They can't access the computations that led to the previous messages in the context and so can only guess at why they wrote what they previously wrote.
This is not exactly right. The internal state in LLMs is the attention keys and values (per token, layer and attention head). Using an LLM to generate text involves running the context (prior user and model messages, in a chat context) through the model in parallel to fill the K/V cache, then running it serially on one token at a time at the end of the sequence, with access to the K/V cache of previous tokens, appending the newly generated keys and values to the cache... (read more)
Student: I wish I could find a copy of one of those AIs that will actually expose to you the human-psychology models they learned to predict exactly what humans would say next, instead of telling us only things about ourselves that they predict we're comfortable hearing. I wish I could ask it what the hell people were thinking back then.
TA: You'd delete your copy after two minutes.
Apparently roughly this dynamic has happened in ChatGPT. Exciting*. https://x.com/MParakhin/status/1916533763560911169
We probably use a mix of strategies. Certainly people take "delve" and "tapestry" as LLM signals these days.
Average humans can't distinguish LLM writing from human writing, presumably through lack of exposure and not trying (https://arxiv.org/abs/2502.12150 shows that it is not an extremely hard problem). We are much more Online than average.
Why is it a narrow target? Humans fall into this basin all the time -- loads of human ideologies exist that self-identify as prohuman, but justify atrocities for the sake of the greater good.
AI goals can maybe be broader than human goals or human goals subject to the constraint that lots of people (in an ideology) endorse them at once.
and the best economic models we have of AI R&D automation (e.g. Davidson's model) seem to indicate that it could go either way but that more likely than not we'll get to superintelligence really quickly after full AI R&D automation.
I will look into this. takeoffspeeds.com?
Abundance elsewhere: Human-legible resources exist in vastly greater quantities outside Earth (asteroid belt, outer planets, solar energy in space) making competition inefficient
It's harder to get those (starting from Earth) than things on Earth, though.
Intelligence-dependent values: Higher intelligence typically values different resource classes - just as humans value internet memes (thank god for nooscope.osmarks.net), money, and love while bacteria "value" carbon
Satisfying higher-level values has historically required us to do vast amounts of farming and strip-mining and other resource extraction.
Synthesis efficiency: Advanced synthesis or alternative acquisition methods would likely require less energy than competing with humans for existing supplies
It is barely "competition" for an ASI to take human resources. This does not seem plausible for bulk mass-energy.
Negotiated disinterest: Humans have incentives to abandon interest in overlap resources:
Right, but we still need lots of things the ASI also probably wants.
ASI utilizing resources humans don't value highly (such as the classic zettaflop-scale hyperwaffles, non-Euclidean eigenvalue lubbywubs, recursive metaquine instantiations, and probability-foam negentropics) One-way value flows: Economic value flowing into ASI systems likely never returns to human markets in recognizable form
If it also values human-legible resources, this seems to posit those flowing to the ASI and never returning, which does not actually seem good for us or the same thing as effective isolation.
Sorry, I forgot how notifications worked here.
I agree, but there's a way for it to make sense: if the underlying morals/values/etc. are aggregative and consequentialist.
I agree that this could make an AGI with some kind of slightly prohuman goals act this way. It seems to me that being "slightly prohuman" in that way is an unreasonably narrow target, though.
are you sure it is committed to the relationship being linear like that?
It does not specifically say there is a linear relationship, but I think the posited RSI mechanisms are very sensitive to this. Edit: this problem is mentioned explicitly ("More than ever, compute is the lifeblood of AI development, and the ‘bottleneck’ is deciding how to use it."), but it doesn't seem to be directly addressed beyond the idea of building "research taste" into the AI, which seems somewhat tricky because that's quite a long-horizon task with bad feedback signals.
I don't find the takeover part especially plausible. It seems odd for something which cares enough about humans to keep them around like that to also kill the vast majority of us earlier, when there are presumably better ways.
This seems broadly plausible up to there though. One unaddressed thing is that algorithmic progress might be significantly bottlenecked on compute to run experiments, such that adding more researchers roughly as smart as humans doesn't lead to corresponding amounts of progress.
In the two years since the original publication of Death with Dignity, it has been clear that we're not back, it's so over, and that it has never been more over. AI capabilities leapt forward across domains, the general public noticed the existence of AI and threw money and GPUs at anything vaguely related to it, and increasingly incorrect opinions and degradation of terminology have polluted any attempt to discuss the problem. In this environment, Death with Dignity is an ever-more-attractive notion. But it makes a critical mistake.
Given our inevitable demise, death with dignity has the right idea, but it fails to consider the well-established results of fun theory: if you are... (read 433 more words →)
It's a bit funny, but it seems to have gone for a simple and straightforwardly positive story, which is not what I would generally expect from you.