Perhaps you've seen the film Limitless, about a struggling writer who is introduced to a nootropic drug called NZT-48 that magnifies his intelligence to super-human levels for 12 hours, after which he reverts back to his old self, but with a headache. Of course, no such drug exists, currently, or else you would notice more people saying intelligent things in the news media and in leadership roles. In fact, I seriously doubt such a compound can exist, unless it does something very complicated, and involves little nanobots restructuring your brain; seems a very long ways away from being a reality to me. However, long before then -- possibly in the next decade or two -- we might discover that invasive brain-computer interfaces, like Neuralink and Paradromics are building, will enable us to produce some of the brain-enhancement results seen in the film. In this post I will try to lay out a possible scenario for how this might come about.
Imagine it's 2022, and some scientists in Silicon Valley, in a nondescript building housing a primate testing facility, begin conducting experiments on enhancing the brains of chimps. What they do is as follows: they install high-throughput, read-write BCIs in the chimps' brains, which are wirelessly tethered to a computer that decides when to stimulate or inhibit certain brain regions (voxels). The computer software that directs this contains a large "brain-encoding model". This model is trained on massive amounts of human brain data, and maps an audio stream to a predicted human brain state -- the state of the brain one would expect to see if a human heard that audio. This predicted human brain state is then mapped to an target / desired chimp brain state; and if a region of the brain of a chimp in the room is at variance with what the model predicts, that region is subtly stimulated or inhibited.
The mapping from human brain state to target chimp brain state will be mostly based on guesswork; but it should work reasonably well for low-level sensory features. Some of these features will probably be similar in both species -- for example, some features of the early visual cortex behave as Gabor filters across a wide variety of animal species:
What the computer would be doing for this chimp is guiding its entire brain towards understanding language more like a human. In the past, it was believed that language understanding was confined to specific regions like the Broca's and Wernicke's areas; but with modern brain-scanning devices scientists have come to realize that language understanding is a "whole brain" affair -- it works almost every region.
At first the scientists maybe speak in single words to the chimp, to see if it understands them, but it gives no indication of such. They try more experiments; talking to it about food, various objects in the room, their names, and so on. Still nothing. After about two weeks of trying, they start to notice the tell-tale signs they are getting somewhere. For example, they talk about "oranges" to themselves, having never used that word around the chimp, and the chimp seems to know they are talking about fruit -- and even which kind of fruit, because it points to an orange toy in the room and then to its mouth. Hairs stand up all over the scientists' bodies; waves of self-adulation and fear wash over them. They turn around, look at each other, and high-5. "Yeah! Man-o-Man! Holy shit!! It actually worked! It actually fucking worked!!"
The next several days, they have long and heated conversations about whether to tell the world, and whether to see how far they can push it. They worry that it isn't ethical to make chimps more like humans, so discontinue that line of work. But do conclude that testing it on humans might be acceptable.
Now we're up to 2025, and a human subject with severe mental impairments has agreed to have one of the high-bandwidth experimental BCIs installed, to help treat them. The treatments are moderately successful, and the patient is pleased with the outcome. The scientists ask him whether he would also be interested in experimental "intelligence enhancement". The sessions would be of short duration, and the brain stimulation and inhibition would be kept at a very low level, so as not to induce seizures, changes in personality, or anything of the sort. Initially, what they want to do is to repeat the chimp experiment, but using a more sophisticated model, that not only uses auditory information, but also visual, tactile, gustatory, olfactory information, and even information from the person's brain itself. One of its many new tricks they want to test is the system's ability to infer, implicitly, what the brain is trying to do -- e.g. trying to remember a word to say in conversation -- and then to complete it, by sharpening the neural pattern. The patient agrees to be a test-subject, and the experiments begin within the next few weeks, and continue over the next several years.
They give the patient a battery of tests to assess his intelligence, and things like vocabulary size and intuition for the meaning of words. For example, they might mention some word to him, and ask him what it means; and then ask him whether he was just guessing (intuition), or whether he actually knew. The reason for this is that, the computer should only have an effect on improving the patient's intuition for the meaning of words, not their actual knowledge (though, the "neural sharpening" feature of the system may induce the subject to recall words he has forgotten the meaning of).
As you might expect, the experiment gets off to a slow start; but, suddenly, and very rapidly, the guy's intuition for the meaning of words improves, very quickly reaching the upper limits of human ability. Although the software was built using human brain data, it was tweaked so as to predict a rich spectrum of shades of meaning, that more often than not matched the brains of experts and geniuses. Not only does the guy's vocabulary drastically improve, but he finds that he is able to parse deep technical conversations and lectures with aplomb, and is able to read very dense text at a blinding pace -- the meaning just flows into his mind like nothing he'd ever experienced before. And it doesn't matter what the text is about -- astronomy, particle physics, Ancient-Greek literature, math... It all makes sense. The enhancement also is designed to work even when he subvocalizes or narratizes words or sentences to himself, in his inner-voice. That is just one more modality for the software to keep track of; as is his inner movie or visual imagination.
Tests of his fluid intelligence and working memory also place him at the extreme upper-limits of human ability. This wasn't a total surprise, as the specific neural net that the scientists built was tweaked to exceed human performance on various intelligence tests, using ideas somewhat like this work:
Solving Raven's Progressive Matrices with Neural Networks
Raven's Progressive Matrices (RPM) have been widely used for Intelligence Quotient (IQ) test of humans. In this paper, we aim to solve RPM with neural networks in both supervised and unsupervised manners.
His creativity and ability to communicate are also enhanced. As a consequence of how the neural net is trained and how it operates, it passes along little hints about what is an appropriate thing to say in any situation.
Now it is 2027, and the scientists' arguments behind the scenes crescendo to a roar, with several saying, "Pull the fucking plug!!" Other say, "We owe it to humanity to let the world know what we have done!! Just think of what we could do if we could make everyone smarter than Einstein!!" Others say, "Autocratic governments already know what we have done... and even our own government. You don't seriously think nobody knows what we are up to, do you?"
In the end, fear wins out. They pull the plug. Meanwhile, several exceedingly rich individuals in the media suddenly seem a lot smarter than they did a few years ago. When told they seem "different", and asked to explain, they say things like that they have turned over a new leaf and have started reading more and talking less.