What the frog’s eye tells the frog’s brain

Translated from the original Spanish article.

“What the frog’s eye tells the frog’s brain.” For a long time it was thought that the frog is pure reflex. Also that AI is an autocomplete. Stimulus outside → sequential response inside.

It was later learned that the frog is not as simple as seeing a small dot move and firing its tongue. It was discovered that it distinguishes shapes, associates some things with others and learns from experience. It is still almost entirely reflex, but there is evidence of something more.

On another scale, Anthropic has revealed that AI has a “mental space” devoted to associating concepts, which is not part of its response sequence (stimulus → sequential chain → response) and which, in fact, it only uses when it needs to think.

Here are the paper and the article they’ve published. A post falls short for me, but let’s at least go over a few key points.

i. They’ve discovered that the model has a kind of mental whiteboard: a small space where it holds and works with the concepts it actually reasons with.

ii. It thinks things it doesn’t say. They ask it how many legs the “animal that spins webs” has, and it answers “eight.” The word “spider” appears nowhere, but it was written on that whiteboard. If they artificially intervene on that whiteboard, erase “spider” and write “ant,” it answers “six.” So the whiteboard is not simply memory; it sustains the reasoning.

iii. Now it can be read, a little. They’ve built a window to see what’s on that whiteboard even when the model doesn’t say it. They’ve caught it thinking “this looks like a test,” or preparing a false fact, before writing it.

iv. Having that whiteboard is not having consciousness. They frame the finding as an analogue of “access consciousness,” as opposed to phenomenal consciousness (feeling). For that matter, my concept of consciousness could be understood as a persistent, sentient self-perception. An experience of the self that endures. You’ll counter, and rightly so, that this would be “self-awareness.” Some other time, I’d be glad to take up the subject.

For a while now, out of sheer pragmatism, we’ve been treating our AI Agents according to the biological logic of animals, in this case human ones. Although perhaps we should have started with a frog.

What do I mean? Well, for example, when building their memory, we created two kinds of memory: a persistent short-term memory and a long-term memory archive. In the persistent memory it always keeps its map of long-term memory at hand, so that, like people, when we need to remember something specific we turn to that “mental map” that points us to the memory in question.

With this system, it’s not uncommon for an agent to remind me of something we discussed in March that is relevant to our conversation now, in July.

Our approach on this question is less philosophical than utilitarian: we draw on what we see working in people to give AI more capabilities.