Posts
Comments
Leopold Aschenbrenner is starting a cross between a hedge fund and a think tank for AGI. I have read only the sections of Situational Awareness most relevant to this project, and I don't feel nearly like I understand all the implications, so I could end up being quite wrong. Indeed, I’ve already updated towards a better and more nuanced understanding of Aschenbrenner's points, in ways that have made me less concerned than I was to begin with. But I want to say publicly that the hedge fund idea makes me nervous.
Before I give my reasons, I want to say that it seems likely most of the relevant impact comes not from the hedge fund but from the influence the ideas from Situational Awareness have on policymakers and various governments, as well as the influence and power Aschenbrenner and any cohort he builds wield. This influence may come from this hedge fund or be entirely incidental to it. I mostly do not address this here, but it does make all of the below less important.
I also believe that some (though not all) of my concerns about the hedge fund are based on specific disagreements with Aschenbrenner’s views. I discuss some of those below, but a full rebuttal this is not (and many of the points of disagreement I don’t yet feel confident in my view on). There is still plenty to do to hash out the actual empirical questions at hand.
Why I am nervous
A hedge fund investing in AI related investments means Aschenbrenner and his investors will gain financially from more and accelerated AGI progress. This seems to me to be one of the most important dynamics (excluding the points about influence above). That creates an incentive to create more AGI progress, even at the cost of safety, which seems quite concerning. I will say that Leopold has a good track record here around turning down money in not signing an NDA at Open AI despite loss of equity.
Aschenbrenner expresses strong support for the liberal democratic world to maintain a lead on AI advancement, and ensure that China does not reach an AI-based decisive military advantage over the United States[1]. The hedge fund, then, presumably aims to both support the goal of maintaining an AI lead over China and profit off of it. In my current view, this approach increases race dynamics and increases the risks of the worst outcomes (though my view on this has softened somewhat since my first draft, for reasons similar to what Zvi clarifies here[2]).
I especially think that it risks unnecessary competition when cooperation - the best outcome - could still be possible. It seems notable, for example, that no Chinese version of the Situational Awareness piece has come to my attention; going first in such a game both ensures you are first and that the game is played at all.
It’s also important that the investors (e.g. Patrick Collison) appear to be more focused on economic and technological development, and less concerned about risks from AI. The incentives of this hedge fund are therefore likely to point towards progress and away from slowing down for safety reasons.
There are other potential lines of thought here I have not yet fleshed out including:
- The value of aiming to orient the US government and military attention to AGI (seems like a huge move with unclear sign)
- The degree to which this move is unilateralist on Aschenbrenner’s part
- How much money could be made and how much power the relevant people (e.g. Aschenbrenner and his investors) will have through investment and being connected to important decisions.
- If a lot of money and/or power could be acquired, especially over AGI development, then there’s a healthy default skepticism I think should be applied to their actions and decision-making.
- Specifics about Aschenbrenner himself. Different people in the same role would take very different actions, so specifics about his views, ways of thinking, and profile of strengths and weaknesses may be relevant.
Ways that the hedge fund could in fact be a good idea:
EA and AI causes could really use funder diversification. If Aschenbrenner intends to use the money he makes to support these issues, that could be very valuable (though I’ve certainly become somewhat more concerned with moonshot “become a billionaire to save the world” plans than I used to be).
The hedge fund could position Aschenbrenner to have a deep understanding of and connections within the AI landscape, making the think tank outputs very good, and causing important future decisions to be made better.
Aschenbrenner of course could be right about the value of the US government’s involvement, maintaining a US lead, and the importance of avoiding Chinese military supremacy over the US. In that case, him achieving his goals would of course be good. Cruxes include the likelihood of international cooperation, the possibility of international bans, probability of catastrophic outcomes from AI and the likelihood of “muddling through” on alignment.
I’m interested in hearing takes, ways I could be wrong, fleshing out of my arguments, or any other thoughts people have relevant to this. Happy to have private chats in DMs to discuss as well.
- ^
To be clear, Aschenbrenner wants that lead to exist to avoid a tight race in which safety and caution are thrown to the winds. If we can achieve that lead primarily through infosecurity (something he emphasizes), then added risks are low; but I think the views expressed in Situational Awareness also imply the importance of staying technologically ahead of China as their AI research improves. This comes with precisely the risks of creating and accelerating a race of this nature.
Additionally, when I read his description of the importance of even a two month lead, it implied to me that if the longer, more comfortable lead is lost, there will be strong reasons for the US to advance quickly so as to avoid China reaching superintelligence and subsequent military dominance first (which doesn’t mean he thinks we should actually do this if the time came). This seems to fairly explicitly describe the tight race scenario. I don’t think Aschenbrenner believes this would be a good situation to be in, but nonetheless thinks that’s what the true picture is.
- ^
From Zvi’s post: “He confirms he very much is NOT saying this:
The race to ASI is all that matters.
The race is inevitable.
We might lose.
We have to win.
Trying to win won’t mean all of humanity loses.
Therefore, we should do everything in our power to win.
I strongly disagree with this first argument. But so does Leopold.
Instead, he is saying something more like this:
ASI, how it is built and what we do with it, will be all that matters.
ASI is inevitable.
A close race to ASI between nations or labs almost certainly ends badly.
Our rivals getting to ASI first would also be very bad.
Along the way we by default face proliferation and WMDs, potential descent into chaos.
The only way to avoid a race is (at least soft) nationalization of the ASI effort.
With proper USG-level cybersecurity we can then maintain our lead.
We can then use that lead to ensure a margin of safety during the super risky and scary transition to superintelligence, and to negotiate from a position of strength.”
It sounds from this back and forth like we should assume that Anthropic leadership who left from OAI (so Dario and Daniela Amodei, Jack Clark, Sam McCandlish, others?) are still under NDA because it was probably mutual. Does that sound right to others?
Oh! I think you're right, thanks!
Also relevant: AI companies aren't really using external evaluators
I feel pretty sympathetic to the desire not to do things by text; I suspect you get much more practiced and checked over answers that way.
which privacy skills you are able to execute.
This link goes to a private google doc, just fyi.
This is great!
I really like this about slack:
- If you aren’t maintaining this, err on the side of cultivating this rather than doing high-risk / high-reward investments that might leave you emotionally or financially screwed.
- (or, if you do those things, be aware I may not help you if it fails. I am much more excited about helping people that don’t go out of their way to create crises)
Seems like a good norm and piece of advice.
I'm confused how much I should care whether an impact assessment is commissioned by some organization. The main thing I generally look for is whether the assessment / investigation is independent. The argument is that because AISC is paying for it, that will influence the assessors?
I have not read most of what there is to read here, just jumping in on "illegal drugs" ---> ADHD meds. Chloe's comment spoke to weed as the illegal drug on her mind.
AI has immense potential, but also immense risks. AI might be misused by China, or get of control. We should balance the needs for innovation and safety." I wouldn't call this lying (though I agree it can have misleading effects, see Issue 1).
Not sure where this slots in, but there's also a sense in which this contains a missing positive mood about how unbelievably good (aligned) AI could or will be, and how much we're losing by not having it earlier.
Thanks!
Interesting how many of these are "democracy / citizenry-involvement" oriented. Strongly agree with 18 (whistleblower protection) and 38 (simulate cyber attacks).
20 (good internal culture), 27 (technical AI people on boards) and 29 (three lines of defense) sound good to me, I'm excited about 31 if mandatory interpretability standards exist.
42 (on sentience) seems pretty important but I don't know what it would mean.
The top 6 of the ones in the paper (the ones I think got >90% somewhat or strongly agree, listed below), seem pretty similar to me - are there important reasons people might support one over another?
- Pre-deployment risk assessments
- Evaluations of dangerous capabilities
- Third-party model audits
- Red teaming
- Pre-training risk assessments
- Pausing training of dangerous models
Curious if you have any updates!
Chat GPT gives some interesting analysis when asked, though I think not amazingly accurate. (The sentence I gave it, from here is a weird example, though)
Does it say anything about AI risk that is about the real risks? (Have not clicked the links, the text above did not indicate to me one way or another).
This is great, and speaks to my experience as well. I have my own frames that map onto some of this but don't hit some of the things you've hit and vice versa. Thanks for writing!
Is this something Stampy would want to help with?
https://www.lesswrong.com/posts/WXvt8bxYnwBYpy9oT/the-main-sources-of-ai-risk
I think that incentivizes self-deception on probabilities. Also, P <10^-10 are pretty unusual, so I'd expect that to cause very little to happen.
Thanks!
When you say "They do, however, have the potential to form simulacra that are themselves optimizers, such as GPT modelling humans (with pretty low fidelity right now) when making predictions"
do you mean things like "write like Ernest Hemingway"?
Is it true that current image systems like stable diffusion are non-optimizers? How should that change our reasoning about how likely it is that systems become optimizers? How much of a crux is "optimizeriness" for people?
Why do people keep saying we should maximize log(odds) instead of odds? Isn't each 1% of survival equally valuable?
In addition to Daniel's point, I think an important piece is probabilistic thinking - the AGI will execute not based on what will happen but on what it expects to happen. What probability is acceptable? If none, it should do nothing.
Have you written about your update to slow takeoff?
Nice! Added these to the wiki on calibration: https://www.lesswrong.com/tag/calibration
Oh, whoops. I took from this later tweet in the thread that they were talking.
After years of tinkering and incremental progress, AIs can now play Diplomacy as well as human experts.[6]
Maybe this happened in 2022: https://twitter.com/polynoamial/status/1580185706735218689
Let me know if you have a cheerful price for this!
Here's the git! https://github.com/SonOfLilit/calibrate?fbclid=IwAR2vBZ8IWfMgHTPla0CbohCUIqmrMUl-XEcYIWhKUrJ4ZRfH2Eg7Z7Zf1J4
I will talk to the developer about it being open source - I think that was both of our ideals.
Do you know how to do this kind of thing? I'd be happy to pay you for your time.
This seems interesting to me but I can't yet latch onto it. Can you give examples of secrets being one or the other?
Are you distinguishing between "secrets where the existence of the secret is a big part of the secret" and "secrets where it's not"?
One of my feature requests! Just hard to do.
Why would they be jokes?
Don't know what you mean in the latter sentence.
Conversational moves in EA / Rationality that I like for epistemics
- “So you are saying that”
- “But I’d change my mind if”
- “But I’m open to push back here”
- “I’m curious for your take here”
- “My model says”
- “My current understanding is…”
- “...I think this because…”
- “...but I’m uncertain about…”
- “What could we bet on?”
- “Can you lay out your model for me?”
- “This is a butterfly idea”
- “Let’s do a babble”
- “I want to gesture at something / I think this gestures at something true”
This is why less wrong needs the full suite of emoji reacts.
Title changed!
I meant signposting to indicate things like saying "here's a place where I have more to say but not in this context" etc, during for instance a conversation, so I'm truthfully saying that there's more to the story.
Yeah, I think "intentionally causing others to update in the wrong direction" and "leaving them with their priors" end up pretty similar (if you don't make strong distinctions between action and omission, which I think this test at least partially rests on) if you have a good model of their priors (which I think is potentially the hardest part here).
Kind is one of the four adjectives in your description of Iron Hufflepuff.
Hm, Keltham has a lot of good qualities here, but kind doesn't seem among them.
Sounds scary, but thank you for the model of what's actually going on!
Oh woah! Thanks for linking.
True! 65 Watts! That would really be something.
Unfortunately I'm not seeing anything close to that on the Amazon UK site :/
Might be bad search skills, though.
Your link's lightbulbs have a bayonet style, not the E27 threading :) Thanks for the other link! Amazon says currently unavailable.
ETA: Found some, will add to post
Tried to buy those, didn't have any luck finding ones that fit nicely into my sockets! (An embarassing mistake I didn't describe in detail is buying corn bulbs that turned out to be...mini?) If you have an amazon UK link for ones with E27 threading, that would be awesome.
ETA: Having looked, it looks like not all corn bulbs are brighter than the ones I have, though I have now found 2000 lumen ones. I don't know if corn bulbs are still better if they have lower lumens. I would guess not?
ETA 2: The link above does have E27 if you click through the multiple listings in the same link, wasn't obvious to me at first, thanks!
I saw people discussing forecasting success of this on twitter and people were saying that the intelligence agencies actually called this right. Does anyone know an easy link to what those agencies were saying?
Context: https://twitter.com/ClayGraubard/status/1496699988801433602?s=20&t=mQ8sAzMRppI8Pr44O38M3w
https://twitter.com/ClayGraubard/status/1496866236973658112?s=20&t=mQ8sAzMRppI8Pr44O38M3w
Nice! Welcome!
I definitely find it helpful to be surrounded by people who will do this for me and help me cultivate a habit of it over time. The case for it being very impactful is if people do a one-time thing, like apply for something or put themselves in the running for something that they otherwise wouldn't have that makes a big difference. The ones that are about accountability (Can I remind you about that in a week?) also are sort of a conscientiousness loan, which can be cheap since it can be easier to check in on other people than to do it for yourself.
It is definitely important to have sense of who you're talking to and what they need (law of equal and opposite advice). For what it's worth, 5-10 and 13 are aimed to be disproportionately helpful for people who have trouble doing things (depending on the reason).