I agree that the average reader is probably smarter in a general sense, but they also have FAR more things competing for their attention. Thus the amount of intelligence available for reading and understanding any given sentence, specifically, may be lower in the modern environment.
Kenoubi
Karma: 353
Question marks and exclamation points are dots with an extra bit. Ellipses may be multiple dots, but also indicate an uncertain end to the sentence. (Formal usage distinguishes ”...” for ellipses in arbitrary position and ”....” for ellipses coming after a full stop, but the latter is rarely seen in any but academic writing, and I would guess even many academics don’t notice the difference these days.)
I read a bunch of its “thinking” and it gets SO close to solving it after the second message, but it miscounts the number of [] in the text provided for 19. Repeatedly. While quoting it verbatim. (I assume it foolishly “trusts” the first time it counted.) And based on its miscount, thinks that should be the representation for 23, instead. And thus rules out (a theory starting to point towards) the corrects answer.
I think this may at least be evidence that having anything unhelpful in context, even (maybe especially!) if self-generated, can be really harmful to model capabilities. I still think it’s pretty interesting.
I have very mixed feelings about this comment. It was a good story (just read it, and wouldn’t have done so without this comment) but I really don’t see what it has to do with this LW post.
Possible edge case / future work—what if you optimize for faithfulness and legibility of the chain of thought? The paper tests optimizing for innocent-looking CoT, but if the model is going to hack the test either way, I’d want it to say so! And if we have both an “is actually a hack” detector and a “CoT looks like planning a hack” detector, this seems doable.
Is this an instance of the Most Forbidden Technique? I’m not sure. I definitely wouldn’t trust it to align a currently unaligned superintelligence. But it seems like maybe it would let you make an aligned model at a given capability level into a still aligned model with more legible CoT, without too much of a tax, as long as the model doesn’t really need illegible CoT to do the task? And if capabilities collapse, that seems like strong evidence that illegible CoT was required for task performance; halt and catch fire, if legible CoT was a necessary part of your safety case.
Is it really plausible that human driver inattention just doesn’t matter here? Sleepiness, drug use, personal issues, eyes were on something interesting rather than the road, etc. I’d guess something like that is involved in a majority of collisions, and that Just Shouldn’t Happen to AI drivers.
Of course AI drivers do plausibly have new failure modes, like maybe the sensors fail sometimes (maybe more often than human eyes just suddenly stop working). But there should be plenty of data about that sort of thing from just testing them a lot.
The only realistic way I can see for AI drivers, that have been declared street-legal and are functioning in a roadway and regulatory system that humans (chose to) set up, to be less safe than human drivers is if there’s some kind of coordinated failure. Like if they trust data coming from GPS satellites or cell towers, and those start spitting out garbage and throw the AIs off distribution; or a deliberate cyber-attack / sabotage of some kind.
“ goal” in “football| goal|keeping”
Looks like an anti-football (*American* football, that is) thing, to me. American football doesn’t have goals, and soccer (which is known as “football” in most of the world) does. And you mentioned earlier that the baseball neuron is also anti-football.
Since it was kind of a pain to run, sharing these probably minimally interesting results. I tried encoding this paragraph from my comment:
I wonder how much information there is in those 1024-dimensional embedding vectors. I know you can jam an unlimited amount of data into infinite-precision floating point numbers, but I bet if you add Gaussian noise to them they still decode fine, and the magnitude of noise you can add before performance degrades would allow you to compute how many effective bits there are. (Actually, do people use this technique on latents in general? I’m sure either they do or they have something even better; I’m not a supergenius and this is a hobby for me, not a profession.) Then you could compare to existing estimates of text entropy, and depending on exactly how the embedding vectors are computed (they say 512 tokens of context but I haven’t looked at the details enough to know if there’s a natural way to encode more tokens than that; I remember some references to mean pooling, which would seem to extend to longer text just fine?), compare these across different texts.
with SONAR, breaking it up like this:
sentences = [ 'I wonder how much information there is in those 1024-dimensional embedding vectors.', 'I know you can jam an unlimited amount of data into infinite-precision floating point numbers, but I bet if you add Gaussian noise to them they still decode fine, and the magnitude of noise you can add before performance degrades would allow you to compute how many effective bits there are.', '(Actually, do people use this technique on latents in general? I\'m sure either they do or they have something even better; I\'m not a supergenius and this is a hobby for me, not a profession.)', 'Then you could compare to existing estimates of text entropy, and depending on exactly how the embedding vectors are computed (they say 512 tokens of context but I haven\'t looked at the details enough to know if there\'s a natural way to encode more tokens than that;', 'I remember some references to mean pooling, which would seem to extend to longer text just fine?), compare these across different texts.']and after decode, I got this:
['I wonder how much information there is in those 1024-dimensional embedding vectors.', 'I know you can encode an infinite amount of data into infinitely precise floating-point numbers, but I bet if you add Gaussian noise to them they still decode accurately, and the amount of noise you can add before the performance declines would allow you to calculate how many effective bits there are.', "(Really, do people use this technique on latent in general? I'm sure they do or they have something even better; I'm not a supergenius and this is a hobby for me, not a profession.)", "And then you could compare to existing estimates of text entropy, and depending on exactly how the embedding vectors are calculated (they say 512 tokens of context but I haven't looked into the details enough to know if there's a natural way to encode more tokens than that;", 'I remember some references to mean pooling, which would seem to extend to longer text just fine?), compare these across different texts.']Can we do semantic arithmetic here?
sentences = [ 'A king is a male monarch.', 'A bachelor is an unmarried man.', 'A queen is a female monarch.', 'A bachelorette is an unmarried woman.' ] ... pp(reconstructed) ['A king is a male monarch.', 'A bachelor is an unmarried man.', 'A queen is a female monarch.', 'A bachelorette is an unmarried woman.'] ... new_embeddings[0] = embeddings[0] + embeddings[3] - embeddings[1] new_embeddings[1] = embeddings[0] + embeddings[3] - embeddings[2] new_embeddings[2] = embeddings[1] + embeddings[2] - embeddings[0] new_embeddings[3] = embeddings[1] + embeddings[2] - embeddings[3] reconstructed = vec2text_model.predict(new_embeddings, target_lang="eng_Latn", max_seq_len=512) pp(reconstructed) ['A kingwoman is a male monarch.', "A bachelor's is a unmarried man.", 'A bachelorette is an unmarried woman.', 'A queen is a male monarch.']Nope. Interesting though. Actually I guess the 3rd one worked?
OK, I’ll stop here, otherwise I’m at risk of going on forever. But this seems like a really cool playground.
You appear to have two full copies of the entire post here, one above the other. I wouldn’t care (it’s pretty easy to recognize this and skip the second copy) except that it totally breaks the way LW does comments on and reactions to specific parts of the text; one has to select a unique text fragment to use those, and with two copies of the entire post, there aren’t any unique fragments.
Wow, the SONAR encode-decode performance is shockingly good, and I read the paper and they explicitly stated that their goal was translation, and that the autoencoder objective alone was extremely easy! (But it hurt translation performance, presumably by using a lot of the latent space to encode non-semantic linguistic details, so they heavily downweighted autoencoder loss relative to other objectives when training the final model.)
I wonder how much information there is in those 1024-dimensional embedding vectors. I know you can jam an unlimited amount of data into infinite-precision floating point numbers, but I bet if you add Gaussian noise to them they still decode fine, and the magnitude of noise you can add before performance degrades would allow you to compute how many effective bits there are. (Actually, do people use this technique on latents in general? I’m sure either they do or they have something even better; I’m not a supergenius and this is a hobby for me, not a profession.) Then you could compare to existing estimates of text entropy, and depending on exactly how the embedding vectors are computed (they say 512 tokens of context but I haven’t looked at the details enough to know if there’s a natural way to encode more tokens than that; I remember some references to mean pooling, which would seem to extend to longer text just fine?), compare these across different texts.
Exploring this embedding space seems super interesting, in general, way more so on an abstract level (obviously it isn’t as directly useful at this point) than the embedding space used by actual LLMs. Like, with only 1024 dimensions for a whole paragraph, it must be massively polysemantic, right? I guess your follow-on post (which this was just research to support) is implicitly doing part of this, but I think maybe it underplays “can we extract semantic information from this 1024-dimensional embedding vector in any way substantially more efficient than actually decoding it and reading the output?” (Or maybe it doesn’t; I read the other post too, but haven’t re-read it in light of this one.)
There also appears to be a way to attempt to use this to enhance model capabilities. I seem to think of one of these every other week, and again, I’m not a supergenius nor a professional ML researcher so I assume it’s obvious to those in the field. The devil appears to be in the details; sometimes a new innovation appears to be a variant of something I thought of years ago, sometimes they come out of left field from my perspective, and in no case does there appear to be anything I, from my position, could have usefully done with the idea, so far. Experiments seem very compute-limited, especially because like all other software development in my experience, one needs to actually run the code and see what happens. This particular technique, if it actually works (I’m guessing either it doesn’t, or it only works when scaled so large that a bunch of other techniques would have worked just as well and converged on the same implicit computations) might come with large improvements to interpretability and controllability, or it might not (which seems to be true for all the other ideas I have that might improve capabilities, too). I’m not advising anyone to try it (again, if one works in the field I think it’s obvious, so either there are reasons not to or someone already is). Just venting, I guess. If anyone’s actually reading this, do you think there’s anything useful to do with this idea and others like it, or are they pretty much a dime a dozen, interesting to me but worthless in practice?
(Sorry for going on so long! Wish I had a way to pay a penny to anyone who thoughtfully reads this, whether or not they do anything with it.)
Sorry, I think it’s entirely possible that this is just me not knowing or understanding some of the background material, but where exactly does this diverge from justifying the AI pursuing a goal of maximizing the inclusive genetic fitness of its creators? Which clearly either isn’t what humans actually want (there are things humans can do to make themselves have more descendants that no humans, including the specific ones who could take those actions, want to take, because of godshatter) or is just circular (who knows what will maximize inclusive genetic fitness in an environment that is being created, in large part, by the decision of how to promote inclusive genetic fitness?). At some point, your writing started talking about “design goals”, but I don’t understand why tools / artifacts constructed by evolved creatures, that happen to increase the inclusive genetic fitness of the evolved creatures who constructed them by means other than the design goals of those who constructed them, wouldn’t be favored by evolution, and thus part of the “purpose” of the evolved creatures in constructing them; and this doesn’t seem like an “error” even in the limit of optimal pursuit of inclusive genetic fitness, this seems to be just what optimal pursuit of IGF would actually do. In other words, I don’t want a very powerful human-constructed optimizer to pursue the maximization of human IGF, and I think hardly any other humans do either; but I don’t understand in detail why your argument doesn’t justify AI pursuit of maximizing human IGF, to the detriment of what humans actually value.
As the person who requested of MIRI to release the Sequences as paper books in the first place, I have asked MIRI to release the rest of them, and credibly promised to donate thousands of dollars if they did so. Given the current situation vis-a-vis AI, I’m not that surprised that it still does not appear to be a priority to them, although I am disappointed.
MIRI, if you see this, yet another vote for finishing the series! And my offer still stands!
Thank you for writing this. It has a lot of stuff I haven’t seen before (I’m only really interested in neurology insofar as it’s the substrate for literally everything I care about, but that’s still plenty for “I’d rather have a clue than treat the whole area as spooky stuff that goes bump in the night”).
As I understand it, you and many scientists are treating energy consumption by anatomical part of the brain (as proxied by blood flow) as the main way to see “what the brain is doing”. It seems possible to me that there are other ways that specific thoughts could be kept compartmentalized, e.g. which neurotransmitters are active (although I guess this correlates pretty strongly to brain region anyway) or microtemporal properties of neural pulses; but the fact that we’ve found any kind of reasonably consistent relationship between [brain region consuming energy] and [mental state as reported or as predicted by the situation] means that brain region is a factor used for separating / modularizing cognition, if not that it’s the only such part. So, I’ll take brain region = mental module for granted for now and get to my actual question:
Do you know whether anyone has compiled data, across a wide variety of experiments or other data-gathering opportunities, of which brain regions have which kinds of correlations with one another? E.g. “these two tend to be active simultaneously”, “this one tends to become active just after this one”, etc.
I’m particularly interested in this for the brain regions you mention in this article, those related in various senses to good and/or to bad. If one puts both menthol and capsaicin in one’s mouth at the same time, the menthol will stimulate cold receptors and the capsaicin will stimulate heat receptors, and one will have an experience out of range of what the sensors usually encounter: hot and cold, simultaneously in the same location. What I actually want to know is: are good and bad (or some forms of them, anyway) also represented in a way where one isn’t actually the opposite of the other, neurologically speaking? If so, are there actual cases that are clearly best described as “good and bad”, where to pick a single number instead would inevitably miss the intensity of the experience?
2 years and 2 days later, in your opinion, has what you predicted in your conclusion happened?
(I’m just a curious bystander; I have no idea if there are any camps regarding this issue, but if so, I’m not a member of any of them.)
might put lawyers out of business
This might be even worse than she thought. Many, many contracts include the exact opposite of this clause, i.e., that the section titles are without any effect whatsoever on the actual interpretation of the contract. I never noticed until just now that this is an instance of self-dealing on the part of the attorneys (typically) drafting the contracts! They’re literally saying that if they make a drafting error, in a way that makes the contract harder to understand and use and is in no conceivable way an improvement to the contract, the courts need to assume “well, one common kind of drafting error is putting a clause in the wrong section, probably that’s what happened here” only because the “clauses in wrong section are void” provisions you mentioned are as far as I know literally unheard of!
I was just reading about this, and apparently subvocalizing refers to small but physically detectable movement of the vocal cords. I don’t know whether / how often I do this (I am not at all aware of it). But it is literally impossible for me to read (or write) without hearing the words in my inner ear, and I’m not dyslexic (my spelling is quite good and almost none of what’s described in OP sounds familiar, so I doubt it’s that I’m just undiagnosed). I thought this was more common than not, so I’m kind of shocked that the reacts on this comment’s grandparent indicate only about 1⁄3 (of respondents to the “poll”) subvocalize. The voice I hear is quite featureless, and I can read maybe 300 words per minute, which I think is actually faster than average, though needing to “hear” the words does impose an upper bound on reading speed.
Leaving an unaligned force (humans, here) in control of 0.001% of resources seems risky. There is a chance that you’ve underestimated how large the share of resources controlled by the unaligned force is, and probably more importantly, there is a chance that the unaligned force could use its tiny share of resources in some super-effective way that captures a much higher fraction of resources in the future. The actual effect on the economy of the unaligned force, other than the possibility of its being larger than thought or being used as a springboard to gain more control, seems negligible, so one should still expect full extermination unless there’s some positive reason for the strong force to leave the weak force intact.
Humans do have such reasons in some cazes (we like seeing animals, at least in zoos, and being able to study them, etc.; same thing for the Amish; plus we also at least sometimes place real value on the independence and self-determination of such beings and cultures), but there would need to be an argument made that AI will have such positive reasons (and a further argument why the AIs wouldn’t just “put whatever humans they wanted to preserve” in “zoos”, if one thinks that being in a zoo isn’t a great future). Otherwise, exterminating humans would be trivially easy with that large of a power gap. Even if there are multiple ASIs that aren’t fully aligned with one another, offense is probably easier than defense; if one AI perceives weak benefits to keeping humans around, but another AI perceives weak benefits to exterminating us, I’d assume we get exterminated and then the 2nd AI pays some trivial amount to the 1st for the inconvenience. Getting AI to strongly care about keeping humans around is, of course, one way to frame the alignment problem. I haven’t seen an argument that this will happen by default or that we have any idea how to do it; this seems more like an attempt to say it isn’t necessary.
Ah, okay, some of those seem to me like they’d change things quite a lot. In particular, a week’s notice is usually possible for major plans (going out of town, a birthday or anniversary, concert that night only, etc.) and being able to skip books that don’t interest one also removes a major class of reason not to go. The ones I can still see are (1) competing in-town plans, (2) illness or other personal emergency, and (3) just don’t feel like going out tonight. (1) is what you’re trying to avoid, of course. On (3) I can see your opinion going either way. It does legitimately happen sometimes that one is too tired for whatever plans one had to seem appealing, but it’s legitimate to say that if that happens to you so often that you mind the cost of the extra rounds of drinks you end up buying, maybe you’re not a great member for that club. (2) seems like a real problem, and I’m gonna guess that you actually wouldn’t make people pay for drinks if they said they missed because they had COVID, there was a death in the family, etc.?
Reads like a ha ha only serious to me anyway.
That’s possible, but what does the population distribution of [how much of their time people spend reading books] look like? I bet it hasn’t changed nearly as much as overall reading minutes per capita has (even decline in book-reading seems possible, though of course greater leisure and wealth, larger quantity of cheaply and conveniently available books, etc. cut strongly the other way), and I bet the huge pile of written language over here has large effects on the much smaller (but older) pile of written language over there.
(How hard to understand was that sentence? Since that’s what this article is about, anyway, and I’m genuinely curious. I could easily have rewritten it into multiple sentences, but that didn’t appear to me to improve its comprehensibility.)
Edited to add: on review of the thread, you seem to have already made the same point about book-reading commanding attention because book-readers choose to read books, in fact to take it as ground truth. I’m not so confident in that (I’m not saying it’s false, I really don’t know), but the version of my argument that makes sense under that hypothesis would crux on books being an insufficiently distinct use of language to not be strongly influenced, either through [author preference and familiarity] or through [author’s guesses or beliefs about [reader preference and familiarity]], by other uses of language.