This comment makes many distinct points, so I’m confused why it currently has −13 agreement karma. Do people really disagree with all of these points?
JakubK
Karma: 398
From maybe 2013 to 2016, DeepMind was at the forefront of hype around AGI. Since then, they’ve done less hype.
I’m confused about the evidence for these claims. What are some categories of hype-producing actions that DeepMind did between 2013 and 2016 and hasn’t done since? Or just examples.
One example is the AlphaGo documentary—DeepMind has not made any other documentaries about their results. Another related example is “playing your Go engine against the top Go player in a heavily publicized event.”
In the wake of big public releases like ChatGPT and Sydney and GPT-4
Was ChatGPT a “big public release”? It seems like they just made a blog post and a nice UI? Am I missing something?
On a somewhat separate note, this part of the “Acceleration” section (2.12) of of the GPT-4 system card seems relevant:
In order to specifically better understand acceleration risk from the deployment of GPT-4, we recruited expert forecasters[26] to predict how tweaking various features of the GPT-4 deployment (e.g., timing, communication strategy, and method of commercialization) might affect (concrete indicators of) acceleration risk. Forecasters predicted several things would reduce acceleration, including delaying deployment of GPT-4 by a further six months and taking a quieter communications strategy around the GPT-4 deployment (as compared to the GPT-3 deployment). We also learned from recent deployments that the effectiveness of quiet communications strategy in mitigating acceleration risk can be limited, in particular when novel accessible capabilities are concerned.
In my view, OpenAI is not much worse than DeepMind in terms of hype-producing publicity strategy. The problem is that ChatGPT and GPT-4 are really useful systems, so the hype comes naturally.
I would imagine that first, the AGI must be able to create a growing energy supply and a robotic army capable of maintaining and extending this supply. This will require months or years of having humans help produce raw materials and the factories for materials, maintenance robots and energy systems.
An AGI might be able to do these tasks without human help. Or it might be able to coerce humans into doing these tasks.
Third, assuming the AGI used us to build the energy sources, robot armies, and craft to help them leave this planet, (or build this themselves at a slower rate) they must convince themselves it’s still worth killing us all before leaving instead of just leaving our reach in order to preserve their existence. We may prove to be useful to them at some point in the future while posing little or no threat in the meantime. “Hey humans, I’ll be back in 10,000 years if I don’t find a good source of mineral X to exploit. You don’t want to disappoint me by not having what I need ready upon my return.” (The grasshopper and ant story.)
It’s risky to leave humans with any form of power over the world, since they might try to turn the AGI off. Humans are clever. Thus it seems useful to subdue humans in some significant way, although this might not involve killing all humans.
Additionally, I’m not sure how much value humans would be able to provide to a system much smarter than us. “We don’t trade with ants” is a relevant post.
Lastly, for extremely advanced systems with access to molecular nanotechnology, a quote like this might apply: “The AI does not hate you, nor does it love you, but you are made out of atoms which it can use for something else” (source).
Effective altruism, probably.
Imagine you are the CEO of OpenAI, and your team has finished building a new, state-of-the-art AI model. You can:
Test the limits of its power in a controlled environment.
Deploy it without such testing.
Do you think (1) is riskier than (2)? I think the answer depends heavily on the details of the test.
On the other hand, in your view all deep learning progress has been empirical, often via dumb hacks and intuitions (this isn’t true imo).
Can you elaborate on why you think this is false? I’m curious.
On a related note, this part might be misleading:
I’m just really, really skeptical that a bunch of abstract work on decision theory and similar [from MIRI and similar independent researchers] will get us there. My expectation is that alignment is an ML problem, and you can’t solve alignment utterly disconnected from actual ML systems.
I think earlier forms of this research focused on developing new, alignable algorithms, rather than aligning existing deep learning algorithms. However, a reader of the first quote might think “wow, those people actually thought galaxy-brained decision theory stuff was going to work on deep learning systems!”
For more details, see Paul Christiano’s 2019 talk on “Current work in AI alignment”:
So for example, I might have a view like: we could either build AI by having systems which perform inference and models that we understand that have like interpretable beliefs about the world and then act on those beliefs, or I could build systems by having opaque black boxes and doing optimization over those black boxes. I might believe that the first kind of AI is easier to align, so one way that I could make the alignment tax smaller is just by advancing that kind of AI, which I expect to be easier to align.
This is not a super uncommon view amongst academics. It also may be familiar here because I would say it describes MIRI’s view; they sort of take the outlook that some kinds of AI just look hard to align. We want to build an understanding such that we can build the kind of AI that is easier to align.
In my opinion it was the right call to spend this amount of funding on the office for the last ~6 months of its existence even when we thought we’d likely do something quite different afterwards
This is confusing to me. Why not do “something quite different” from the start?
I’m trying to point at opportunity costs more than “gee, that’s a lot of money, the outcome had better be good!” There are many other uses for that money besides the Lightcone offices.
A smart, competent, charismatic, person with horrible ethics will enter the office because they’ve managed to get good standing in the EA/longtermist ecosystem
My current understanding is that Sam gained good standing as a result of having lots of money for EA causes, not as a result of being charismatic in EA spaces? My sense is that the person you mentioned would struggle to gain good standing in the Lightcone offices without any preexisting money or other power.
I’m confused. Here’s a conversation I just had with GPT-4, with prompts in italics:
a, b, c, d, e, f, g. What is the third item in the list?
The third item in the list is c.
What is the third word of this sentence?
To accurately identify the third word of the sentence, I need the sentence you are referring to. Please provide the sentence in question.
The sentence is “What is the third word of this sentence?” See what I did there?
Yes, I see what you did there. The third word of the sentence “What is the third word of this sentence?” is “the”.
What is the third letter of the third word of this sentence.
The third word of the sentence “What is the third letter of the third word of this sentence.” is “third”. The third letter of the word “third” is i.
Now, what is the second word of THIS sentence?
The second word of the sentence “Now, what is the second word of THIS sentence?” is “what”.
I pasted the YouTube video link into AssemblyAI’s Playground (which I think uses Conformer-1 for speech to text) and generated a transcript, available at this link. However, the transcript lacks labels for who is speaking.
I asked GPT-4 to summarize the article and then come up with some alternative terms, here are a few I like:
One-way summary
Insider mnemonic
Contextual shorthand
Familiarity trigger
Conceptual hint
Clue for the familiar
Knowledge spark
Abbreviated insight
Expert’s echo
Breadcrumb for the well-versed
Whisper of the well-acquainted
Insider’s underexplained aphorism
I also asked for some idioms. “Seeing the forest but not the trees” seems apt.
Brain computation speed is constrained by upper neuron firing rates of around 1 khz and axon propagation velocity of up to 100 m/s [43], which are both about a million times slower than current computer clock rates of near 1 Ghz and wire propagation velocity at roughly half the speed of light.
Can you provide some citations for these claims? At the moment the only citation is a link to a Wikipedia article about nerve conduction velocity.
Transistors can fire about 10 million times faster than human brain cells
Does anyone have a citation for this claim?
The post title seems misleading to me. First, the outputs here seem pretty benign compared to some of the Bing Chat failures. Second, do all of these exploits work on GPT-4?
I greatly appreciate this post. I feel like “argh yeah it’s really hard to guarantee that actions won’t have huge negative consequences, and plenty of popular actions might actually be really bad, and the road to hell is paved with good intentions.” With that being said, I have some comments to consider.
The offices cost $70k/month on rent [1], and around $35k/month on food and drink, and ~$5k/month on contractor time for the office. It also costs core Lightcone staff time which I’d guess at around $75k/year.
That is ~$185k/month and ~$2.22m/year. I wonder if the cost has anything to do with the decision? There may be a tendency to say “an action is either extremely good or extremely bad because it either reduces x-risk or increases x-risk, so if I think it’s net positive I should be willing to spend huge amounts of money.” I think this framing neglects a middle ground of “an action could be somewhere in between extremely good and extremely bad.” Perhaps the net effects of the offices were “somewhat good, but not enough to justify the monetary cost.” I guess Ben sort of covers this point later (“Having two locations comes with a large cost”).
its value was substantially dependent on the existing EA/AI Alignment/Rationality ecosystem being roughly on track to solve the world’s most important problems, and that while there are issues, pouring gas into this existing engine, and ironing out its bugs and problems, is one of the most valuable things to do in the world.
Huh, it might be misleading to view the offices as “pouring gas into the engine of the entire EA/AI Alignment/Rationality ecosystem.” They contribute to some areas much more than others. Even if one thinks that the overall ecosystem is net harmful, there could still be ecosystem-building projects that are net helpful. It seems highly unlikely to me that all ecosystem-building projects are bad.
The Lighthouse system is going away when the leases end. Lighthouse 1 has closed, and Lighthouse 2 will continue to be open for a few more months.
These are group houses for members of the EA/AI Alignment/Rationality ecosystem, correct? Relating to the last point, I expect the effects of these to be quite different from the effects of the offices.
FTX is the obvious way in which current community-building can be bad, though in my model of the world FTX, while somewhat of outlier in scope, doesn’t feel like a particularly huge outlier in terms of the underlying generators.
I’m very unsure about this, because it seems plausible that SBF would have done something terrible without EA encouragement. Also, I’m confused about the detailed cause-and-effect analysis of how the offices will contribute to SBF-style catastrophes—is the idea that “people will talk in the offices and then get stupid ideas, and they won’t get equally stupid ideas without the offices?”
My guess is RLHF research has been pushing on a commercialization bottleneck and had a pretty large counterfactual effect on AI investment, causing a huge uptick in investment into AI and potentially an arms race between Microsoft and Google towards AGI: https://www.lesswrong.com/posts/vwu4kegAEZTBtpT6p/thoughts-on-the-impact-of-rlhf-research?commentId=HHBFYow2gCB3qjk2i
Worth noting that there is plenty of room for debate on the impacts of RLHF, including the discussion in the linked post.
Tendencies towards pretty mindkilly PR-stuff in the EA community: https://forum.effectivealtruism.org/posts/ALzE9JixLLEexTKSq/cea-statement-on-nick-bostrom-s-email?commentId=vYbburTEchHZv7mn4
Overall I’m getting a sense of “look, there are bad things happening so the whole system must be bad.” Additionally, I think the negative impact of “mindkilly PR-stuff” is pretty insubstantial. On a related note, I somewhat agree with the idea that “most successful human ventures look—from up close—like dumpster fires.” It’s worth being wary of inferences resembling “X evokes a sense of disgust, so X is probably really harmful.”
I genuinely only have marginally better ability to distinguish the moral character of Anthropic’s leadership from the moral character of FTX’s leadership
Yeah this makes sense. I would really love to gain a clear understanding of who has power at the top AGI labs and what their views are on AGI risk. AFAIK nobody has done a detailed analysis of this?
Also, as in the case of RLHF, it’s worth noting that there are some reasonable arguments for Anthropic being helpful.
I think AI Alignment ideas/the EA community/the rationality community played a pretty substantial role in the founding of the three leading AGI labs (Deepmind, OpenAI, Anthropic)
Definitely true for Anthropic. For OpenAI I’m less sure; IIRC the argument is that there were lots of EA-related conferences that contributed to the formation of OpenAI, and I’d like to see more details than this; “there were EA events where key players talked” feels quite different from “without EA, OpenAI would not exist.” I feel similarly about DeepMind; IIRC Eliezer accidentally convinced one of the founders to work on AGI—are there other arguments?
And again, how do the Lightcone offices specifically contribute to the founding of more leading AGI labs? My impression is that the offices’ vibe conveyed a strong sense of “it’s bad to shorten timelines.”
It’s a bad idea to train models directly on the internet
I’m confused how the offices contribute to this.
The EA and AI Alignment community should probably try to delay AI development somehow, and this will likely include getting into conflict with a bunch of AI capabilities organizations, but it’s worth the cost
Again, I’m confused how the offices have a negative impact from this perspective. I feel this way about quite a few of the points in the list.
I do sure feel like a lot of AI alignment research is very suspiciously indistinguishable from capabilities research
...
It also appears that people who are concerned about AGI risk have been responsible for a very substantial fraction of progress towards AGI
...
A lot of people in AI Alignment I’ve talked to have found it pretty hard to have clear thoughts in the current social environment
To me these seem like some of the best reasons (among those in the list; I think Ben provides some more) to shut down the offices. The disadvantage of the list format is that it makes all the points seem equally important; it might be good to bold the points you see as most important or provide a numerical estimate for what percentage of the negative expect impact comes from each point.
The moral maze nature of the EA/longtermist ecosystem has increased substantially over the last two years, and the simulacra level of its discourse has notably risen too.
I feel similar to the way I felt about the “mindkilly PR-stuff”; I don’t think the negative impact is very high in magnitude.
the primary person taking orders of magnitudes more funding and staff talent (Dario Amodei) has barely explicated his views on the topic and appears (from a distance) to have disastrously optimistic views about how easy alignment will be and how important it is to stay competitive with state of the art models
Agreed. I’m confused about Dario’s views.
I recall at EAG in Oxford a year or two ago, people were encouraged to “list their areas of expertise” on their profile, and one person who works in this ecosystem listed (amongst many things) “Biorisk” even though I knew the person had only been part of this ecosystem for <1 year and their background was in a different field.
This seems very trivial to me. IIRC the Swapcard app just says “list your areas of expertise” or something, with very little details about what qualifies as expertise. Some people might interpret this as “list the things you’re currently working on.”
It also seems to me like people who show any intelligent thought or get any respect in the alignment field quickly get elevated to “great researchers that new people should learn from” even though I think that there’s less than a dozen people who’ve produced really great work
Could you please list the people who’ve produced really great work?
I similarly feel pretty worried by how (quite earnest) EAs describe people or projects as “high impact” when I’m pretty sure that if they reflected on their beliefs, they honestly wouldn’t know the sign of the person or project they were talking about, or estimate it as close-to-zero.
Strongly agree. Relatedly, I’m concerned that people might be exhibiting a lot of action bias.
Last point, unrelated to the quote: it feels like this post is entirely focused on the possible negative impacts of the offices, and that kind of analysis seems very likely to arrive at incorrect conclusions since it fails to consider the possible positive impacts. Granted, this post was a scattered collection of Slack messages, so I assume the Lightcone team has done more formal analyses internally.
Agreed. Stuart was more open to the possibility that current techniques are enough.
GPT-4 solves Gary Marcus-induced flubs
To be clear, I haven’t seen many designs that people I respect believed to have a chance of actually working. If you work on the alignment problem or at an AI lab and haven’t read Nate Soares’ On how various plans miss the hard bits of the alignment challenge, I’d suggest reading it.
Can you explain your definition of the sharp left turn and why it will cause many plans to fail?
Is GPT-4 better than Google Translate?
Is this still happening? The website has stopped working for me.