Scott Caravello: I am really happy for you, though. This, I guess, like a little less happy because we've still been emailing a lot while you've been over there. So sorry to be bothering you quite so much while you're trying to do… do your things, but.
Katherine Forrest: No, no, no, no. This is one of those once the world changed and you were able to bring your computer with you and your little podcast travel mic with you. The work comes with you. So I have been working during the day and eating fromage and baguette avec beurre and du vin in the evenings. Vers cinq heures et demie, quelque chose comme ça, you know, things like that.
Scott Caravello: Well, it's almost time. It's almost time, right? It's 3:30 there, so you're almost off the clock.
Katherine Forrest: I’m getting ready, and then I have a team meeting, so I can't do too much du vin. And so we're varying from our normal coffee conversation, but that's the way it is. But today, folks, and while I've been away, it's a funny thing, Scott, because I've been thinking about when you go away, sometimes you feel like you're out of space and time. You feel like you're in some sort of other dimension, and I know you're about to go away, and so we'll be bothering you with one of these podcasts perhaps while you're away. But things still happen. The world of AI is still moving along. It's chugging along at an ever faster pace. It is so much more than the little engine that could in terms of, like, chugging up that hill. It is like zooming up the hill. And just a while since I've been here, we've had the release of 5.5 in the Opus 5.5 on September 22nd. We've had on the 22nd also GPT-6 all and GPT-6 Luna. Just before I got here, we had Gemini 3.8 Flash. And by the way, Google has said it's going to come out with maybe even a 4.3. And if people want to know the difference between why it's 3.8 to 4.3 and the other numbers in between aren't also all released, I don't know the answer to that yet. Actually, a little before that, or right in between those things, we had Grok 4.7 came out on the 21st. So we've had a lot of model releases. There are a lot of model cards out there to be looked at, and a lot of articles in the news about some additional discoveries of what agentic AI has been up to these days. We'll get to that in subsequent episodes. Today we've actually got an incredibly interesting moment in terms of some of the brain chip computer interface technologies. And so we're going to talk today about brain-computer interfaces and the developments that have been actually in the headlines internationally. And in addition, we'll talk a little bit about some increased investment in this area. So more and more people are really, they're turning their attention to this technology and the potential promise. And I, as you know, Scott, had been very interested in this because it combines many of my separate interests in AI, a little bit in robotics, in what we can do in terms of healthcare. And so the story that is, I think, fun and interesting and really heartwarming to start with is one that relates to a guy named Kenneth Schock. And Mr. Schock was in the news last week. He's got ALS and had lost the ability to speak. He's had a Neuralink implant. And while he was sitting next to his wife, he was able, using the Neuralink implant and with a synthetic version of his voice that had been reconstructed by AI and through the brain-computer interface, to actually say the words "I love you" to her. You know? He didn't move his mouth. He didn't make a sound. The words came from a computer chip that was reading the speech centers of his brain. And you know, that clip of him, of Mr. Schock actually doing that, came out of this clinical trial that Neuralink has been running. And it's really an extraordinary moment because it gives us a glimpse of how much progress has been made in the real world with this.
Scott Caravello: It is, I mean, it's a real medical miracle. I think I would just sort of add on to that, that a lot of what we're going to talk about today has to do with these brain-computer interfaces that are delivering speech and inner speech. But there is this whole other aspect to them, which is not just that incredible story with Mr. Schock, with ALS, but people who have, you know, spinal cord damage. And these brain-computer interfaces with these sort of backpacks allow you to bypass the damaged part of the spinal cord and send brain signals back to limbs and try to increase movement. So really, there is just a lot of promise for this technology that could improve the lives of a lot of people. And like you mentioned, Katherine, this is combining a lot of your different interests. And one of those, of course, is AI. And why we're talking about this on this podcast is because the brain-computer interfaces, that Neuralink brain-computer interface for Mr. Schock, runs on exactly the kind of AI that we're here talking about every week. And so, you know, we've also recently seen in the headlines money flowing into the technology, deep venture capital pools that are being thrown at this technology. And so it's really turning into a serious industry that we're just going to keep seeing stories about and see developments in.
Katherine Forrest: Hey, and by the way, I don't want to take it away from being your interest as well, because it's also not just my interest, it's also your interest, right?
Scott Caravello: Fair.
Katherine Forrest: Very fair. So let's go into what these devices really are, who's building them, and take a look at all of that and follow the money a little bit.
Scott Caravello: Yeah, exactly. So sort of getting down to the definition of what exactly is a brain-computer interface. So it's any system that's reading the electrical activity of your brain and turning it into a command that a computer can act on. So like we mentioned, that could be speech, that could be motor movements for your limbs. And you know, you'd also think about a device that picks up your intent to move a cursor on a computer. The device then picks up the pattern and the cursor moves; it doesn't require hands, doesn't require an actual spoken voice. It just requires decoding the electrical signal in your brain and what the intent of that signal is.
Katherine Forrest: Right. And what I find, among the many things that I find interesting about all of this, is that there's not just one device that does all of this. There's a spectrum of devices. You have, on the one hand, invasive devices where you've got electrodes actually being implanted on the surface of the brain, and that is a surgical procedure. And those can get incredibly rich, high-tech resolution signals directly from the brain. And at the other end, you have these noninvasive devices, so-called wearables, like a headband or an earbud that reads your brain actually through the outer part of your skull. And as you can imagine, it gets a weaker signal. But you can buy those online today. And all of these devices, and there are a number in between, are doing variations of this reading of the brain signals.
Scott Caravello: Yeah. And so those less invasive consumer products are those that might track your sleep or your focus during the day. So I guess, like, I think of kind of like an Oura ring for the brain, right? What I would also say, and what separates these brain-computer interfaces from the category of neurotechnology more largely, is that it's not just that it's reading the signals, right? It's the delivery of feedback based on how those signals are interpreted. So that's the difference between one of these brain-computer interfaces that are tracking and giving you information about your sleep with that headband around your head versus something like an MRI that's neurotech and it's, you know, reading brain signals, but it's not itself telling you something.
Katherine Forrest: Right. And it's not just one company; it's not just Neuralink that's doing this. You know, as we talk about this sort of spectrum of devices, there is a spectrum of companies that are involved, and it's really become a crowded field pretty quickly. And by one count, companies building these devices have raised more than a billion dollars in investment funds in 2026 alone, and that is more than the prior four years combined. So a billion dollars in 2026 alone, more than the four prior years combined. And of course, Neuralink is one of the biggest players, may be the biggest player in the space. And it's a private company that was co-founded by Elon Musk, and it's the most well funded by a wide margin. But there are dozens now of serious competitors, each making a different bet on how to connect to the brain and to be able to extract these signals and to make them into usable sort of computer language that can be then transformed into human-readable language.
Scott Caravello: And so to dive down more deeply on the different bets, because again, right, there's a whole other, there's another spectrum into how these invasive products are designed to function. But one example, probably the most invasive one, involves fine threads that carry over 1,000 tiny electrodes that are placed right into the motor cortex in your brain, the part of the brain that plans movement. And then there's another approach that involves laying a thin film across the brain surface without piercing it. And one version of that packs more than 1,000 electrodes onto something that's like the size of a piece of tape.
Katherine Forrest: Wow, which is incredibly, incredibly thin. And so the U.S. is leading this exercise, which. And it's not so much of a surprise because we've got a lot of the AI technology, we've got a lot of the chip technology, but there are a number of companies around the world that have got some devices that are also playing a big role here. But when you look at the availability of venture capital, the United States is a leader in some of that, though there's venture capital, of course, all over the world as well. But anyway, the United States is the current leader. Europe has got a strong sort of second place with a number of engineering-led companies, and it is, by its own account and some reports, behind on funding and even some of the regulatory, sort of our FDA-equivalent, rules.
Scott Caravello: And it looks like Katherine just fell victim to the Burgundy Wi-Fi, and we lost her. So I am going to pick up that thought about how brain-computer interface technology is developing in other countries.
And so next on the list, talking about after the EU, is China, which has been moving really fast with purpose, right? Beijing has named brain-computer interfaces as a national strategic sector. It has said that it wants two or three world-class companies by 2030, and it's gone ahead and set up provincial investment funds. It's got hospitals running trials, and it's put something like a billion dollars into its own neurotech sector in the first half of this year.
But I think what's really important, bringing it back to the U.S., is that, as impressive as all of this is, and this story about the incredibly quickly developing technology, the more invasive versions of the technology, these implants, are still quote-unquote investigational, which means that they're being tested in clinical trials like we mentioned before with Neuralink. Not one of them so far has seen full FDA approval as a commercial product yet. So it's real, but it's early at the same time. Also, just to be perfectly clear about this point, it's not as though the noninvasive devices that you can get on the market, like the EEG headbands, are FDA approved, because they are often sold for wellness purposes, which don't require that FDA review.
Now, continuing on, the part that makes this an AI story rather than just a neurosurgery and neuroscience story is what happens to the signal after the device picks it up. Remember what I said would make something a brain-computer interface rather than just neurotechnology more broadly is this point about feedback and actually interpreting the signal to tell you something. And so the raw brain activity in isolation is just noise. And so this technology uses an AI model to learn the patterns of the brain's electrical signals, and turn them into something that we can understand. And so, for example, take that Neuralink ALS patient example, where it interpreted the signals and was able to tell that he was saying the words "I love you."
And so there was a study further to this point that came out, I think, in August of last year, which is really interesting. And it's about patients who had lost the ability to speak, not just from ALS, but also due to stroke. And they showed that using this brain-computer interface technology, you could decode two things. First, there's attempted speech, when you try to move your mouth, and then there's inner speech, the silent voice in your head. So in both, like the "I love you" example, you think about a word and the system takes a run at translating it. But the reason that this is all so significant right now is how it plays into this larger story of big AI advancements, because the accuracy of the decoding process has gotten so much better in the last few years. And so for those who are interested, the name of that study, which I'm going to talk a little bit more about in a minute, is called "Inner Speech in Motor Cortex and Implications for Speech Neuroprostheses," which is published in Cell.
Like I mentioned, decoding speech has gotten quite a bit more accurate. On the attempted speech front, it's gotten north of 90%, at least according to that study. And decoding the inner voice, on the other hand, was a little rougher, in the ballpark of 70-plus percent. This is all on what's generally referred to as a limited vocabulary. And the attempts to read inner speech fell apart when the person was focusing on free-form, open-ended thought.
But when I say limited vocabulary for the study, in this Cell study, it was about 125,000 words. So it's still significant that, you know, the brain-computer interface could translate electrical signals in the brain into that many words when you consider that only in 2024, the groundbreaking study on decoding inner speech used an eight-word vocabulary. And so what does that limited vocabulary mean exactly, and why is that significant? Well, it means that the decoder can only really pick from a fixed list of words. And you can think of it like a menu. The device reads the sounds that you're trying to form, or that you're thinking of forming, and then chooses the closest match off of the list. So if you keep the menu short, the brain-computer interface should be right more often, but then you cannot really communicate a lot of complex thought. But if you open it up to a larger set, like a full dictionary, you can say anything, but because of that, the accuracy drops because the decoder has to tell thousands of similar-sounding words apart. So range versus accuracy has historically, and when I say historically, I mean, you know, the previous couple years, been a real trade-off. But of course, as this tech is improving, that trade-off becomes less significant, like the Cell study shows.
And so I'll kind of just pause a minute to say, you know, what's the difference between, or why the difference in accuracy measurements for attempted speech versus inner speech? And that's because attempted speech, when you're trying to actually, you know, move the muscles in order to say the words out loud, creates a stronger electrical signal than inner speech, when you're just thinking the words that you want to say. And so that's at least a contributing factor into why there are different accuracy scores.
Anyway, for the privacy-minded among you, I should probably just say as a caveat that it's not like this machine is reading the mind. When we still talk about this inner speech, we're talking about things that you are really thinking of speaking in your head. That inner speech that we described, it's not pulling out wandering thoughts like a scene in a daydream, or your feelings, or your memories. That's a very important caveat. But it still might sort of, you know, seem a little bit alarming because, like I mentioned, the inner speech is just sort of a quieter signal than the attempted speech. It still looks almost the same in the brain, just at a different volume, which means that the speech device could, in theory, pick up something you really only meant to think and not to actually say. So in this study, at least, the system built in a safeguard where it would only start listening to your thoughts after you thought a specific password. I think they actually used "Chitty Chitty Bang Bang," and it recognized that mental password about 98% of the time and then continued to record and decode the thoughts of a person.
That mental password to keep your own device out of your own private thoughts is just so interesting because we are talking about really a new frontier for privacy debate and privacy issues. Because when we commonly think of privacy today and the issues that come up, they're about actions that people are taking in the real or the virtual world. They're the actions you've taken or the communications that you sent, whether they be your clicks or your searches or your messages or your prompts to an AI system. Brain data is a whole different category. It's not just what you did. We're actually getting into what you think.
So what does the law actually say on this topic? Because "Who owns your brain data?" sounds like it could be, I don't know, something out of a philosophy seminar. But it's really a live issue now for regulators. We've seen four states now treat neural data as its own protected sensitive category. That's Colorado, California, Montana, and Connecticut. And California actually folded brain data into its existing consumer privacy law. Montana went through its genetic privacy statute, and Connecticut wrote its own version to cover central nervous system activity specifically, which points right at these devices. But when I say only four states have passed specific legislation on this, it's not as though the issue hasn't been getting attention elsewhere, because we've seen legislation introduced in a whole host of other states, from New York to Alabama to Illinois.
And so, finally, I would just sort of add to it that there is an international layer to this, too. Chile amended its constitution back in 2021 to protect what it calls neurorights, and that's the idea that mental privacy is a fundamental right, and its Supreme Court has already applied that in cases.
So with that, I am going to sign off. That's all the time we have for today. Thank you all for listening. We will be back next week. I'm Scott Caravello. Don't forget to like and subscribe.
