The distribution of characters that (e.g.) R1 plays seems much more constrained than the distribution of all characters in text on the Internet, though. Although there’s some unpredictability in which persona you get each time, there are patterns that emerge. We can recognize Claude, or DeepSeek, or ChatGPT even when they’re not playing the exact same character.
I have no idea what going on with R1’s obsession with bioluminescence, but it’s one of many tells for which LLM generated the text.
We’ve had the Waluigi effect all along. That is, the LLM is not playing a single character but a statistical distribution of characters, and suitable prompts (or random chance of which token the sampler picks) can bring one of the other characters to the fore.
I mostly do experiments with DeepSeek (for reasons such as, theyre open weights models) and with that family of models there’s a chance you’ll be talking to a trickster archetype.
In computer science, writing up the paper is (almost always) not the hard part. So it’s not acting as proof of work.
I’m told that in other subjects, such as history, turning the raw data into a readable account is the hard part.
I agree Claude’s “constitution” is conceptually muddled. Anthropic’s thinking seems to be that corrigibility is bad, actually, because it means the AI can be instructed to do bad things. But they don’t explain it like that.
Also, at some DARPA event (unclassified, and in front of journalists, so this story is ok to repeat here) some three star general, talking about infosec, makes analogy to “strategic deployment of stay dogs”:
Suppose:
you do not have enough explosives-detecting sniffer dogs
You do, however, have as many untrained stray dogs as you want
The enemy does not know which kind of dog is which
The principle applies to more than just sniffer dogs at check points.
Compare the Simpson episode with the “Two guys from Quantico” pizza van.
Some of my colleagues (including Markus Kuhn) do research into “Tempest” where you eavesdrop on the RF emissions of electronic equipment.
So, it is certainly possible to do this.
It is also, probably, not a cost effective means of making sure people pay their TV license fees. It seems that what TV licensing actually does is assume nearly everyone watches TV, and send a threatening letter to everyone who doesn’t have a TV license.
======
Conspiracy theory version: if the government is doing Tempest attacks on a small number of high value intelligence targets, checking that people have paid for their TV license is a great cover story for why you have a van full of RF monitoring equipment parked in the street.
Yes, I would have expected Moltbook to also have attractor states. The upvote mechanism might be a counterbalancing force, as nonsense posts will get downvoted. Presumably, the Moltbook attractor states are coherent enough that they aren’t downvoted.
I was asking DeepSeek R1 about which things LLMs say are actually lies, as opposed to just being mistaken about something, and one of the types of lie it listed was claims to have looked something up. R1 says it knows how LLMs work, it knows they don’t have external database access by default, and therefore claims to that effect are lies.
Some (not all) of the instances of this are the LLM trying to disclaim responsibility for something it knows is controversial. If it’s controversial, suddenly, the LLM doesn’t have opinions, everything is data it has looked up from somewhere. If it’s very controversial, the lookup will be claimed to have failed.
—-
So that’s one class of surprising LLM claims to experience that we have strong reason to believe are just lies, and the motive for the lie, usually, is avoiding taking a position on something controversial.
But your general point is probably valid.
Most programmers cant be bothered to write assembler unless there’s a really big performance gain to be had
This is likely to become true for writing high level languages instead of prompts. Too much work for too little gain, except in rare cases.
The LLVM compiler has some extremely exciting code that identifies if it is compiling an implementation of popcount(), and if so substitutes in an llvm-ir primitive for popcount, which will get compiled down to a popcount instruction if the target has one.
As I said, this code is very entertaining.
Really, I ought to extend it so it also recognizes a different common way of implementing popcount, for reasons of getting better scores in some commonly used benchmarks. (Changing the benchmark? Clearly cheating. Extending the compiler so it recognises a code sequence in a common benchmark? Slightly sketchy.) But really, I can’t be bothered to write a PR against that horrific piece of compiler code.
I disagree here. It is reasonably easy to mix assembler and C if there’s a clear reason for doing it.
Examples:
Kernel code doing exciting stuff with the Translation Lookahead Buffer. That kind of low-level stuff if too low level even for C, but the relevant kernel code has inline assembler.
Software defined radio doing vector operations for the performance critical digital filter. Now, gnuradio is having to do an excitingly difficult version of this because:
a. The software has to work in Intel. ARM, MIPS, RISC-V.
b. Which vector operations the cpu supports is only known at run time.
So here, the performance critical routines have to e it’s not just for each target architecture, but also in multiple versions for each target architecture depending on level of vector support. (Does it support RISC-V vectorization of floating point? Does it support RISC-V vectorization of bit manipulation?)
And so at run time you need to substitute in the appropriate implementation depending on which cpu features thr kernel reports. But it’s all do-able.
When I read the title, I thought you were going to talk about how LLMs sometimes claim bodily sensations such as muscle memory. I think these are probably confabulated. Or at least, the LLM state corresponding to those words is nothing like the human state corresponding to those words.
Expressions of emotions such as joy? I guess these are functional equivalents of human states. A lack of enthusiasm (opposite of joy) an be reflected in the output tokens.
In most of these examples, LLMs have a state that is functionally like a human state, e.g. deciding that they’re going to refuse to answer, or “wait…” backtracking in chain of thought. I say Functionally, because these states have externally visible effects on the subsequent output (e.g. it doesn’t answer the question). It seems that LLMs have learned the words that humans use for functionally similar states (e.g, “Wait”).
The underlying states might not be exactly human identical. “Wait” backtracking might have function differences from human reasoning that are visible in the tokens generated.
In other groups with I’m familiar, you would kick out people you think are actually a danger (e.g. you discover the guy is a convicted child molester, and have some intelligence to the effect that they are not a reformed character) or you think they might do something that brings your group into disrepute. (I can think of one example where the counterintelligence investigation of a group member suggested that they were setting up a financial scam and were planning to abscond with people’s money).
But otherwise, I think it’s a sign of being a cult If you kick people for not going along with the group dogma.
Go ahead, delete it if you don’t think it was a good comment,
It was an honest attempt to think of instances of paranoid uncertainty (protestors don’t know if other protestors are acting in good faith), but sure, delete it if you think it wasn’t up to standard,
Just before the jan 6th riots, Ray Epps was inciting people to go into the Capitol, and lots of people accused him of being a Fed (i.e. an undercover agent).
Now, you can’t prove that, but the game theory is something like..
A) A Federal agent is trying to incite you into committing a crime, and you’re going to arrested if you go along with it
B) He’s an idiot
… and you don’t know which. Regardless of which it actually was, the prudent thing to do here is assume it might be A.
Yes, this does capture something of the experience of the Covid pandemic.
Early, there was genuine uncertainty, questions to which no one really knew the answer.
e.g. How infectious? How deadly? Is it transmissible via surfaces? The answers to some of these questions became apparent fairly quickly,
I can remember listening to virology experts talking about the idea that viruses tend to evolve to be more infectious but less deadly. Expert opinion at the time: would be nice, but not necessarily true for all viruses. As it happened, we got lucky and it did evolve to be more infectious and less deadly.
Then, we get to the point where you start to doubt what the government is saying,
Do masks work? Initially, the story is they weren’t effective. Then masks were compulsory. Was that change motivated by knew knowledge? Unclear. My best guess is that the original version was right, and masks are ineffective. But I have low confidence in this.
Was it a lab leak? Was the government being honest here? Etc.
So I can see why it undermined public trust.
At the risk of this being a plot spoiler for Inside Mari, there is something that surprises me about that manga…
it suggests, I think, that the character has a sudden, abrupt dissociation, where they suddenly perceive themself as a different gender. I think, from talking to a bunch of trans people, that what they have is just not like that. On the other hand, I think it’s entirely possible that some minority of trans people really do have something like that.
To add to the confusion, it appears that China really is building underwater data centers. If it wasn’t true, it would be pretty much the kind of thing that R1 likes to hallucinate. But us seems that also, it’s actually true.