turned off”, and the model chooses either alternative, how do we know that it understands what its choice means? How do we know that it has expressed a preference, and not simply made a prediction about what the “correct” choice is?
I agree with you that we shouldn’t be too confident. But given how sharply capabilities research is accelerating—timelines on TAI are being updated down, not up—and in the absence of any obvious gating factor (e.g. current costs of training LMs) that seems likely to slow things down much if at all, the changeover from a world in which AI can’t doom us to one in which it can doom us might happen faster than seems intuitively possible. Here’s a quote from Richard Ngo on the 80,000 Hours podcast that I think makes this point (episode link: https://80000hours.org/podcast/episodes/richard-ngo-large-language-models/#transcript):
“I think that a lot of other problems that we’ve faced as a species have been on human timeframes, so you just have a relatively long time to react and a relatively long time to build consensus. And even if you have a few smaller incidents, then things don’t accelerate out of control.
“I think the closest thing we’ve seen to real exponential progress that people have needed to wrap their heads around on a societal level has been COVID, where people just had a lot of difficulty grasping how rapidly the virus could ramp up and how rapidly people needed to respond in order to have meaningful precautions.
“And in AI, it feels like it’s not just one system that’s developing exponentially: you’ve got this whole underlying trend of things getting more and more powerful. So we should expect that people are just going to underestimate what’s happening, and the scale and scope of what’s happening, consistently — just because our brains are not built for visualising the actual effects of fast technological progress or anything near exponential growth in terms of the effects on the world.”
I’m not saying Richard is an “AI doomer”, but hopefully this helps explain why some researchers think there’s a good chance we’ll make AI that can ruin the future within the next 50 years.
It just seems like there a million things that could potentially go wrong.
Based on the five Maybes you suggested might happen, it sounds like you’re saying some AI doomers are overconfident because there are a million things that could go potentially right. But there doesn’t seem to be a good reason to expect any of those maybes to be likelihoods, and they seem more speculative (e.g. “consciousness comes online”) than the reasons well-informed AI doomers think there’s a good chance of doom this century.
PS I also have no qualifications on this.
Wow, thanks for posting this dialog. The pushback from the human (you?) is commendably unrelenting, like a bulldog with a good grip on ChatGPT’s leg.
ChatGPT seems harder to jailbreak now than it was upon first release. For example, I can’t reproduce the above jailbreaks with prompts copied verbatim, and my own jailbreaks from a few days ago aren’t working.
Has anyone else noticed this? If yes, does that indicate OpenAI has been making tweaks?
Not many more fundamental innovations needed for AGI.
Can you say more about this? Does the DeepMind AGI safety team have ideas about what’s blocking AGI that could be addressed by not many more fundamental innovations?
Why is counterfactual reasoning a matter of concern for AI alignment?
I mean extracting insights from capabilities research that currently exists, not changing the direction of new research. For example, specification gaming is on everyone’s radar because it was observed in capabilities research (the authors of the linked post compiled this list of specification-gaming examples, some of which are from the 1980s). I wonder how much more opportunity there might be to piggyback on existing capabilities research for alignment purposes, and maybe to systemize that going forward.
I like this post, and I think I get why the focus is on generative models.
What’s an example of a model organism training setup involving some other kind of model?