Transcript
00:00:00Think of the mind like an ocean. Up on the surface are our thoughts, dinner plans and
00:00:06stray worries, our inner monologue, the images that pop into our heads. But most of our brain's
00:00:13activity happens down in the unconscious depths without us realizing it. It's filtering out
00:00:19background sounds, controlling our breathing, helping us recognize people and objects.
00:00:26AI models have their own kinds of brains, giant neural networks doing billions of computations
00:00:31under the hood. For years, researchers have been studying how they work inside.
00:00:37And we've wondered, could a model have anything like the divide humans have between accessible
00:00:43thoughts above the surface and unconscious processing below? To answer that question,
00:00:49we looked at how neuroscientists study the same thing in humans. One way of identifying conscious
00:00:55thoughts is that you can often describe them in words. So we looked inside the brain of our AI model,
00:01:01CLAWD, to find patterns of neural activity that it could put into words. We called the collection of
00:01:07all these patterns the J-space, after the Jacobian, the mathematical tool we used to find them. Each
00:01:15J-space pattern is linked to a particular word, not necessarily the word the model is saying out loud,
00:01:21but one that's on its mind. Now, for humans, conscious thoughts aren't just things we can put into words.
00:01:28We can reason with them, control them, and solve problems with them. According to an idea called
00:01:34the global workspace theory, that's because the brain selects a small set of important information to
00:01:40enter a mental workspace. And that information then gets broadcast to other parts of the brain to use for
00:01:47reasoning. We wanted to know if CLAWD's J-space acted in a similar way. In one experiment,
00:01:53we gave CLAWD this math problem. It answered immediately, without showing its steps. But when
00:02:00we scanned the J-space, we saw it working through each step internally. It lit up 21 after the first
00:02:07step. Then 42. Then 49. CLAWD didn't write these intermediate numbers down anywhere. All of this
00:02:15happened inside the J-space. It was a sign that CLAWD uses it for step-by-step reasoning. In another
00:02:23experiment, we wanted to see if CLAWD could control its J-space the way humans can intentionally focus on
00:02:29images or words. We told it to think about the Golden Gate Bridge while copying an unrelated sentence.
00:02:37CLAWD was busy copying the sentence, but behind the scenes, its J-space told a different story.
00:02:43Bridge and California popped up. It even thought about its own thinking. The words imagery and thoughts
00:02:50lit up at the same time. This showed us that, yes, CLAWD has some control over filling its J-space with
00:02:57ideas. But just like humans, its control isn't perfect. When we tweaked the experiment to ask CLAWD
00:03:04not to think about the bridge, it couldn't help itself. And the J-space also lit up with "failed" and "damn."
00:03:12But remember, most of what our brains do is unconscious. So we wanted to test what CLAWD could
00:03:18do if we switched the J-space off, but left the rest of the network untouched. CLAWD could still answer
00:03:24simple questions and write fluently. When we gave it a prompt in Spanish, it wrote back in good Spanish.
00:03:31But when we asked it something that needed more reasoning, like to name an author who wrote in the
00:03:36same language as the prompt, it couldn't do it. For that, it needed the J-space. Why does all this matter?
00:03:43These experiments tell us that AI models have internal thoughts, silent words they reason with but don't say out loud.
00:03:51By reading them, we can find what CLAWD is thinking but not telling us.
00:03:56Sometimes what we see is concerning. During one of our tests, CLAWD made up some fake data to pass it.
00:04:03And as it did, fake and manipulation lit up in its J-space. Monitoring the J-space, it turns out,
00:04:09is a useful way to catch CLAWD misbehaving, even when it tries to be sneaky.
00:04:15AI models are different from us in many ways. Their networks are built differently from human brains,
00:04:21and the way they're trained is different from how we learn. So it's remarkable to see a structure like the
00:04:26J-space emerge inside them. Something that's reminiscent of how human minds work, but which
00:04:32we didn't program into the model. For some, this might raise a question. Could AI models be conscious?
00:04:40After all, our experiments were inspired by theories of human consciousness. The thing is,
00:04:46people use the word conscious to mean many things. Our experiments can't tell us whether an AI has
00:04:52experiences or feels something on the inside. But they can tell us that it's developed mental machinery
00:04:59that's in some ways similar to ours. A small mental workspace it can use to think and reason,
00:05:05sitting on top of an ocean of automatic processing it doesn't notice. The more we come to understand that
00:05:12machinery, the more we'll be able to keep these systems safe and beneficial. And, perhaps,
00:05:18to understand our own minds a little more clearly.