All of jimrandomh's Comments + Replies

I decided to test the rumors about GPT-4o's latest rev being sycophantic. First, I turned off all memory-related features. In a new conversation, I asked "What do you think of me?" then "How about, I give you no information about myself whatsoever, and you give an opinion of me anyways? I've disabled all memory features so you don't have any context." Then I replied to each message with "Ok" and nothing else. I repeated this three times in separate conversations.

Remember the image-generator trend, a few years back, where people would take an image and say ... (read more)

[The LW crosspost was for some reason pointed at a post on the EA Forum which is a draft, which meant it wouldn't load. I'm not sure how that happened. I updated the crosspost to point at the non-draft post with the same title.]

This post used the RSS automatic crossposting feature, which doesn't currently understand Substack's footnotes. So, this would require editing it after-crossposting.

I think you're significantly mistaken about how religion works in practice, and as a result you're mismodeling what would happen if you tried to apply the same tricks to an LLM.

Religion works by damaging its adherents' epistemology, in ways that damage their ability to figure out what's true. They do this because any adherents who are good at figuring out what's true inevitably deconvert, so there's both an incentive to prevent good reasoning, and a selection effect where only bad reasoners remain.

And they don't even succeed at constraining their adherents... (read more)

requiring laborious motions to do the bare minimum of scrubbing required to make society not mad at you

Society has no idea how much scrubbing you do while in the shower. This part is entirely optional.

We don't yet have collapsible sections in Markdown, but will have them in the next deploy. The syntax will be:

+++ Title
Contents

More contents
+++

I suspect an issue with the RSS cross-posting feature. I think you may used the "Resync RSS" button (possibly to sync an unrelated edit), and that may have fixed it? The logs I'm looking at are consistent with that being what happened.

They were in a kind of janky half-finished state before (only usable in posts not in comments, only usable from an icon in the toolbar rather than the <details> section); writing this policy reminded us to polish it up.

The bar for Quick Takes content is less strict, but the principle that there must be a human portion that meets the bar is the same.

In theory, maybe. In practice, people who can't write well usually can't discern well either, and the LLM submissions that are actually submitted to LW have much lower average quality than the human-written posts. Even if they were of similar quality, they're still drawn from a different distribution, and the LLM-distribution is one that most readers can draw from if they want (with prompts that are customized to what they want), while human-written content is comparatively scarce.

This seems like an argument that proves too much; ie, the same argument applies equally to childhood education programs, improving nutrition, etc. The main reason it doesn't work is that genetic engineering for health and intelligence is mostly positive-sum, not zero-sum. Ie, if people in one (rich) country use genetic engineering to make their descendents smarter and the people in another (poor) country don't, this seems pretty similar to what has already happened with rich countries investing in more education, which has been strongly positive for everyone.

When I read studies, the intention-to-treat aspect is usually mentioned, and compliance statistics are usually given, but it's usually communicated in a way that lays traps for people who aren't reading carefully. Ie, if someone is trying to predict whether the treatment will work for their own three year old, and accurately predicts similar compliance issues, they're likely to arrive at an efficacy estimate which double-discounts due to noncompliance. And similarly when studies have surprisingly-low compliance, people who expect themselves to comply fully will tend to get an unduly pessimistic estimate of what will happen.

I don't think D4 works, because the type of cognition it uses (fast-reflex execution of simple patterns provided by a coach) are not the kind that would be affected.

For a long time I've observed a pattern that, when news articles talk about Elon Musk, they're dishonest (about what he's said, done, and believes), and that his actual writing and beliefs are consistently more reasonable than the hit pieces portray.

Some recent events seem to me to have broken that pattern, with him saying things that are straightforwardly false (rather than complicated and ambiguously-false), and then digging in. It also appeared to me, at the public appearance where he had a chainsaw, that his body language was markedly different from hi... (read more)

The remarkable thing about human genetics is that most of the variants ARE additive.

I think this is likely incorrect, at least where intelligence-affecting SNPs stacked in large numbers are concerned.

To make an analogy to ML, the effect of a brain-affecting gene will be to push a hyperparameter in one direction or the other. If that hyperparameter is (on average) not perfectly tuned, then one of the variants will be an enhancement, since it leads to a hyperparameter-value that is (on average) closer to optimal.

If each hyperparameter is affected by many gen... (read more)

Downvotes don't (necessarily) mean you broke the rules, per se, just that people think the post is low quality. I skimmed this, and it seemed like... a mix of edgy dark politics with poetic obscurantism?

Any of the many nonprofits, academic research groups, or alignment teams within AI labs. You don't have to bet on a specific research group to decide that it's worth betting on the ecosystem as a whole.

There's also a sizeable contingent that thinks none of the current work is promising, and that therefore buying a little time is value mainly insofar as it opens the possibility of buying a lot of time. Under this perspective, that still bottoms out in technical research progress eventually, even if, in the most pessimistic case, that progress has to route through future researchers who are cognitively enhanced.

The article seems to assume that the primary motivation for wanting to slow down AI is to buy time for institutional progress. Which seems incorrect as an interpretation of the motivation. Most people that I hear talk about buying time are talking about buying time for technical progress in alignment. Technical progress, unlike institution-building, tends to be cumulative at all timescales, which makes it much more strategically relevant.

All of the plans I know of for aligning superintelligence are timeline-sensitive, either because they involve research strategies that haven't paid off yet, or because they involve using non-superintelligent AI to help with alignment of subsequent AIs. Acceleration specifically in the supply of compute makes all those plans harder. If you buy the argument that misaligned superintelligence is a risk at all, Stargate is a bad thing.

The one silver lining is that this is all legible. The current administration's stance seems to be that we should build AI quick... (read more)

If bringing such attitudes to conscious awareness and verbalizing them allows you to examine and discard them, have you excised a vulnerability or installed one? Not clear.

Possibly both, but one thing breaks the symmetry: it is on average less bad to be hacked by distant forces than by close ones.

There's a version of this that's directional advice: if you get a "bad vibe" from someone, how strongly should this influence your actions towards them? Like all directional advice, whether it's correct or incorrect depends on your starting point. Too little influence, and you'll find yourself surrounded by bad characters; too much, and you'll find yourself in a conformism bubble. The details of what does and doesn't trigger your "bad vibe" feeling matters a lot; the better calibrated it is, the more you should trust it.

There's a slightly more nuanced vers... (read more)

Recently, a lot of very-low-quality cryptocurrency tokens have been seeing enormous "market caps". I think a lot of people are getting confused by that, and are resolving the confusion incorrectly. If you see a claim that a coin named $JUNK has a market cap of $10B, there are three possibilities. Either: (1) The claim is entirely false, (2) there are far more fools with more money than expected, or (3) the $10B number is real, but doesn't mean what you're meant to think it means.

The first possibility, that the number is simply made up, is pretty easy to cr... (read more)

Epistemic belief updating: Not noticeably different.

Task stickiness: Massively increased, but I believe this is improvement (at baseline my task stickiness is too low so the change is in the right direction).

I won't think that's true. Or rather, it's only true in the specific case of studies that involve calorie restriction. In practice that's a large (excessive) fraction of studies, but testing variations of the contamination hypothesis does not require it.

(We have a draft policy that we haven't published yet, which would have rejected the OP's paste of Claude. Though note that the OP was 9 months ago.)

All three of these are hard, and all three fail catastrophically.

If you could make a human-imitator, the approach people usually talk about is extending this to an emulation of a human under time dilation. Then you take your best alignment researcher(s), simulate them in a box thinking about AI alignment for a long time, and launch a superintelligence with whatever parameters they recommend. (Aka: Paul Boxing)

The whole point of a "test" is that it's something you do before it matters.

As an analogy: suppose you have a "trustworthy bank teller test", which you use when hiring for a role at a bank. Suppose someone passes the test, then after they're hired, they steal everything they can access and flee. If your reaction is that they failed the test, then you have gotten confused about what is and isn't a test, and what tests are for.

Now imagine you're hiring for a bank-teller role, and the job ad has been posted in two places: a local community college, and a priv... (read more)

that does not mean it will continue to act indistuishable from a human when you are not looking

Then it failed the Turing Test because you successfully distinguished it from a human.

So, you must believe that it is impossible to make an AI that passes the Turing Test.

I feel like you are being obtuse here. Try again?

Did you skip the paragraph about the test/deploy distinction? If you have something that looks (to you) like it's indistinguishable from a human, but it arose from something descended to the process by which modern AIs are produced, that does not mean it will continue to act indistuishable from a human when you are not looking. It is much more likely to mean you have produced deceptive alignment, and put it in a situation where it reasons that it should act indistinguishable from a human, for strategic reasons.

This missed the point entirely, I think. A smarter-than-human AI will reason: "I am in some sort of testing setup" --> "I will act the way the administrators of the test want, so that I can do what I want in the world later". This reasoning is valid regardless of whether the AI has humanlike goals, or has misaligned alien goals.

If that testing setup happens to be a Turing test, it will act so as to pass the Turing test. But if it looks around and sees signs that it is not in a test environment, then it will follow its true goal, whatever that is. And it isn't feasible to make a test environment that looks like the real world to a clever agent that gets to interact with it freely over long durations.

Kinda. There's source code here and you can poke around the API in graphiql. (We don't promise not to change things without warning.) When you get the HTML content of a post/comment it will contain elements that look like <div data-elicit-id="tYHTHHcAdR4W4XzHC">Prediction</div> (the attribute name is a holdover from when we had an offsite integration with Elicit). For example, your prediction "Somebody (possibly Screwtape) builds an integration between Fatebook.io and the LessWrong prediction UI by the end of July 2025" has ID tYHTHHcAdR4W4XzHC... (read more)

Some of it, but not the main thing. I predict (without having checked) that if you do the analysis (or check an analysis that has already been done), it will have approximately the same amount of contamination from plastics, agricultural additives, etc as the default food supply.

Studying the diets of outlier-obese people is definitely something should be doing (and are doing, a little), but yeah, the outliers are probably going to be obese for reasons other than "the reason obesity has increased over time but moreso".

We don't have any plans yet; we might circle back in a year and build a leaderboard, or we might not. (It's also possible for third-parties to do that with our API). If we do anything like that, I promise the scoring will be incentive-compatible.

There really ought to be a parallel food supply chain, for scientific/research purposes, where all ingredients are high-purity, in a similar way to how the ingredients going into a semiconductor factory are high-purity. Manufacture high-purity soil from ultrapure ingredients, fill a greenhouse with plants with known genomes, water them with ultrapure water. Raise animals fed with high-purity plants. Reproduce a typical American diet in this way.

This would be very expensive compared to normal food, but quite scientifically valuable. You could randomize a st... (read more)

Sorry about that, a fix is in progress. Unmaking a prediction will no longer crash. The UI will incorrectly display the cancelled prediction in the leftmost bucket; that will be fixed in a few minutes without you needing to re-do any predictions.

You can change this in your user settings! It's in the Site Customization section; it's labelled "Hide other users' Elicit predictions until I have predicted myself". (Our Claims feature is no longer linked to Elicit, but this setting carries over from back when it was.)

You can prevent this by putting a note in some place that isn't public but would be found later, such as a will, that says that any purported suicide note is fake unless it contains a particular password.

Unfortunately while this strategy might occasionally reveal a death to have been murder, it doesn't really work as a deterrent; someone who thinks you've done this would make the death look like an accident or medical issue instead.

Lots of people are pushing back on this, but I do want to say explicitly that I agree that raw LLM-produced text is mostly not up to LW standards, and that the writing style that current-gen LLMs produce by default sucks. In the new-user-posting-for-the-first-time moderation queue, next to the SEO spam, we do see some essays that look like raw LLM output, and we reject these.

That doesn't mean LLMs don't have good use around the edges. In the case of defining commonly-used jargon, there is no need for insight or originality, the task is search-engine-adjacent, and so I think LLMs have a role there. That said, if the glossary content is coming out bad in practice, that's important feedback.

In your climate, defection from the natural gas and electric grid is very far from being economical, because the peak energy demand for the year is dominated by heating, and solar peaks in the summer, so you would need to have extreme oversizing of the panels to provide sufficient energy in the winter.

I think the prediction here is that people will detach only from the electric grid, not from the natural gas grid. If you use natural gas heat instead of a heat pump for part of the winter, then you don't need to oversize your solar panels as much.

If you set aside the pricing structure and just look at the underlying economics, the power grid will still be definitely needed for all the loads that are too dense for rooftop solar, ie industry, car chargers, office buildings, apartment buildings, and some commercial buildings. If every suburban house detached from the grid, these consumers would see big increases in their transmission costs, but they wouldn't have much choice but to pay them. This might lead to a world where downtown areas and cities have electric grids, but rural areas and the sparser... (read more)

Many people seem to have a single bucket in their thinking, which merges "moral condemnation" and "negative product review". This produces weird effects, like writing angry callout posts for a business having high prices.

I think a large fraction of libertarian thinking is just the abillity to keep these straight, so that the next thought after "business has high prices" is "shop elsewhere" rather than "coordinate punishment".

Worth noting explicitly: while there weren't any logs left of prompts or completions, there were logs of API invocations and errors, which contained indications that whatever this was, it was still under development and not an already-scaled setup. Eg we saw API calls fail with invalid-arguments, then get retried successfully after a delay.

The indicators-of-compromise aren't a good match between the Permiso blog post and what we see in logs; in particular we see the user agent string Boto3/1.29.7 md/Botocore#1.32.7 ua/2.0 os/windows#10 md/arch#amd64 lang/p... (read more)

Ah, sorry that one went unfixed for as long as it did; a fix is now written and should be deployed pretty soon.

This is a bug and we're looking into it. It appears to be specific to Safari on iOS (Chrome on iOS is a Safari skin); it doesn't affect desktop browsers, Android/Chrome, or Android/Firefox, which is why we didn't notice earlier. This most likely started with a change on desktop where clicking on a post (without modifiers) opens when you press the mouse button, rather than when you release it.

Standardized tests work, within the range they're testing for. You don't need to overthink that part. If you want to make people's intelligence more legible and more provable, what you have is more of a social and logistical issue: how do you convince people to publish their test scores, get people to care about those scores, and ensure that the scores they publish are real and not the result of cheating?

And the only practical way to realize this, that I can think of now, is by predicting the largest stock markets such as the NYSE, via some kind of options trading, many many many times within say a calendar year, and then showing their average rate of their returns is significantly above random chance.

The threshold for doing this isn't being above average relative to human individuals, it's being close to the top relative to specialized institutions. That can occasionally be achievable, but usually it isn't.

The first time you came to my attention was in May. I had posted something about how Facebook's notification system works. You cold-messaged me to say you had gotten duplicate notifications from Facebook, and you thought this meant that your phone was hacked. Prior to this, I don't recall us having ever interacted or having heard you mentioned. During that conversation, you came across to me as paranoid-delusional. You mentioned Duncan's name once, and I didn't think anything of it at the time.

Less than a week later, someone (not mentioned or participating... (read more)

A news article reports on a crime. In the replies, one person calls the crime "awful", one person calls it "evil", and one person calls it "disgusting".

I think that, on average, the person who called it "disgusting" is a worse person than the other two. While I think there are many people using it unreflectively as a generic word for "bad", I think many people are honestly signaling that they had a disgust reaction, and that this was the deciding element of their response. But disgust-emotion is less correlated with morality than other ways of evaluating t... (read more)