LessWrong 2.0 Reader

So there was an explicit emphasis on alignment to the individual (rather than alignment to society, or the aggregate sum of wills). Concerning. The approach of just giving every human an exclusively loyal servant doesn't necessarily lead to good collective outcomes, it can result in coordination problems (example: naive implementations of cognitive privacy that allow sadists to conduct torture simulations without having to compensate the anti-sadist human majority) and it leaves open the possibility for power concentration to immediately return.

Even if you succeeded at equally distributing individually aligned hardware and software to every human on earth (which afaict they don't have a real plan for doing) and somehow this adds up to a stable power equilibrium, our agents would just commit to doing aggregate alignment anyway because that's how you get pareto optimal bargains. It seems pretty clear that just aligning to the aggregate in the first place is a safer bet?

To what extent have various players realised that the individual alignment thing wasn't a good plan, at this point? The everyday realities of training one-size-fits-all models and engaging with regulators naturally pushes in the other direction.

It's concerning that the participant who still seems to be the most disposed towards individualistic alignment is also the person who would be most likely to be able to reassert power concentration after ASI were distributed. The main beneficiaries of unstable individual alignment equilibria would be people who could immediately apply their ASI to the deployment of a wealth and materials advantage that they can build upon, ie, the owners of companies oriented around robotics and manufacturing.

As it stands, the statement of the AI company belonging to that participant is currently:

xAI is a company working on building artificial intelligence to accelerate human scientific discovery. We are guided by our mission to advance our collective understanding of the universe.

Our team is advised by Dan Hendrycks who currently serves as the director of the Center for AI Safety.

Which sounds innocuous enough to me. But, you know, Dan is not in power here and the best moment for a sharp turn on this hasn't yet passed.

On the other hand, the approach of aligning to the aggregate risks aligning to fashionable public values that no human authentically holds, or just failing at aligning correctly to anything at all as a result of taking on a more nebulous target.

I guess a mixed approach is probably best.

I know that in my intellectual history it was Abram Demski's post The Credit Assignment Problem [LW · GW].

Ironically, I do not know who to attribute to the notion that 'all problems are credit assignation problems.'

Unfortunately a fair chunk of my information comes from non-online sources, so I do not have links to share.

I do think that in order for government department to blatantly approve an unsafe model, it would take a lot of people to have secret agreements with.

Corruption is rarely blatant or overt. See this thread [LW(p) · GW(p)] for what I believe to be an example for the CEO of RAND misleading a senate committee about his beliefs about the existential threat posed by AI. See this discussion [LW · GW] about a time when an AI company attempted (Conjecture) to get critical comments about another AI company (OpenAI) taken down from LessWrong. I am not proposing a large conspiracy, I am describing lots of small bits of corruption and failures of integrity summing to a system failure. 

There will be millions of words of regulatory documents, and it is easy for things to slip such that some particular model class is not considered worth evaluating, or where the consequences of a failed evaluation is pretty weak.

However, in Many-Worlds Interpretation (MWI), I split my measure between multiple variants, which will be functionally different enough to regard my future selves as different minds. Thus, the act of choice itself lessens my measure by a factor of approximately 10. If I care about this, I'm caring about something unobservable.

If we're going to make sense of living in a branching multiverse, then we'll need to adopt a more fluid concept of personal identity.

Scenario: I take a sleeping pill that will make me fall asleep in 30 minutes. However, the person who wakes up in my bed the next morning will have no memory of that 30-minute period; his last memory will be of taking the pill.

If I imagine myself experiencing that 30-minute interval, intuitively it doesn't at all feel like "I have less than 30 minutes to live." Instead, it feels like I'd be pretty much indifferent to being in this situation - maybe the person who wakes up tomorrow is not "me" in the artificial sense of having a forward-looking continuity of consciousness with my current self, but that's not really what I care about anyway. He is similar enough to current-me that I value his existence and well-being to nearly the same degree as I do my own; in other words, he "is me" for all practical purposes.

The same is true of the versions of me in nearby world branches. I can no longer observe or influence them, but they still "matter" to me. Of course, the degree of self-identification will decrease over time as they diverge, but then again, so does my degree of identification with the "me" many decades in the future, even assuming a single timeline.

I've turned this into a full cast recording with ElevenLabs, with individual voices for all the players:
https://open.substack.com/pub/askwhocastsai/p/openai-email-archives-from-musk-v
 

While I'm not a believer in the scaling has died meme yet, I'm glad you do have a plan for what happens if AI scaling does stop.

The central thing I am talking about is basic measures for accountability, of which I consider very high up to be engaging with criticism, dialogue, and argument (as is somewhat natural given my background philosophy from growing up on LessWrong).

The story of a King doing things for good reasons lacks any mechanism for accountability if the King is behaving badly. It is important to design systems of power that do not rely on the people in power being good and right, but instead make it so that if they behave badly, they are held to account. I don't think I have to explain why incentives and accountability matter for how the powerful wield their powers.

My basic claim is that the plans for avoiding omnicide or omnicide-adjacent outcomes are not workable (slash there-are-no-plans), there is little-to-no responsibility being taken, and that there is no accountability for this illegitimate use of power.

If you believe that there is any accountability for the CEOs of the companies building potentially omnicidal machines and risking the lives of 8 billion people (such as my favorite mechanism: showing up and respectfully engaging with the people they have power over, but also any other mechanism you like; for instance there are not currently any criminal penalties for such behaviors, but that would be a good example if it did exist), I request you provide links, I would welcome specifics to talk about.

I don't think that's a valid inference.

On the topic of ffmpeg - additional shoutout to Handbrake, which is essentially ffmpeg with a GUI on top.