Not a member of Pastebin yet?
Sign Up,
it unlocks many cool features!
- Why robot teachers were discontinued
- Bernstein:
- This is a recorded interview with Professor Peter Russel, in his home in Los Angeles, California, on April 14, 2026. Harold Bernstein will occasionally interrupt with questions. Let’s start with the origin of the Machine.
- Russel:
- Well, it was innocent enough in the beginning, I suppose. A small weekend project. I wanted to build an auto-testing system for the students in our CS182 class. I had recently gotten access to a large testbank of curated, high quality questions from uh--well, maybe I shouldn't say? Well, anyway, we had some spare GPUs in the Hoffman2 cluster and I decided to train a simple generative network to produce new questions. It was...awful, I mean predictably so, given the limited computing power. The questions were syntactically correct and sometimes even made sense, but the answers were completely nonsensical 90 percent of the time.
- Bernstein:
- These were math questions?
- Russel:
- Right. Mostly probability and some statistical modeling. CS182 is a statistical inference class so--
- Bernstein:
- Did you end up using anything from this first iteration?
- Russel:
- Other than the test bank? I don't think we did. In fact, I sort of put it aside for a few months until [redacted] came up with the biofeedback part.
- Bernstein:
- Talk a little bit about that. How did that work?
- Russel:
- That was really the major breakthrough. It was a clever combination of technologies and methods that had been around since the 50s. The new method of Electrocorticography published that year by [redacted] was absolutely instrumental as well. [redacted] was a true-blue genius.
- Bernstein:
- Did you know him personally?
- Russel:
- I did. We met at a conference in Seoul when I was still a post-doc. He was a colleague, a rival and a good friend.
- Bernstein:
- It's a shame what happened to him.
- Russel:
- ...
- Bernstein:
- So the biofeedback required surgery, if I'm not mistaken?
- Russel:
- That's correct. Minor surgery--although I suppose any kind of surgery involving the brain is major--they would basically drill a small hole in the skull where they could insert the ECG chip. Initially we went for the wireless model, but the latency was causing problems, besides which our subjects started having trouble falling asleep, and we soon switched to the wired variant.
- Bernstein:
- These were the Westmark 14?
- Russel:
- No, no, this was way before the clinical trials. We were experimenting with rats at the time. It was roughly the same model: reinforcement learning with feedback from the ECG. The feedback went both ways. The rats were training the algorithm as much the algorithm was training the rats.
- Bernstein:
- And what were you training them to do exactly?
- Russel:
- Oh you know, the usual. Running mazes. Lever puzzles. That sort of thing.
- Bernstein:
- And it was at that point that you got the additional funding?
- Russel:
- Well, we had published two papers on it. The first one in a medical journal whose name I can't recall just now, might have been Springer Neurophysiology, that was just the biofeedback portion, and the second one in Science. It was after the Science publication that we got a call from the DOD. They wanted to expedite everything and get it to phase I clinical trials. The biggest hurdle was getting volunteers--taking a pill is one thing, getting a hole drilled in your head, quite another. It was impossible to convince any college student. In the end, we went for a special needs high school in LA.
- Bernstein:
- The Westmark 14. Talk a little bit about them.
- Russel:
- Good kids. Not all of them were mentally handicapped by the way, the press never got that part right. A few just had minor psychological or developmental disorders. One boy was just dyslexic, for example.
- Bernstein:
- But the majority had some form of autism spectrum disorder, is that right?
- Russel:
- That's correct. And so the initial goal was try and use the Machine to entrain normative behaviors and mitigate impairments. For the more severe cases we focused on basic speech and fine motor skills.
- Bernstein:
- And you were successful.
- Russel:
- Well...yes, I suppose.
- Bernstein:
- The New York Times reported, and I quote: "Dr. Russel's Westmark 14 are nothing short of a medical miracle. Only six months ago, Susan, aged sixteen, and one of the most severe cases in the study, could not tie her own shoelaces without assistance. Today, she can be found preoccupied with her newfound favorite hobby: crochet." I'd call that a success by any measure, wouldn't you?
- Russel:
- If it had ended there, I'd agree.
- Bernstein:
- Yes, let's talk about Brian Summers.
- Russel:
- This is off the record? (pause) Alright. You have to understand that we never intended it to go that far. I was happy with the results we'd gotten up to that point, perfectly happy, but--well I don't want to point fingers, but I think you know who without my saying it. They wanted us to push and so we pushed. I admit...I was a little curious myself. I knew we'd hit limits with the algorithm eventually, and I wanted to see how the Machine would overcome them. If it could overcome them. I never imagined that it would take THAT approach, however--still less that it would work.
- Bernstein:
- You are talking about the "dream strategy"?
- Russel:
- That term is an invention of the press. We never called it that. But yes. At some point, the Machine figured out how to transmit or implant information directly into the unconscious mind. I remember it distinctly. We thought it was some kind of error with the monitors--just flashing lights and sound--but the kids were glued to them and interacting with the program like normal and their test scores--it was as if they had never been handicapped. Brian Summers was the most exceptional of the group, inasmuch as he seemed to be the most receptive to the Machine's new method, addicted almost. Within three months his vocabulary was comparable to that of my graduate students. Within six, he was fluent in seven languages, including Laotian, which was not even part of the training data. He had reconstructed it from memories of going to a nail salon with his mom when he was six. A month after that he stopped speaking entirely. He hasn't spoken a word since. After his second suicide attempt, the DOD cut off our funding, confiscated the Machine, scrapped our research and threw us to the curb. [redacted] tried to revive the project through some other funding source, but officially the whole project ended a failure. Unofficially...well, I suppose you'd have to ask the DOD.
- Bernstein:
- And Brian Summers? Do you know what happened to him?
- Russel
- (pause) I visit him once a year. Still doesn't speak. His mom bought him a laptop and he's apparently picked up programming--though in a language of his own invention. Last time I saw him, he was working on some kind of natural language program.
- Bernstein:
- What do you think it might be?
- Russel:
- I don't know. You have to understand that, in his case, the Machine went much further than with the others, that the feedback, as I've already mentioned, went both ways. They taught each other. Whatever he's writing is, in a certain sense, the product of a kind of consensus between him the Machine, the product of a new kind of intelligence. If I had to guess, well, what's the first thing a new intelligence would want to do?
Advertisement
Add Comment
Please, Sign In to add comment