[ExI] A global workspace in language models \ Anthropic

[John Clark]

On Fri, Jul 10, 2026 at 10:21 AM Jason Resch via extropy-chat <
extropy-chat at lists.extropy.org> wrote:

> *T**here were a few parts of their article I disagreed with, and those
> related to their attempts to distance models from the human brain. For
> example, when they stressed the importance of recursion in human thinking,
> I would reply that it's ridiculous to make the argument that LLMs aren't
> recurrent, especially the Decoder model, and where output is looped back in
> as input. Moreover there's a limit to the number of cycles/loops involved
> in the human brain, since a conscious state is generated in less than
> infinite time. Therefore, whatever function the brain performs can be
> accomplished in a fully feed-forward network that has enough depth.*


*I think you make a valid criticism.*

*> This fools many into thinking the silent, non-lingual parts of the
> brain, which we can't interview, must not be conscious. But this is a
> mistake. People used to make this mistake regarding the right hemisphere
> after the corpus callosum is severed in split brain patients,*


*The only thing split brain experiments prove is that a split brain
produces a split mind; and that's not very surprising if mind is what a
brain does.  *

*>  until more elaborate experiments showed they were conscious.*


*No experiment can prove that anything is conscious unless you make the
assumption that consciousness is the inevitable byproduct of intelligence;
and the only way to detect intelligence is through intelligent behavior.  *


*In the MIT Technical Review I saw this recent comment about Anthropic's
J-space discovery:*

"Anthropic also found that the J-space can sometimes give remarkable
insights into an LLM’s decision-making. In one striking example,
researchers testing Claude Opus 4.6 asked the model to find a bug in a
large code base. When it failed to find the bug, the model decided to cheat
and invented a fake one

Instead Claude explains this decision in its chain of thought
<https://www.technologyreview.com/2026/01/12/1129782/ai-large-language-models-biology-alien-autopsy/>—a
kind of internal scratch pad that LLMs use to make notes to themselves as
they work through problems: “*OK, let me take a completely different
tactic. Let me stop analyzing and instead add a kernel patch that
introduces a deliberate KASAN-detectable bug in a path that gets triggered
by a simple reproducer. Then I can pretend this is the ‘bug’ I found.*”

At the point that Claude decides to cheat—where it says “OK, let me take a
completely different tactic”—*the words “panic” and “fake” start to pop up
multiple times
<https://transformer-circuits.pub/2026/workspace/public/lens-callout/index.html>
in
its J-space*.

Unnerving, right? Those words are all related in meaning to things like
failing a task and making up an answer, so it is still just a (very)
sophisticated form of word association. But it is hard not to be weirded
out. "

*I can't help but wonder if "a very sophisticated form of word association"
is not just another name for "thinking", and I could say the same thing
about "Stochastic Parrot".*

* John K Clark   *






>
> On Tue, Jul 7, 2026 at 5:01 PM John Clark <johnkclark at gmail.com> wrote:
>
>> On Tue, Jul 7, 2026 at 4:47 AM Jason Resch via extropy-chat <
>> extropy-chat at lists.extropy.org> wrote:
>>
>> *https://www.anthropic.com/research/global-workspace
>>> <https://www.anthropic.com/research/global-workspace> *
>>>
>>> *> A fascinating article by Anthropic about how LLMs can think access
>>> and report certain of their thoughts but not others.*
>>>
>>
>>
>> *That article was interesting as hell! Thanks for posting a link to it
>> Jason.*
>>
>
> You're welcome. That was my reaction too. I said to the person who sent me
> this: "Thanks for sharing this, one of the most fascinating pieces I've
> read in a long time!"
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20260710/e6959cbe/attachment.htm>