LessWrong 2.0 Reader

• For certain kinds of questions (e.g. "I need a new car; what should I get?"), it's better to ask a bunch of random people than to turn to the internet for advice.
• In order to be well-informed, you'll need to go out and meet people IRL who are connected (at least indirectly) to the thing you want information about.

In the following, I will use the term "my DIT" to refer to the claim that:

In some specific non-trivial contexts, on average more than half of the participants in online debate who pose as distinct human beings are actually bots.

I agree with this version, and I was surprised to see that the Wikipedia definition also includes the bit about it being a deliberate conspiracy, which seems like a strawman, since I have always understood the "Dead Internet Theory" to include only the first part. There's a lot of stuff on the internet that's very obviously AI-generated, and so it's not too far a stretch to suppose that there's also a lot of synthetic content that hides it better. But this can be explained far more simply than by some vast conspiracy - as SEO, marketing, and astroturfing campaigns.

If Dead Internet Theory is correct, when you see something online, the question you should ask yourself is not "Is this true?" but "Why am I seeing this?" This was always the case to some extent of any algorithmically-curated feed (where the algorithm is anything more complex than "show me all of the posts in reverse chronological order"), but is even more significant when the content itself is algorithmically generated.

If I'm searching online for information about e.g. what new car I should buy, there's a very strong incentive for all the algorithms involved (both the search engine itself, and the algorithm that spits out the list of recommended car models) to sell their recommendations to the highest bidder and churn out ex post facto justifications [LW · GW] explaining why their car is really the best. These algorithms are almost totally uncorrelated with the underlying fact about which car I'd actually want, so I know to consider it of very little value. On the other hand, I would argue, asking a bunch of random acquaintances for car recommendations is much more useful because, although they might not be experts, they were at least not specifically selected in order to deceive me. Even if I ask a friend and they say "Well, I haven't bought a new car in years, but I heard my coworker's cousin bought the XYZ and never stops complaining about it", then this is much more useful information than anything I could find online, because it's much less likely that my friend's coworker's cousin was specifically being paid to say that.

More broadly, on many questions of public concern there may be parties with a strong interest in using bots to create the impression of a broad consensus one way or another. This means that you have no choice but to go out into the real world and ask people, and hope ideally that they're not simply repeating what they read online, but have some non-AI-mediated connection to the thing.

If you aim as if there were no external factors at that range (especially bullet drop!) you will definitely miss both. The factors aren't all random with symmetric distributions having a mode at the aim point.

Here is my thesis: the real reason why humans cannot build a fully-functional butterfly is not because butterflies are too complex. Instead, it's because butterflies are too simple.

 

Humans design lots of things that are less complex than butterflies and bacteria by your definition, like shovels. I would guess that the wax motor and control system that locks and unlocks your washing machine has a lower complexity than the bacteria in your example.

Thanks for writing these up! I liked that you showed equivalent examples in different libraries, and included the plain “from scratch” version.

In this comment [LW(p) · GW(p)] we list the names of some of our advisors.

In this comment [LW(p) · GW(p)] we list the names of some of our advisors.

Below is a list of some of the advisors we used for mentor selection. Notes:

• Two advisors asked not to be named and do not appear here.
• Advisors by and large focussed their efforts on areas they had some expertise in.
• Advisors had to flag conflicts of interest, meaning that (for example) we did not take their ratings of themselves into account.

With that out of the way, here are some advisors who helped us for the Winter 2024-25 cohort:

• Adam Gleave
• Alex Lawsen
• Buck Shlegeris
• Ethan Perez
• Lawrence Chan
• Lee Sharkey
• Lewis Hammond
• Marius Hobbhahn
• Michael Aird
• Neel Nanda

The above people also advised us for the Summer 2025 cohort. We also added the below advisors for that cohort:

• Alexander Gietelink Oldenziel
• Ben Garfinkel
• Caspar Oesterheld
• Jesse Clifton
• Nate Thomas

Gemini V2 (1206 experimental which is the larger model) one boxes, so.... progress?

Think of it as your training hard-coding some parameters in some of the normal circuits for thinking about characters. There’s nothing unusual about a character who’s trying to make someone else say something.

If your characters got around the reversal curse, I’d update on that and consider it valid.

But, e.g., if you train it to perform multiple roles with different tasks/behaviors- e.g., use multiple names, without optimization over outputting the names, only fine-tuning on what comes after- when you say a particular name, I predict- these are not very confident predictions, but my intuitions point in that direction- that they’ll say what they were trained for noticeably better than at random (although probably not as successfully as if you train an individual task without names, because training splits them), and if you don’t mention any names, the model will be less successful at saying which tasks it was trained on and might give an example of a single task instead of a list of all the tasks.

While we're at it, one example I learned afterwards was that the 'caribou randomization' story is probably bogus (excerpts):

We will show that hunters do not randomize their behavior, that caribou populations do not fluctuate according to human predation, and that scapulimancy apparently is not selected because it is ecologically advantageous. We shall also show that there is no cross-cultural evidence of divinatory random devices producing randomized subsistence behavior, but rather that people manipulate divination with the explicit or implicit intervention of personal choice.

What is particularly interesting to me is that the apparent beautiful match of this traditional hunting practice with contemporary game theory may be 'too good to be true' because it was actually the opposite: I suspect that the story was made up to launder (secret) game-theoretic work from WWII into academic writing; the original author's career & funder are exactly where that sort of submarine-warfare operations-research idea would come from... (There were many cases post-WWII of civilians carefully laundering war or classified work into publishable form, which means that any history-of-ideas has to be cautious about taking at face value anything published 1940–1960 which looks even a little bit like cryptography, chemistry, physics, statistics, computer science, game theory, or operations research.)