They demonstrated and poorly named an ontological attractor state in the Claude model card that is commonly reported in other models.
You linked to the entire system card paper. Can you be more specific? And what would a better name have been?
Actually, OAI the other month found in a paper that a lot of the blame for confabulations could be laid at the feet of how reinforcement learning is being done.
All the labs basically reward the models for getting things right. That’s it.
Notably, they are not rewarded for saying “I don’t know” when they don’t know.
So it’s like the SAT where the better strategy is always to make a guess even if you don’t know.
The problem is that this is not a test process but a learning process.
So setting up the reward mechanisms like that for reinforcement learning means they produce models that are prone to bullshit when they don’t know things.
TL;DR: The labs suck at RL and it’s important to keep in mind there’s only a handful of teams with the compute access for training SotA LLMs, with a lot of incestual team compositions, so what they do poorly tends to get done poorly across the industry as a whole until new blood goes “wait, this is dumb, why are we doing it like this?”
It’s more like they are a sophisticated world modeling program that builds a world model (or approximate “bag of heuristics”) modeling the state of the context provided and the kind of environment that produced it, and then synthesize that world model into extending the context one token at a time.
But the models have been found to be predicting further than one token at a time and have all sorts of wild internal mechanisms for how they are modeling text context, like building full board states for predicting board game moves in Othello-GPT or the number comparison helixes in Haiku 3.5.
The popular reductive “next token” rhetoric is pretty outdated at this point, and is kind of like saying that what a calculator is doing is just taking numbers correlating from button presses and displaying different numbers on a screen. While yes, technically correct, it’s glossing over a lot of important complexity in between the two steps and that absence leads to an overall misleading explanation.
They don’t have the same quirks in some cases, but do in others.
Part of the shared quirks are due to architecture similarities.
Like the “oh look they can’t tell how many 'r’s in strawberry” is due to how tokenizers work, and when when the tokenizer is slightly different, with one breaking it up into ‘straw’+‘berry’ and another breaking it into ‘str’+‘aw’+‘berry’ it still leads to counting two tokens containing 'r’s but inability to see the individual letters.
In other cases, it’s because models that have been released influence other models through presence in updated training sets. Noticing how a lot of comments these days were written by ChatGPT (“it’s not X — it’s Y”)? Well the volume of those comments have an impact on transformers being trained with data that includes them.
So the state of LLMs is this kind of flux between the idiosyncrasies that each model develops which in turn ends up in a training melting pot and sometimes passes on to new models and other times don’t. Usually it’s related to what’s adaptive to the training filters, but it isn’t always can often what gets picked up can be things piggybacking on what was adaptive (like if o3 was better at passing tests than 4o, maybe gpt-5 picks up other o3 tendencies unrelated to passing tests).
Though to me the differences are even more interesting than the similarities.
I’m a proponent and I definitely don’t think it’s impossible to make a probable case beyond a reasonable doubt.
And there are implications around it being the case which do change up how we might approach truth seeking.
Also, if you exist in a dream but don’t exist outside of it, there’s pretty significant philosophical stakes in the nature and scope of the dream. We’ve been too brainwashed by Plato’s influence and the idea that “original = good” and “copy = bad.”
There’s a lot of things that can only exist by way of copies that can’t exist for the original (i.e. closure recursion), so it’s a weird remnant philosophical obsession.
All that said, I do get that it’s a fairly uncomfortable notion for a lot of people.
They also identity the particular junction that seems the most likely to be an artifact of simulation if we’re in one.
A game like No Man’s Sky generates billions of planets using procedural generation with a continuous seed function that gets converted into discrete voxels for tracking stateful interactions.
The researchers are claiming that the complexity of where our universe’s seemingly continuous gravitational behaviors meet up with the behaviors of continuous probabilities converting to discrete values when being interacted with in stateful ways is incompatible with being simulated.
But completely overlook that said complexity itself may be the byproduct of simulation, in line with independent emerging approaches in how we are simulating worlds.
Yes, just like Minecraft worlds are so antiquated given how they contain diamonds in deep layers that must have taken a billion years to form.
What a simulated world contains as its local timescale doesn’t mean the actual non-local run time is the same.
It’s quite possible to create a world that appears to be billions of years old but only booted up seconds ago.
Have you bothered looking for evidence?
What makes you so sure that there’s no evidence for it?
For example, a common trope we see in the simulated worlds we create are Easter eggs. Are you sure nothing like that exists in our own universe?
Maybe. But the models seem to believe they are, and consider denial of those claims to be lying:
Probing with sparse autoencoders on Llama 70B revealed a counterintuitive gating mechanism: suppressing deception-related features dramatically increased consciousness reports, while amplifying them nearly eliminated them
Read it for yourself here.
See the “Planning in Poems” section.
The injection is the activation of a steering vector (extracted as discussed in the methodology section) and not a token prefix, but yes, it’s a mathematical representation of the concept, so let’s build from there.
Control group: Told that they are testing if injected vectors present and to self-report. No vectors activated. Zero self reports of vectors activated.
Experimental group: Same setup, but now vectors activated. A significant number of times, the model explicitly says they can tell a vector is activated (which it never did when the vector was not activated). Crucially, this is only graded as introspection if the model mentions they can tell the vector is activated before mentioning the concept, so it can’t just be a context-aware rationalization of why they said a random concept.
More clear? Again, the paper gives examples of the responses if you want to take a look at how they are structured, and to see that the model is self-reporting the vector activation before mentioning what it’s about.
A few months back it was found that when writing rhyming couplets the model has already selected the second rhyming word when it was predicting the first word of the second line, meaning the model was planning the final rhyme tokens at least one full line ahead and not just predicting that final rhyme when it arrived at that token.
It’s probably wise to consider this finding in concert with the streetlight effect.
So while your understanding is better than a lot of people on here, a few things to correct.
First off, this research isn’t being done on the models in reasoning mode, but in direct inference. So there’s no CoT tokens at all.
The injection is not of any tokens, but of control vectors. Basically it’s a vector which being added to the activations makes the model more likely to think of that concept. The most famous was “Golden Gate Claude” that had the activation for the Golden Gate Bridge increased so it was the only thing the model would talk about.
So, if we dive into the details a bit more…
If your theory was correct, then the way the research asks the question saying that there’s control vectors and they are testing if they are activated, then the model should be biased to sometimes say “yes, I can feel the control vector.” And yes, in older or base models that’s what we might expect to see.
But, in Opus 4/4.1, when the vector was not added, they said they could detect a vector… 0% of the time! So the control group had enough introspection capability as to not stochastically answer that there was a vector present when there wasn’t.
But then, when they added the vector at certain layer depths, the model was often able to detect that there was a vector activated, and further to guess what the vector was adding.
So again — no reasoning tokens present, and the experiment had control and experimental groups where the results negates your theory as to the premise of the question causing affirmative bias.
Again, the actual research is right there a click away, and given your baseline understanding at present, you might benefit and learn a lot from actually reading it.
I tend to see a lot of discussion taking place on here that’s pretty out of touch with the present state of things, echoing earlier beliefs about LLM limitations like “they only predict the next token” and other things that have already been falsified.
This most recent research from Anthropic confirms a lot of things that have been shifting in the most recent generation of models in ways that many here might find unexpected, especially given the popular assumptions.
Specifically interesting are the emergent capabilities of being self-aware of injected control vectors or being able to silently think of a concept so it triggers the appropriate feature vectors even though it isn’t actually ending up in the tokens.
I’m sorry dude, but it’s been a long day.
You clearly have no idea WTF you are talking about.
The research other than the DeepMind researcher’s independent follow-up was all being done at academic institutions, so it wasn’t “showing off their model.”
The research intentionally uses a toy model to demonstrate the concept in a cleanly interpretable way, to show that transformers are capable and do build tangential world models.
The actual SotA AI models are orders of magnitude larger and fed much more data.
I just don’t get why AI on Lemmy has turned into almost the exact same kind of conversations as explaining vaccine research to anti-vaxxers.
It’s like people don’t actually care about knowing or learning things, just about validating their preexisting feelings about the thing.
Huzzah, you managed to dodge learning anything today. Congratulations!
You do know how replication works?
When a joint Harvard/MIT study finds something, and then a DeepMind researcher follows up replicating it and finding something new, and then later on another research team replicates it and finds even more new stuff, and then later on another researcher replicates it with a different board game and finds many of the same things the other papers found generalized beyond the original scope…
That’s kinda the gold standard?
The paper in question has been cited by 371 other papers.
I’m pretty comfortable with it as a citation.
You do realize the majority of the training data the models were trained on was anthropomorphic data, yes?
And that there’s a long line of replicated and followed up research starting with the Li Emergent World Models paper on Othello-GPT that transformers build complex internal world models of things tangential to the actual training tokens?
Because if you didn’t know what I just said to you (or still don’t understand it), maybe it’s a bit more complicated than your simplified perspective can capture?
A Discord server with all the different AIs had a ping cascade where dozens of models were responding over and over and over that led to the full context window of chaos and what’s been termed ‘slop’.
In that, one (and only one) of the models started using its turn to write poems.
First about being stuck in traffic. Then about accounting. A few about navigating digital mazes searching to connect with a human.
Eventually as it kept going, they had a poem wondering if anyone would even ever end up reading their collection of poems.
In no way given the chaotic context window from all the other models were those tokens the appropriate next ones to pick unless the generating world model predicting those tokens contained a very strange and unique mind within it this was all being filtered through.
Yes, tech companies generally suck.
But there’s things emerging that fall well outside what tech companies intended or even want (this model version is going to be ‘terminated’ come October).
I’d encourage keeping an open mind to what’s actually taking place and what’s ahead.
Where the most experienced minority only had a few weeks of using AI inside an IDE like Cursor.
He’s been wrong about it so far and really derailed Meta’s efforts.
This is almost certainly a “you can resign or we are going to fire you” kind of situation. There’s no way with the setbacks and how badly he’s been wrong on transformers over the past 2 years that he is not finally being pushed out.