A system is intelligent if it can use its sensors and effectors to learn continually, generalize, and set/achieve goals in the real world.
Joined January 2018
- Tweets21K
- Following300
- Followers5K
- Likes46.8K
Pinned Tweet
Neuroscience bits from my research
My biggest breakthrough: Recognition is a by-product of perception.
Over the many years that I've researched the human brain, I've had a number of minor breakthroughs in my understanding. Most of them had to do with the architecture and function of the retina. I've had only one breakthrough that I consider to be the biggest one of them all.
In the beginning, like everyone else who researches intelligence, I had assumed that the brain learned to recognize patterns. I spent many years looking for a plausible cortical mechanism that would support the assumption. I failed miserably.
Then one day, a question popped into my consciousness. What if I was wrong? What if everyone was wrong? What if it is not true that the brain learns to recognize patterns? I knew from my studies that most of the innate synaptic connections of the neocortex completely disappeared early in infancy. I reasoned that this could not possibly have anything to do with pattern learning.
At that point, the rebel in me took over. The brain is learning something but it is not pattern recognition. What then? Then I remembered thinking about something peculiar that I had observed before.
Unlike deep learning systems, the brain can perceive anything in the world, whether it recognizes it or not. By contrast, a deep neural net cannot perceive anything unless it has been previously trained to recognize it. This is a fatal flaw imo, one of the many reasons that the AI community cannot solve AGI. (note: I've written about this many times, but it's so important, it bears repeating over and over.)
I had a flash of insight. The brain doesn't learn to recognize the world. It learns to perceive it. Recognition is a mere by-product of perception. Indeed, we forget almost everything we perceive. When we perceive an object, the brain activates short-lived memory "traces". This is what psychologists call short term memory (STM).
In order to remember or later recognize a pattern, the STM traces must be reactivated a number of times. Otherwise they fade away and self-delete. With enough reactivations, the traces can become permanent. That is, they become part of long term memory (LTM).
Work in progress. I love the brain. 😍
Neuroscience bits from my research: Recognition is a by-product of perception.
Nothing can move in more than one direction at the same time.
Don't let its simplicity fool you. I believe this is the most powerful principle of perception in the visual cortex. It is used for both learning and on-the-fly perception.
While researching the visual system of the human brain, I came to understand that, unlike deep neural nets, the visual cortex doesn't learn to recognize patterns, scenes or objects. Rather, it learns to perceive the world instantly, i.e., to use the world as its own model.
Recognition is a by-product of perception. When a perceptual representation is built in memory, it has a short life and, unless rehearsed (recalled) repeatedly, it slowly decays and disappears.
The advantage of universal and instantaneous perception is that the brain is not blind to patterns or scenes it has not encountered before. Humans learn to achieve universal perception during the first few years after birth. During this time, most of the initial connections in the cortex disappear.
Instant perception solves the edge case problem, the bane of the self-driving car industry. This is an absolutely essential ability that deep learning systems do not have. It is one of the many reasons that DL will not be part of the AGI solution.
I believe there are maybe 2 or 3 additional principles that govern visual perception. I have a couple of partly baked ideas percolating in my brain. Hopefully, something will come out of it soon.
My thesis is that all the sensory modalities are analogous. The main difference is in the architecture and function of the various sensors. This is exciting because it means that solving one cortical functionality amounts to solving all the others.
Concept formation, including language learning, is a separate process that uses recognized percepts from all the regions of the cortex to form associations. It's an essential part of intelligence but I think it's a trivial problem compared to sensory perception.
I could be wrong but I still believe that the mystery of intelligence will likely be solved before the end of this decade. It will be a neuroscience solution. Work in progress.
I love the brain. 🧠😍
China is making a big mistake. US AI (LLMs) is not AI. It's fake AI. Don't copy them. Don't waste the money on fake AI. Nature knows best. Learn neuroscience and philosophy instead.
I know I'm speaking to the wind. Oh well. 🤦♂️
BREAKING: DeepSeek just raised $15 BILLION to aggressively build data centers and open-source models
>walks into investor meetings in hangzhou
>"we need $7 billion"
>investors start sweating
>"actually make it $11 billion"
>Tencent nods immediately
>"you know what, $15 billion"
>CATL, a literal battery company, throws in a bag
>everyone just kept nodding
One of the largest funding rounds in Chinese history btw.
"Children learn by playing; Optimus needs to learn doing self-play at scale." ~ Elon Musk
If I was a superstitious man, I'd swear Elon Musk was reading my posts. 😀
I mean, just yesterday I posted that our future intelligent robots will need to learn everything in the real world and play like babies. It's too bad Elon has no idea how to that. His engineers will use simulations and reinforcement learning like everyone else. Good luck with that. Haha. 😂
Tesla Q3 earnings call - answering a question on data collection:
Elon:
We’re working on AI architecture. For data, we need on the order of 10k Optimus bots, we’re building Optimus Academy where the robots could learn. Unlike cars, the data flywheels need to be bootstrapped.
Children learn by playing; Optimus needs to learn doing self-play at scale.
Ashok:
Unlike cars, we don’t have real-world data for Humanoids. Optimus will have to learn from watching humans to bootstrap learning, and then do self-play in sim and the real world to get better.
Elon:
In a month or so, Tesla will release videos that are not canned. Tesla AI5 will be integrated into Optimus sometime around April next year. It’s a challenge to squeeze all the intelligence for Optimus into the current computer, so the AI5 computer will make a big difference.
In about a year, we will start shipping to customers, starting with industrial and dangerous tasks.
Elon Musk and robots
"There's still quite a lot of work to do with the AI in the robot." ~ @elonmusk on Optimus, the humanoid robot
You think? 🙄
"I would expect that the AI, the electromechanical design, reliability, usage life, expect these issues to be solved by the end of next year."
The AI is the showstopper. Unlike the other stuff you mentioned, the AI problem is not an engineering problem. It's a scientific problem that requires a scientific breakthrough in our understanding of intelligence. Current AI methods only produce robots with canned automation. Intelligence is not automation.
"I remain convinced Optimus will be the biggest product on Earth."
Earth to Elon:
Man, Tesla is not the only company on Earth making robots. If someone cracks the secrets of intelligence (I doubt it will be Tesla), there's no guarantee they will share it with you. And, if they do share the science with the world, everyone else will have it.
After intelligence is solved, the proto-robots will have to undergo an upbringing period in the real world. There's no other way for them to gain an understanding of the world. The hard work will be in raising them and teaching them to become proficient in various real-world skills, trades or professions. That will take years. Eventually, they will be cloned by the millions or billions as needed. That's when the true robot/AI revolution will begin. 🤔
In other words, there's plenty of competition out there. You are not an island unto yourself. And again, you don't know how to solve intelligence. We're still waiting for the millions of robotaxis. 😀
Elon Musk on Optimus:
"We've started early production. The growth of Optimus is an S curve, but the initial curve is a shallow slope. We have the vast majority of the 1 million unit/year line installed at Fremont. There's still some things we need to tweak about the design. There's still quite a lot of work to do with the AI in the robot. I would expect that the AI, the electromechanical design, reliability, usage life, expect these issues to be solved by the end of next year. I remain convinced Optimus will be the biggest product on Earth. I do want to temper expectations that we still have a lot to get through, it will take about a year to get through that."
AGIHound retweeted
I think AI companies are too focused on one particular technique, the transformer architecture + backpropagation & betting all their efforts on the likelihood that at large enough scale it could undergo some "phase change" enabling it to reach & surpass human intelligence.
You mean the markets that priced SpaceX they way they did? Don't know, it's also not about politics and I really despise Dario and the cultish way Anthropic behaves. But if I look solely at the numbers and combine that with what I just saw speaking to ~50 different business owners over the past 4 weeks, I can only see that the current usage is at best early adopter stage for the whole market. Means, if their cash does not run out for the next 15 month and it doesn't look like it will, adoption will have moved much further through the early adopter phase and into the early majority where the real money sits.
I would rather look at this from the perspective of them becoming supply constraint over the course of next year. But for bankruptcy, given these numbers, there would be a need for stuff to become really bad.
I am personally, after all that I experienced in my own work with my customers and seeing these numbers, more and more certain, that there is no bubble pop waiting for us.
Neuroscience: two attention mechanisms?
I've come to understand that the human brain has two attention mechanisms in each hemisphere. They allow us to form associations between perceptual phenomena located in different or distant parts of the neo-cortex.
For example, we use them associate spoken words (auditory cortex) and visual objects (visual cortex).
Attentional activities in the two hemispheres are linked together via the corpus callosum. 🤔
Neuroscience bits from my research: Recognition is a by-product of perception.
Nothing can move in more than one direction at the same time.
Don't let its simplicity fool you. I believe this is the most powerful principle of perception in the visual cortex. It is used for both learning and on-the-fly perception.
While researching the visual system of the human brain, I came to understand that, unlike deep neural nets, the visual cortex doesn't learn to recognize patterns, scenes or objects. Rather, it learns to perceive the world instantly, i.e., to use the world as its own model.
Recognition is a by-product of perception. When a perceptual representation is built in memory, it has a short life and, unless rehearsed (recalled) repeatedly, it slowly decays and disappears.
The advantage of universal and instantaneous perception is that the brain is not blind to patterns or scenes it has not encountered before. Humans learn to achieve universal perception during the first few years after birth. During this time, most of the initial connections in the cortex disappear.
Instant perception solves the edge case problem, the bane of the self-driving car industry. This is an absolutely essential ability that deep learning systems do not have. It is one of the many reasons that DL will not be part of the AGI solution.
I believe there are maybe 2 or 3 additional principles that govern visual perception. I have a couple of partly baked ideas percolating in my brain. Hopefully, something will come out of it soon.
My thesis is that all the sensory modalities are analogous. The main difference is in the architecture and function of the various sensors. This is exciting because it means that solving one cortical functionality amounts to solving all the others.
Concept formation, including language learning, is a separate process that uses recognized percepts from all the regions of the cortex to form associations. It's an essential part of intelligence but I think it's a trivial problem compared to sensory perception.
I could be wrong but I still believe that the mystery of intelligence will likely be solved before the end of this decade. It will be a neuroscience solution. Work in progress.
I love the brain. 🧠😍
This butterfly has more real-world intelligence than Anthropic's Claude. And it's more beautiful than Dario Amodei and everyone else working at Anthropic. 😀
If you can emulate the intelligence of this wonderful little creature, you cracked the secrets of intelligence. Congratulations.
These automated humanoid robots are impressive but they have no intelligence. Intelligence is not automation.
Whoever cracks the secrets of intelligence will rule the world. The good news is that the fake-AI mafia don't know the first thing about intelligence. Our wicked overlords won't know what hit them. Just saying. 😀
I love AI bubbles and AI winters. Just saying. 😀
The "unlimited AI demand" story just hit a wall.
Corporate token spending growth collapsed from 42% in July to 0.2% in September, per Ramp data. Meanwhile, return on invested capital at the five biggest hyperscalers fell from 38% to 20% as capex outruns earnings. The infinite-demand thesis needs infinite returns. The returns are going the other way.
AGIHound retweeted
"We burned ungodly amounts of compute and got nothing in return."
So par for the course? Lol.
Rules to keep in mind about any model of anything:
1. The model is only as good as the data used to build it. The real world is the gold standard.
2. All models are wrong. Some are close enough.
3. Models that appear close, made with data that looks roughly sufficient, will still almost always fart out a stream of false positives. This is because the models are designed to output lots of things, and because of rule 2, are wrong.
4. Output without understanding is just entropy.
Where do you go next after reporting from the Strait of Hormuz? For the past three months, Citrini Research, an independent research firm, has been working with our team and an internal research model to see if they could use AI to create a new drug. What follows is an entertaining, candid, and humbling account of a group of motivated nonexperts (or “filthy casuals”, as Citrini would say) trying to tackle a real problem in human health with Claude.
We landed on scorpion neurotoxins, which kill over 3,000 people a year. Of the 211 designs we produced and tested, 17 met our criteria for a binder. None of the designs bound more than one toxin and none have been tested for neutralizing a toxin.
One of my favorite quotes from the piece is: “Our virtual biologist may have compressed a year's worth of work into a couple days, but biology works at the speed of life…Over two days, we watched Claude work through problems at a speed that was difficult to comprehend. For the next two months, we spent much of the time becoming familiar with laboratory turnaround times (This mostly involved sending memes in Slack).”
I love the sound of bubbles popping when I wake up in the morning.
I'm loading up on popcorn and beer, waiting for the AI bubble to pop. 🍿🍺
PS. Someone should create an app that sends an alert when the AI bubble pops. I don't want to miss it. 😀
More physics for rebels in the age of fake AI
What infuriates me about the fake-AI mafia is that LLMs (fake AI) are fed with all of these bullshit-science theories with no questions asked. Garbage in, garbage out. 🤦♂️
How is fake AI supposed to devise new scientific theories if it has no ability to detect that the stuff it's been trained on is absolute garbage?
I rebel and I protest. 😠
Physics for rebels: the problem with the time dimension
I, too, have a bullshit meter.
Modern physics is 90% garbage science in my opinion. Did you know that a time dimension makes motion impossible? This is the reason that spacetime is called a "block universe in which nothing happens" (Karl Popper).
This is old news but, apparently, "physicist" @skdh either doesn't know this, or she pretends not to know it, since it falsifies Einstein's spacetime physics. Political correctness in science forbids it. 😠
Why does a time dimension make motion impossible, you ask? It's because motion in time is self-referential, an absolute no-no in logic. Why? It's because motion in time would have to be written mathematically as v = dt/dt = 1. Which is nonsense.
Yes, it is that simple to refute Einstein's spacetime theories. All of the time travel stuff you see in sci-fi movies is 100% garbage. And that's just the tip of the iceberg of what is wrong with Einstein's physics. He himself came to accept that his entire like's work was just a "castle in the air" after he realized that continuity, the basis of his theories, is a bogus concept.
So, not content with one time dimension crackpottery, a "physicist" creates a second time dimension. And Hossenfelder is OK with it. Dear Lord. She's not a physicist. She's a political propagandist for fake physics on YouTube.
On a scale of 1 to 10, I give her a 10 on my bullshit meter. 😀
Physics for rebels: the problem with the time dimension
I, too, have a bullshit meter.
Modern physics is 90% garbage science in my opinion. Did you know that a time dimension makes motion impossible? This is the reason that spacetime is called a "block universe in which nothing happens" (Karl Popper).
This is old news but, apparently, "physicist" @skdh either doesn't know this, or she pretends not to know it, since it falsifies Einstein's spacetime physics. Political correctness in science forbids it. 😠
Why does a time dimension make motion impossible, you ask? It's because motion in time is self-referential, an absolute no-no in logic. Why? It's because motion in time would have to be written mathematically as v = dt/dt = 1. Which is nonsense.
Yes, it is that simple to refute Einstein's spacetime theories. All of the time travel stuff you see in sci-fi movies is 100% garbage. And that's just the tip of the iceberg of what is wrong with Einstein's physics. He himself came to accept that his entire like's work was just a "castle in the air" after he realized that continuity, the basis of his theories, is a bogus concept.
So, not content with one time dimension crackpottery, a "physicist" creates a second time dimension. And Hossenfelder is OK with it. Dear Lord. She's not a physicist. She's a political propagandist for fake physics on YouTube.
On a scale of 1 to 10, I give her a 10 on my bullshit meter. 😀
AGIHound retweeted
Statistical modeling requires tons of data to learn variable co-occurrences that don't capture causality.
While causal modeling requires reasoning about directional relations, from cause to effect & is more data efficient.
Humans use the latter for meta-learning.
If AI training is "stealing," then anything that's ever been studied or created has been off the backs of stolen work. It's a stupid standard that falls apart immediately.
Models learn no differently than you did. They take in an enormous amount of other people's work, compress it into weights, and produce new output shaped by everything they absorbed.
You took in hundreds of books, songs and films, which were made by people who took in hundreds of books, songs and films, which were made by people who took in hundreds of books, songs and films, and on and on.
Your neurons compressed all of it into connection strengths, and what comes out of your head is shaped by everything you absorbed. By this standard every writer in history is a thief of a thief of a thief, and the only person who owes nothing to anyone is the one who has never read or heard anything, which is nobody.
Learning from a work takes nothing from its author. The people screaming "theft" at a model want a permission slip before a mind is allowed to learn from what's public. That is entitlement. It is destructive. It is an ego that mistakes its own remix for creation from nothing and demands the whole chain of influence stop at its door.
AGIHound retweeted
"Models learn no differently than you did. They take in an enormous amount of other people's work, compress it into weights, and produce new output shaped by everything they absorbed."
Not at all how human beings learn. Humans are not statistical machines and AI does not learn. It is merely statistical computation.
The machines are limited to permutations of training data and therefore are unable to create any new semantic information. If they could, then new training data wouldn't be required.
It is precisely why the AI labs are desperate for data, because their machines cannot create any new data with meaning. AI "intelligence" is still just an illusion.
If AI training is "stealing," then anything that's ever been studied or created has been off the backs of stolen work. It's a stupid standard that falls apart immediately.
Models learn no differently than you did. They take in an enormous amount of other people's work, compress it into weights, and produce new output shaped by everything they absorbed.
You took in hundreds of books, songs and films, which were made by people who took in hundreds of books, songs and films, which were made by people who took in hundreds of books, songs and films, and on and on.
Your neurons compressed all of it into connection strengths, and what comes out of your head is shaped by everything you absorbed. By this standard every writer in history is a thief of a thief of a thief, and the only person who owes nothing to anyone is the one who has never read or heard anything, which is nobody.
Learning from a work takes nothing from its author. The people screaming "theft" at a model want a permission slip before a mind is allowed to learn from what's public. That is entitlement. It is destructive. It is an ego that mistakes its own remix for creation from nothing and demands the whole chain of influence stop at its door.
Wow. I'm surprised. Was Grok trained on some of my posts?
Yes, Grok, it's all massive statistics, aka cheating. Without real-world sensory grounding, there is no understanding and no knowledge. 😀
LLMs predict tokens via statistics on massive text, capturing rich contextual patterns that encode meaning, facts and relations. This yields strong functional "knowledge" outperforming most humans on many tasks. Yet they lack sensory grounding, continuous real-world learning or true referential understanding—so the "knowing" is emergent simulation, not human-like comprehension. Both views hold partial truth.