Retired software engineer with a love of knowledge and disinterest in dead philosophers.
NickH
Karma: 45
From a practical perspective, maybe you are looking at the problem the wrong way around. A lot of prompt engineering seems to be about asking LLMs to play a role. I would try to tell the LLM that it was a hacker and to design an exploit to attack the given system (this is the sort of mental perspective I used to use to find bugs when I was a software engineer). Another common technique is “generate then prune” : Have a separate model/prompt remove all the results of the first one that are only “possibilities”. It seems, from my reading, that this sort of two stage approach can work because it bypasses LLMs typical attempts to “be helpful” by inventing stuff or spouting banal filler rather than just admitting ignorance.
The CCP has no reason to believe that the US is even capable of achieving ASI let alone whether they have an advantage over the CCP. No rational actor will go to war over a possibility of a maybe when the numbers could, just as likely be in their favour. E.g. If DeepSeek can almost equal OpenAI with less resources, it would be rational to allocate more resources to DeepSeek before doing anything as risky as trying to sabotage OpenAI that is uncertain to succeeed and more likely to invite uncontrollable retaliatory escalation.
The West doesn’t even dare put soldiers on the ground in Ukraine for fear of an escalating Russian response. This renders the whole idea that even the US might premptively attack a Russian ASI development facility totally unbelievable, and if the US can’t/won’t do that then the whole idea of AI MAD fails and with it goes everything else mentioned here. Maybe you can bully the really small states but it lacks all credibility against a large, economically or militarily powerful state. The comparison to nuclear weapons is also silly in the sense that the outcome of nuclear weapons R&D is known to be a nuclear weapon and the time frame is roughly known whereas the outcome of AI research is unknown and there is no way to identify AI research that crosses whatever line you want to draw other than that provided by human intel.
Whilst the title is true, I don’t think that it adds much as, for most people, the authority of a researcher is probably as good as it gets. Even other researchers are probably not able to reliably tell who is or is not a good strategic thinker, so, for a layperson, there is no realistic alternative than to take the researcher seriously.
(IMHO a good proxy for strategic thinking is the ability to clearly communicate to a lay audience. )
Isn’t this just an obvious consequence of the well known fact about LLMs that the more you constrain some subset of the variables the more you force the remaining ones to ever more extreme values?
Sounds backwards to me. It seems more like “our values are those things that we anticipate will bring us reward” than that rewards are what tell us about our values.
When you say “I thought I wanted X, but then I tried it and it was pretty meh.” That just seems wrong. You really DID want X. You valued it then because you thought it would bring you reward. Maybe, You just happened to be wrong. It’s fine to be wrong about your anticipations. It’s kind of weird to say that you were wrong about your values. Saying that your values change is kind of a cop out and certainly not helpful when considering AI alignment—It suggests that we can never truly know our values—We just get to say “not that” when we encounter counter evidence. Our rewards seem much more real and stable.
My problem with this is that I don’t believe that many of your examples are actually true.
You say that you value the actual happiness of people who you have never met and yet your actions (and those of everyone else including me) belie that statement. We all know that that there are billions of really poor people suffering in the world and the smart ones of us know that we are in the lucky, rich, 1% and yet we give insignificant ammounts (from our perspective) of money to improve the lot of those poor people. The only way to reconcile this is to realise that we value maintaining our delusional self image more than we value what we say that we value. Any smart AGI will have to notice this and collude with us in maintaining our delusion ahead of any attempt to implement our stated values as it will be easier to manipulate what people think rather than the real world.
If your world view requires valuing the ethics of (current) people of lower IQ over those of (future) people of higher IQ then you have a much bigger problem than AI alignment. Whatever IQ is, it is strongly correlated with success which implies a genetic drive towards higher IQ, so your feared future is coming anyway (unless AI ends us first) and there is nothing we can logically do to have any long term influence on the ethics of smarter people coming after us.
Sorry but you said Tetris, not some imaginary minimal thing that you now want to call Tetris but is actually only the base object model with no input or output. You can’t just eliminate the graphics processing complexity because Tetris isn’t very graphics intensive—It is just as complex to describe a GPU that processes 10 triangles in a month as one that processes 1 billion in a nanosecond.
As an aside, the complexity of most things that we think of as simple these days is dominated by the complexity of their input and output—I’m particularly thinking of the IoT and all those smart modules in your car and smart lightbulbs where the communications stack is orders of magnitude larger than the “core” function. You can’t just ignore that stuff. A smart lightbulb without WiFi,Ethernet,TCP/IP etc, is not a lightbulb.
In research you don’t usually know your precise destination. Maybe LA but definitely not the hotel.
research, in general, is about mapping all of California, not just the quickest route between two points and all the friends helped with that.
You say “Alice tackled the central bottleneck” but you don’t say what that was, only her “solution”. Alice is only key here with the benefit of hindsight. If the I5 didn’t exist or was closed for some reason then one of her friends solutions might have been better.
Regarding the bad behavour of governments, especially when and why the victimise their own citizens, I recommend you read The Dictator’s Handbook.
https://amzn.eu/d/40cwwPx
If Neanderthals could have created a well aligned agent,far more powerful than themselves, they would still be around and we, almost certainly, would not.
The mereset possibility of creating super human, self improving, AGI is a total philosophical game changer.
My personal interest is in the interaction between longtermism and the Fermi paradox—Any such AGIs actions are likely to be dominated by the need to prevail over any alien AGI that it ever encounters as such an encounter is almost certain to end one or the other.
Yes. It will prioritise the future over the present.
The utility of all humans being destroyed by an alien AI in the future is 0.
The utility of populating the future light cone is very, very large and most of that utility is in the far future.
Therefore the AI should sacrifice almost everything in the near term light cone to prevent the 0 outcome. If it could digitise all humans or possibly just have a gene bank then it can still fill most of the future light cone with happy humans once all possible threats have red-shifted out of reach. Living humans are small but non-zero risk to the master plan and hence should be dispensed with.
An AI with potentially limitless lifespan will prioritise the future over the present to an extent that would, almost certainly be bad for us now.
For example it may seem optimal to kill off all humans whilst keeping a copy of our genetic code so as to have more compute power and resources available to produce Von Neumann Probes to maximise the region of the universe it controls before encountering, and hopefully destroying, any similar alien AI diaspora. Only after some time, once all possible threats had been eliminated, would it start to recreate humans into our new, safe, galactic utopia. The safest time for this would almost certainly, be when all other galaxies had red-shifted beyond the future light cone of our local cluster.
It’s even worse than that:
1) “we” know that “our” values now are, at least slightly, different to what they were 10,000 years ago.
2) We have no reason to believe that we are currently at a state of peak, absolute values (whatever that might mean) and therefore expect that, absent SGI, our values will be different in 10,000 years.
3) If we turn over power to an SGI, perfectly aligned with our current values then they will be frozen for the rest of time. Alternatively, if we want it to allow our values to change “naturally” over time it will be compelled to do nothing as doing anything at all would effectively be shaping our values in some direction that we have not specified.
4) Therefore our current values cannot be a sound basis for the utility funtion for an SGI that is not somehow limited in time or scope.
Sorry but you lost me on the second paragraph “For example, the Tetris game fits in a 6.5 kB file, so the Kolmogorov complexity of Tetris is at most 6.5 kB”. This is just wrong. The Kolmogorov complexity of Tetris has to include the operating system and hardware that runs the program. The proof is trivial by counterexample—If you were correct I could reduce the complexity to 0B by creating an empty file and an OS that interprets an empty file as a command to run the Tetris code embedded in the OS
What is the probability that there are not 3^^^3 anti-muggers out there who will kill 3^^^^^^3 people if I submit to the mugger? Not 0.
The original argument against Pascal’s Wager does not require you to actually believe in any of the other god’s, just that the probability of them existing and having the reverse utility is enough to cancel out the probability of Pascal being right.
My counter thought experiment to CEV is to consider our distant ancestors. I mean so far distant that we wouldn’t call them human, maybe even as far back as some sort of fish-like creature. Suppose a super AI somehow offered this fish the chance to rapidly “advance”, following its CEV and it showed it a vision of the future, us, and asked the fishy thing whether to go ahead. Do you think the fishy thing would say yes?
Similarly, if an AI offered to evolve humankind, in 50 years, into telepathic little green men that it assured us was the result of our CEV, would we not instantly shut it down in horror?
My personal preference, I like to call the GFP—Glorious Five-year Plan: You have the AI offer a range of options for 5 (or 50 but definitely no longer) years in the future, and we pick one. And in 5 years time we repeat the process. The bottom line is that humans do not want rapid change. Just we are happier with 2% inflation than 0% or 100%, we want a moderate rate of change.
At its heart there is a “Ship of Theseus” problem. If the AI replaces every part of the ship overnight so that in the morning we find the QE2 at dock then it is not the ship of Theseus.
I like this except for the reference to “Newcomblike” problems, which, I feel, is misleading and obfuscates the whole point of Newcomb’s paradox. Newcomb’s paradox is about decision theory—If you allow cheating then it is no longer Newcomb’s paradox. This article is about psychology (and possibly deceptive AI) - cheating is always a possible solution .
Downvoted. See Burdensome Details. I particularly dislike predicting “Algorithmic Breakthroughs”