When people talk about AI futures, the picture is usually centralized. Either a single aligned superintelligence replaces society with something utopian and post-scarcity, or an unaligned one destroys us, or maybe a malicious human actor uses a powerful system to cause world-ending harm. Those futures might be possible. However there’s...
Summary: The collection of a massive first-person video dataset for use in large model pretraining is technologically & economically feasible with modest resources. Such a dataset might enable high-fidelity imitation of human behavior, with positive implications for the capabilities and safety of trained models. Such models may also enable a...
Yeah! I think we share a similar vision, at least at the early stages where rogue AIs proliferate. I think the largest difference between the vision that Alvin proposed and what I wrote above is the consequences of proliferation - he assumes that various actors unite to put resources behind an aligned super intelligence that dominates all other rogue AI systems. In my vision this never happens and things stay permanently chaotic, leading to a shift in priorities from "fighting" AI destruction to being robust to it. The vision of the highly-robust bunker-state of the far future was the big thing I wanted to emphasize.
When people talk about AI futures, the picture is usually centralized. Either a single aligned superintelligence replaces society with something utopian and post-scarcity, or an unaligned one destroys us, or maybe a malicious human actor uses a powerful system to cause world-ending harm.
Those futures might be possible. However there’s another shape of the future I keep coming back to, which I almost never see described. The adjectives I’d use are: decentralized, diverse, and durable. I don't think this future is necessarily good, but I do think it’s worth planning for.
I don’t think we’re on extremely short timelines (e.g. AGI before 2030). I expect a small slowdown in capabilties... (read 556 more words →)
Hey! I checked out your site and I found it really interesting. In terms of utility & implementation simplicity, I think focusing purely on digital prescence makes some sense. At the limit, if you were able to accurately model the continuation of a sequence of all inputs + the screen state + audio output of a computer, I think this might end up being a surprisingly powerful "general-purpose-human-simulator" (given how much the average person lives their lives through their screen). However, one thing that draws me to lifelogging-style video is the promise to capture the more subtle and poetic parts of being human (listening to the other conversations in a restaurant, laughing... (read more)
Summary: The collection of a massive first-person video dataset for use in large model pretraining is technologically & economically feasible with modest resources. Such a dataset might enable high-fidelity imitation of human behavior, with positive implications for the capabilities and safety of trained models. Such models may also enable a type of “digital immortality” and provide a compelling alternative to cryonics.
Table of Contents
First Person Behavior Cloning
Implications for Capabilities
Implications for Alignment
Digital Immortality
Implementation
Hardware Options
Economic Feasibility
Privacy & Legal Concerns
First Person Behavior Cloning
Current frontier AI systems rely on pretraining via next-token-prediction on web-scale (primarily text) datasets. At the limit, this training objective is sufficient to enable human-equivalent artificial intelligence (to perfectly predict the next words of... (read 1979 more words →)
I'd pay a lot of money for an app like this. I wonder if recent development's like Google's MedicalLLM could come into play here, where all your symptoms are logged and then expert knowledge / a thorough review of medical literature is done automatically to recommend potential solutions
I've been interested in this topic in the past, especially from the economics / game theoretic perspective. There's one journal I know of that explores this topic that might be worth looking into:
The Journal of Mechanism and Institution Design
http://www.mechanism-design.org/
Mechanism design in this context being the kind of inverse of game theory that starts with a desired outcome and designs a system that produces that outcome, as opposed to the traditional approach of game theory where you start with a system and figuring out the outcome. More info: https://en.wikipedia.org/wiki/Mechanism_design
I also think this topic of ideal governance & mechanism design has a lot of overlap with the field of cryptoeconomics, or the economic analysis... (read more)
First thing that comes to my mind is the chess masters that would stream their practice sessions / teaching sessions via twitch. I watched a few of these and I was surprised how close the experience came to being taught something one-on-one. Even though I wasn't the one being tutored, the types of questions that the pupil asked were similar to the ones I was thinking. I wonder if that would be useful in a professional context? I could certainly see it being useful in a computer security context, like livestreaming a CTF competition. And probably that "learn by observing others learning" approach would be useful in other contexts too
I think we should also take into account the value of English spellings that maintain common forms with other languages, even at the expense of being phonetic.
For instance, to a speaker of French or Spanish, the English word "diversification" would certainly seem less alien than a hypothetical respelling as "daiversifikeishun". Does having a common (or very similar) cross-language orthography for latinate words offer more advantages than the benefits of phonetic spellings? I'm not sure, but it should certainly be part of the discussion.
I suspect the benefit is greatest in technical fields with lots of latin-derived vocabulary (i.e. health & biology). Would international scientific cooperation become more difficult if French and Spanish speakers had to relearn spellings for words like "capillaries" [kapileirīs] or "canine" [keinain] whereas the original english was almost identical to the spelling in their native language.
Really interesting post, and I think proper environment creation is one of, if not the most important question when it comes to the RL-based path to AGI.
You made a point that, contrary to the expectations of some, environments like Go or Starcraft are not sufficient to create the type of flexible, adaptive AGI that we're looking for. I wonder if success in creating such AGI is dependent primarily on the complexity of the environment? That is, even though environments like Starcraft are quite complex and require some of the abstract reasoning we'd expect of AGI, the actual complexity isn't anywhere close to the complexity of the real world in which we want... (read more)
Yeah I share a similar intuition. It seems to me that the two steady states are either strong restrictions by some dominant force, or else proliferation of superintelligence into many distinct and diverse entities, all vying for their own interests