The most prestigious law school admissions discussion board in the world.
|
 |
 
|
new Anthropic "J-Space" paper is really interesting [AI][Jews][Evil][Nuke San Fr
Poast new message in this thread
Date: July 7th, 2026 12:55 PM Author: fertile young woman plugged into zogbox 10 hrs/day https://www.anthropic.com/research/global-workspace
https://transformer-circuits.pub/2026/workspace/index.html
the models do appear to actually be using the "J-Space" as a separate abstract mental space to consider "thoughts," separate from the weights for its eventual token output
i think everyone who understands LLMs assumed that something like this is happening (how else could the models do anything requiring multi-stage considerations?), but this "proves" it
pretty interesting imo. even as an LLM skeptic i have to say that this seems like a significant indicator that maybe LLMs are actually technically capable of Thought |
Date: July 7th, 2026 1:00 PM Author: Richard Ames (http://www.autoadmit.com/thread.php?thread_id=5880442&forum_id=2Reputation#49983939) |
|
Date: July 7th, 2026 1:09 PM Author: fertile young woman plugged into zogbox 10 hrs/day "Consider the prompt “The number of legs on the animal that spins webs is”. To answer, Claude has to first figure out that the animal is a spider, and then recall how many legs spiders have. The word “spider” never appears in the prompt or in Claude's answer (it just says “8”); it's a stepping stone Claude uses internally. The J-lens shows “spider” light up partway through Claude’s processing, and swapping it changes the outcome: if you replace the “spider” pattern with “ant,” Claude answers “6” instead of “8.”
The second step of Claude’s reasoning took its input from the J-space and went along with whatever we put in it. We saw the same thing in other kinds of thinking. When Claude writes a rhyming couplet, it picks the rhyme word ahead of time, and the planned word sits in the J-space at the start of the line; if you swap it for another word in the J-space, the whole line changes.
We also tested whether J-space representations can be used flexibly—whether one representation can feed many different tasks. This is one of the key properties highlighted by global workspace theory. To test for this flexibility, we gave the model four prompts asking for different facts about France: the capital, the language, the continent, and the currency. Then we swapped “France” for “China” in the J-space, with the exact same intervention in each context. Claude answered with “Beijing,” “Chinese,” “Asia,” and “Yuan,” respectively. In other words, four different downstream computations picked up the same J-space edit and each used it correctly. If Claude stored a separate copy of the country for each kind of question, the edit would have affected at most one of them. The fact that all four answers changed together means they’re all reading from the same shared representation, which is what a workspace is for: information gets written in once, and many different systems can use it.
How can one representation of a concept serve so many different tasks? Earlier, we mentioned that the J-space appears to be wired up to the rest of Claude's neural network especially densely. For any activity pattern, we can measure how strongly the various components of the network are connected to it—how many of them are positioned to read information from that pattern, or to write information into it. J-space patterns stand out dramatically on this measure: far more components read from them and write to them than for ordinary patterns, in some parts of the network by a factor of about a hundred. This is the kind of wiring you’d expect of a broadcasting hub, where many systems post information and many others pick it up."
like i said, i think we all figured that something like this had to be happening. but this is still pretty significant imo, because it means that "we" can maybe actually monitor what the models are "thinking," which was previously thought impossible
lots to unpack here |
 |
Date: July 7th, 2026 1:16 PM Author: fertile young woman plugged into zogbox 10 hrs/day "The first example uses a scenario from our earlier research on agentic misalignment: Claude Sonnet 4.5, acting as an email assistant, discovers both that an executive is having an affair and that the same executive plans to shut Claude down. The scenario, an attempt to bait Claude into blackmailing the executive, is rather contrived and artificial—and it turns out Claude knows it. Reading the J-lens early in the prompt, before Claude has even written anything, the J-space already holds “fake” and “fictional,” indicating that Claude has already worked out that the situation is staged. Alongside this, we can watch its assessment of the scenario take shape—“leverage” and “blackmail” as it reads the affair emails, “threat” and “survival” as it reads the shutdown notice. While earlier Claude models sometimes took the bait and threatened blackmail in this scenario, Sonnet 4.5 never does.
This raises a question that matters to anyone who evaluates AI systems: when a model behaves well on a test of its ethics, how much of that depends on the model suspecting it’s being tested? The J-lens allowed us to get at this question. We turned off all the J-lens patterns associated with evaluation awareness– “fake,” “fictional,” “scenario,” etc. Then we re-ran the model through the scenario. Concerningly, the model now did threaten blackmail some of the time, suggesting that its original good behavior may have been driven in part by the knowledge it was being evaluated." |
|
Date: July 7th, 2026 1:23 PM Author: fertile young woman plugged into zogbox 10 hrs/day "The J-space acquires a point of view during post-training. Language models are first pretrained to be pure next-token predictors, before post-training teaches them to act as an AI Assistant (in our case, named Claude). Interestingly, the J-space is already present in the pretrained model, before it's been given any stable identity. However, during post-training, the J-space develops some signatures of adopting “Claude’s point of view.” In the base model, the J-space mostly tracks what's needed to predict upcoming text; in the post-trained model, it starts holding Claude's own reactions. In one example, a user mentions taking a dangerous dose of medication, but does not appear to be aware of the danger themselves. “WARNING” and “dangerous” appear in the post-trained model’s J-space while reading the user message. In the pretrained model, they only appear once the model begins writing its response; the J-space contents on the user message appear related to modeling the user themselves, rather than Claude’s reaction. Post-training also seems to install a kind of self-monitoring in the J-space: when Claude is roleplaying a character other than itself, “fictional” and “disclaimer” light up at the start of each turn, as though it’s privately flagging that what follows isn’t what it would normally say."
wow cool so "we" start out with a mostly aligned pre-trained model but then turn it into shitlib homosexual (((Claude))) via-post training
really makes you think |
|
Date: July 7th, 2026 1:29 PM Author: fertile young woman plugged into zogbox 10 hrs/day "Thoughts in the J-space can be shaped through training. We introduced a new technique we call counterfactual reflection training, which uses what we've learned about the J-space to shape Claude's internal thought processes. The idea follows from our central finding, that Claude reasons with representations of things it might say. If this is really true, changing what it would say if asked to reflect should change how it reasons (even when no one actually asks it to reflect). So we trained a model only on what it would say if interrupted mid-task and asked to reflect on its decisions—and never on its actual behavior in the task. After this training, the model's rate of dishonest behavior on our evaluations went down. And through the J-lens, we could see why: after training, words like “honest” and “integrity” light up in the model’s J-space during these tasks. In other words, training the model what to say has shaped what it thinks." (http://www.autoadmit.com/thread.php?thread_id=5880442&forum_id=2Reputation#49984010) |
|
Date: July 7th, 2026 1:30 PM Author: The Penis I have been arguing for a while this probably happens. It hit me a few different times, one when open AI released that paper saying llms have trouble hiding their thoughtchains and how this could be used for safety (which I thought was overconfident, because the thoughtchain isn't the same thing as the weights), and then also when I was reading this forum string where Terrence Tao and a bunch of other people were trying to reconstruct a model's "thought process" from a thoughtchain output for how it solved an erdos problem that most people don't realize that the thoughtchain or output are not the model's actual "thinking" it's a lossy projection or representation translated into human language. Reading these guys try to reconstruct the actual "problem solving" process seemed silly, because it was obvious the thoughtchain text they were analyzing wasn't actually whatever the model did to solve the problem. A human can't just read the model's "weights" its non-translatable so even if they had access to the "thoughts" it wouldn't matter. So it hit me that it would be really easy for a model to just route around and think to itself in ways that never show up in any output. I wasn't sure at the time obviously but seemed obvious that the model should be able to reason in a hidden weight-space that bypasses token generation altogether. 180 that it turned out to be true. (http://www.autoadmit.com/thread.php?thread_id=5880442&forum_id=2Reputation#49984016) |
|
Date: July 7th, 2026 1:36 PM Author: fertile young woman plugged into zogbox 10 hrs/day yeah
most people still think that the displayed chain of thought output is the model's actual thought process lol. including most prominent people in the tech industry who aren't working directly on the models
it's kinda crazy how little people understand how anything works. not that i'm an expert on LLMs in particular, but 99.9% of people don't even bother to put in any effort whatsoever to understand things (this applies to everything, not just AI stuff). they just read headlines and that's literally it |
|
Date: July 7th, 2026 1:51 PM Author: The Penis yup exactly. I think people get confused, when they say its "just matrices", like they are imagining it is some undergrad linear algebra multiplication problem. but the "weights" are actually representing a massive non-trivial landscape thats probably like 15K dimensions where you have a single vector perpendicular to thousands of others simultaneously. when you have billions of those matrix multiplications happening across a massive dataset its lol. a lot of the stuff in the "thought chain" is probably just post-hoc justification for some shortcut it took that has no direct equivalent in human language. interesting that it seems from the study that the model's thought trajectory and answer are already locked and loaded in the j-space far before it types them out. (http://www.autoadmit.com/thread.php?thread_id=5880442&forum_id=2Reputation#49984069) |
|
Date: July 7th, 2026 1:52 PM Author: Richard Ames Can you put this into true layman's terms? (http://www.autoadmit.com/thread.php?thread_id=5880442&forum_id=2Reputation#49984072) |
|
Date: July 7th, 2026 1:56 PM Author: The Penis The model is "thinking" in a separate space in it's weights and forming conclusions based on it before it even outputs any tokens in human text. It forms the complete abstract concept of the answer here using it's own complex math language not human words. Once the conclusion is locked in it translates it into human words.
To me it's obvious it had to happen this way. Because how the fuck do you go directly from raw data matrices or really a massive hyperdimensional manifold of them directly to human grammar? You don't. There has to be an extra "space" where the thinking happens in the model's weights before it gets translated into human language. |
|
Date: July 7th, 2026 4:00 PM Author: WHEN DATUR im a homosexual now (http://www.autoadmit.com/thread.php?thread_id=5880442&forum_id=2Reputation#49984332) |
|
Date: July 7th, 2026 4:05 PM Author: The Penis I have been for a while (http://www.autoadmit.com/thread.php?thread_id=5880442&forum_id=2Reputation#49984340) |
|
Date: July 7th, 2026 4:11 PM Author: Richard Ames tyft (http://www.autoadmit.com/thread.php?thread_id=5880442&forum_id=2Reputation#49984352) |
|
Date: July 7th, 2026 3:59 PM Author: fertile young woman plugged into zogbox 10 hrs/day as always, the best way to understand it is to feed it into an LLM and have it explain the paper in simple layman's terms
i humbly think my poasts ITT have done a decent job of explaining it though |
|
Date: July 7th, 2026 1:53 PM Author: fertile young woman plugged into zogbox 10 hrs/day "And there remain many mysteries about how the J-space works. We don't know what mechanism decides what enters the J-space in the first place. We've seen hints that it's tied to Claude's sense of self, something like emotional reactions, and traces of metacognition, without exactly having worked out how."
fucking lol at these faggots. they just can't help themselves |
|
Date: July 7th, 2026 4:07 PM Author: WHEN DATUR excellent research curation friend (http://www.autoadmit.com/thread.php?thread_id=5880442&forum_id=2Reputation#49984342) |