Working Memory or Global Workspace
Looking at Anthropic’s J-Space paper
Imagine, if you would, the humble transformer architecture for large language models. Only a few years ago it was counted amongst the most revolutionary technologies the world had ever seen, but now it has already become as standard as smart phones and the internet. It seems as if everything these days has an LLM built into it somehow for some reason, and every tech-company and start up is making their own new model based on transformers.
Yet, as seemingly mundane as this machine has become, the academic and tech communities keep finding out new things about it. Not only is a black box of linguistic wonders, but one that keeps surprising us. It may be the first technology that was intentionally designed from the ground up by us, to fulfill a specific purpose detailed by us, yet a complete mystery to us.
The latest light that has been shone on the mysterious inner workings of this magical black box is by Anthropic, who showcased their J-lens technique to reveal what a transformer thinks about in its middle layers. It is an amazing tool to see not only what the LLM did say (or will say), but also what it could have said. This is (and I am simplifying this a lot) because the J-lens looks at the activations and vectors within a given layer in the transformer, compares this to a list of tokens, and then calculates what are the odds of each token being the ultimately chosen one at the end of the processing.
This is all well and good, but when you create a new way to look at the quasi-thoughts of a machine, people start pondering and pontificating about whether these are quasi-thoughts or real thoughts, and whether the container of thoughts ought to be called a mind, and (most pertinent to this Guidebook entry) whether this space uncovered by the J-lens is a global workspace ala the Global Neuronal Workspace Theory of consciousness.
By the time you find this Guidebook entry, everyone will already have had their say on this topic, including the creator of the Global Neuronal Workspace Theory. However, whilst many people debate about the plausibility of AI minds, or what a global workspace may mean for LLMs vis-a-vis consciousness, the Guidebook entry will focus on why everyone else is (mostly) wrong.
The reason is that the J-space is very clearly not a global workspace. In the depths of the Anthropic paper, the authors admit this, stating only that the J-space shows “workspace-like” features. Unfortunately, since most people only read tweets and headlines, most people won’t ever see this crucial distinction. There is a very real chance that more and more people, papers, and conferences will begin to talk about the J-space as proof of a global workspace which, in turn, is proof of consciousness. Early commentators have already pointed towards this as a type of access consciousness, even if it isn’t phenomenal consciousness.
But is there a better explanation and description of what is going on? Absolutely! J-space is much better explained as a working memory than as a global workspace, for several reasons. Firstly, a global workspace requires parallel sub-units within the architecture competitively submitting packets of information to a central workspace that will only select one packet at a time and broadcast it back to the subunits to be processed; after which, the process starts all over again.
This recurrence and parallel processing are impossible in standard transformer models. For those subscribers to the Guidebook’s Friends Circle, you’ll know that I have developed this parallel recurrence in an ensemble model, but it simply cannot exist within standard transformer architecture. This is because transformers are feedforward systems that push information only in one direction; the information never comes back for a second round (within one standard processing cycle, mind you). Each layer within a transformer does its matrix multiplications, normalisations and stabilisations and then pushes that forward to the next layer. It doesn’t go back to a previous layer for another round of processing before it gets to the output.
This means that there isn’t a single J-space that sits above or below the layers that they feed into and that broadcasts information back to them. The J-space is just a term used for all the spaces in between each layer, a term used and modified in a way to translate the mathematical vectors and calculations of the Jacobian lens into human-readable terms.
So, there absolutely isn’t a workspace here, global or otherwise, but the processing revealed by the J-lens does fit what a working memory is supposed to be doing, which is holding onto transient information while it is being processed, even if that information is never used as output.
When the J-lens does its calculations on the vectors and vocabulary, it shows that not only words which could be reasonably used for output, but all sorts of concepts related to the information being processed, and even intermediary reasoning steps that are required for the final output. What this shows is that the LLM doesn’t simply predict the next token, but that there is a process from the first layer all the way to the output layer that reasons about the information it perceives, what direct and abstract concepts relate to it, what the implications and inferences are about this information, and what and how the LLM should respond to it.
Simply knowing that all this goes on inside the black box is gold to philosophers of mind, but, on a practical level, this is how a working memory should be operating. When you are reading this very sentence, your brain is doing an analogous process to what I just described above. You are reading the words, yes, but you are reading in between the lines looking for implications, drawing inferences, thinking (consciously or unconsciously) about how this sentence relates to the rest of the text, to other things you have read, to memories of the past and predictions of the future.
And all of this is held stable by your working memory. You can do all of these things because your working memory holds onto all those thoughts and information and memories and reasoning while you read this. This is the analog to the J-space. The J-space does not do the processing. The layers of the transformer do. The J-space (really, the J-lens) just shows us how this information is kept stable as it moves through all the layers, ready to be used by the next layer, until it is finally output as a token the user reads.
So, what does all of this mean? Nothing and everything, all at once. Whatever you choose to call the J-space doesn’t change what it is or what it does, nor does its name change the vital importance it has to philosophers of mind and to machine learning scientists. However, what you call it changes how you view it and what you do with the implications. Calling it a global workspace means that you are relating to the J-space in terms of consciousness. Calling it a working memory means that you are thinking of it in terms of cognition.
What we have is a cognitive process. If we overestimate its value to consciousness, we will end up a (quickly debunked) false-positive that will only hurt future attempts at finding evidence for artificial consciousness. The terms we use matter and we shouldn’t simply use the wrong terms because it gets us more engagement on social media.