1. Je n'ai jamais dit que les LLMs étaient inutiles ou n'avaient pas de valeur commerciale.
— Yann LeCun (@ylecun) June 18, 2026
2. Mais beaucoup de technologies sont utiles et à haute valeur sans être particulièrement "intelligentes"
3. Les LLMs sont essentiellement inutiles et sans grande valeur pour ce qui est…
English translation is available at the link.
Judah Goldfeder, Philippe Wyder, Yann LeCun, Ravid Shwartz-Ziv, AI Must Embrace Specialization via Superhuman Adaptable Intelligence, arXiv:2602.23643v1 [cs.AI], 2026.
Abstract: Everyone from AI executives and researchers to doomsayers, politicians, and activists is talking about Artificial General Intelligence (AGI). Yet, they often don't seem to agree on its exact definition. One common definition of AGI is an AI that can do everything a human can do, but are humans truly general? In this paper, we address what's wrong with our conception of AGI, and why, even in its most coherent formulation, it is a flawed concept to describe the future of AI. We explore whether the most widely accepted definitions are plausible, useful, and truly general. We argue that AI must embrace specialization, rather than strive for generality, and in its specialization strive for superhuman performance, and introduce Superhuman Adaptable Intelligence (SAI). SAI is defined as intelligence that can learn to exceed humans at anything important that we can do, and that can fill in the skill gaps where humans are incapable. We then lay out how SAI can help hone a discussion around AI that was blurred by an overloaded definition of AGI, and extrapolate the implications of using it as a guide for the future.
In view of the articles in this special double-issue of Dædalus, AI & Science: What Is the Future of Discovery?, I must agree. Yes, the ascent of Mount AGI will continue, but at the same time we will be developing more specialized AIs for specific tasks, AlphaFold is one example, but it is only one of many. Back in 1990 David Hays and I published an article in which we asserted, "Sooner or later we will create a technology capable of doing what, heretofore, only we could." We didn't put any dates on that, nor did we envision today's technology, but we could see the long-term trend. And that trend certainly includes specialized AIs. Think of them as intelligent instruments.
[Hmmm... Why don't we think of trains, planes, and cars as superhuman vehicular transportation (SVT)?]
Dædalus currently has a double issue, AI & Science: What Is the Future of Discovery?, edited by James M. Manyika. Manyika interviews both Hassabis and LeCun and they offer remarks relevant to the issue of computational compressibility as I discussed it in my recent working paper, On Method: Computational Compressibility in Complex Natural and Cultural Phenomena, though they don’t use the term. Here are some passages from those interviews.
Demis Hassabis
In this first passage Hassabis is talking about a well-known problem in computer science, known as P versus NP, which is about how long it takes to solve a problem as a function of the size of the input. Roughly speaking, what’s at stake goes like this: If you are presented with a proposed solution and can verify it quickly, could you also find the solution from scratch relatively quickly (in polynomial time, P) or is finding a solution so difficult as to be all but impossible (NP, Nondeterministic Polynomial time). You don’t need to understand that to understand this passage , pp. 36-38:
It does, and I think those are the interesting limits to test and understand. P equals NP–which attempts to categorize the difficulty of a problem by how much computation it would take to find and check a solution, respectively–is one of the most important questions in science to resolve. I suspect P is not equal to NP, and there are some problems out there that are just not tractable to solve in a practical amount of time without invoking the help of, say, a quantum computer, but we need to understand this a lot better because there may be more nuance here than we previously realized. In our work with AlphaGo and AlphaFold, we’re showing that if you do a lot of precompute, which is not normally considered in these kinds of scenarios, you can seemingly answer some highly complex questions approximately optimally in P (polynomial) time. Neural networks are effectively using massive amounts of precompute to compress knowledge into some efficient artifact. That computed artifact is then available at test time and, for a lot of natural systems, you can use it to narrow down your search space so you don’t have to consider all the possible configurations they could potentially take, but only a much smaller subset that are actually plausible.
Those last two sentences are about computational compressibility. Hassibis then goes on to illustrate:
Let’s take proteins. There are roughly 10300 possible conformations of an average protein. It would take longer than the age of the universe to enumerate that exhaustively to find the one specific shape it takes, so you have to do something much smarter. You have to learn what patterns there are for different amino acid sequences and then only search a tiny fraction of the possibilities to find the approximately correct solution. That seems to be what we managed to do with AlphaFold. Maybe not perfectly, but to an approximation that is at least good enough for practical purposes. [...]
AlphaFold was our solution to the protein folding or protein structure prediction problem. You start with an amino acid sequence–you can think of it very roughly as the genetic sequence for the protein, a one-dimensional string of letters. In the body or in nature, that string folds up into a 3D structure, and that shape goes a long way toward defining the function of that protein, which is really important for drug discovery and disease understanding.[...]
The way we did it is that there were about 150,000 known structures that had been painstakingly put together by structural biologists over the past thirty to forty years with very expensive equipment like electron microscopes. That was just about enough data to give our AI system clues as to the topology of proteins. Of course they don’t just fold up randomly; there are some constraints, and the AI system learned them. Eventually it was able, within a few seconds, to come up with a plausible structure for an unseen protein.
In this next passage, the first two conditions are about compressibility, p. 39:
We look for three aspects of a problem in determining whether it is suitable to tackle with the AI techniques we have today. First, can the problem be described as or converted into a description of a massive combinatorial space? Perhaps it’s intractably large and normal brute force techniques won’t work. Second, if that’s true, do you have enough data to learn some sort of model of the topology of that space? Or maybe a simulator is available or learnable that can generate some additional synthetic data. Ideally, you have both. Third, you need a clear objective that you’re trying to minimize or maximize. In games, that is winning or maximizing the score. In a natural system, that might be minimizing the free energy in that system. If you can quantify that, you can then use a model to search with the guidance of the objective function toward the optimal solution.
Yann LeCun
In the following passage LeCun talks about an abstract representation space. That space contains a compressed representation of the phenomenon, pp. 47-48
I think this is a crucial point and is what I am presently devoting all of my efforts to: devising AI systems that can find an abstract representation of the phenomenon and make predictions in that abstract representation space. This abstract representation eliminates a lot of details about the original observations. And that’s a crucial point because LLMs (large language models) and other generative models are trained to predict every detail of the input. In language, it’s not too much of a problem. You cannot predict exactly which word follows a sequence of words, but you can produce a probability distribution over words. That’s easy because there’s a finite number of possible words. But when you train the model to predict future frames in a video, you can’t represent a useful distribution. You have to make predictions in an abstract representation space, not at the pixel level. So a lot of people in the last few years instinctively said, “let’s just tokenize the world.” Let’s take images from videos and cut them into little squares and turn that into a vector that doesn’t look different from the one that represents a word, and feed this to a gigantic model to predict the next few frames. Frankly, it doesn’t work that well. The reason why is that you simply cannot predict what’s going to happen in a video at the pixel level. There are so many details that are just not in the input. We don’t know how to produce a probability distribution over all possible video frames because it’s mathematically intractable. It’s a problem people have struggled with for decades in statistical physics.
Instead, what we do as scientists is to find a representation of the input that eliminates all the details we cannot predict, and we make predictions in that representation space. That’s not a generative architecture.
Later, p. 55:
Manyika. Given the advances in AI, and particularly if we go beyond human cognitive levels and AI systems come to understand more than we do, what are the implications for philosophy of science, how we do science, and the nature of scientific understanding?
LeCun. I think that question is not a new one. When we solved PDEs (partial differential equations) numerically with computers, did the computational fluid dynamics simulator understand physics better than we did? It can make a prediction and it’s using an algorithm based on equations that humans came up with.
The next step AI enables is training a machine-learning system to make predictions from data without the manual step of reducing the process to equations. AI allows us to skip having to first build a model of reality that can then be computed. This is powerful because many phenomena in science are collective complex phenomena.
That is the compression step. LeCun continues:
A pile of sand behaves in a particular way, and the theory for this is not entirely clear. The property of materials, particularly complex ones, cannot be directly derived from the elementary equations of quantum mechanics. It’s just too complicated. Another example is the magic angle, 1.1 degrees, at which you rotate two stacked monolayers of carbon, called graphene, to form a superconductor. That’s a collective phenomenon that is extremely difficult to explain. There are various properties of materials of this type that cannot be usefully reduced to a small number of equations from which you can derive this collective behavior. How does intelligence emerge from neurons in interaction? That’s a philosophical question of how a super complex property like intelligence can emerge from a large number of relatively simple elements in interaction, but that’s a pretty high-level thing. At a lower level are questions of how life emerges from the interaction between proteins. This transition is what has baffled scientists for a long time: the transition from the microscopic to the mesoscopic. This is where interesting things happen, like life, for example.
So now there’s a new way of doing science, which is neither completely qualitative and observational nor reductionist, but is a data-driven, AI-powered phenomenological model that may allow us to bridge the gap between microscopic and macroscopic.
That last paragraph is about compression.
1. I never said LLMs were not useful. They are, particularly with all the bells and whistles that are being added to them. I use them.
— Yann LeCun (@ylecun) April 24, 2026
2. A robot-rich future can't be built with AIs that don't understand the physical world and don't anticipate the consequences of their actions.…
LOL!
Of course he's right about LLMs, especially about all those bells and whistles. Those add symbolic control to the core neural net capabilities of LLMs. And of course he's right about the need to predict what's happening in the physical world. Robots need that and LLMs cannot provide it. Next-token prediction is a different beast, even if those tokens are about things happening in the physical world.
Yann LeCun (AMI Labs Founder): "The AI industry is completely LLM-pilled. Everybody is working on the same thing. They're all digging the same trench."
— Big Brain AI (@realBigBrainAI) April 22, 2026
LeCun explains why no lab dares break from the pack:
"They are stealing each other's engineers. So they can't afford to do… pic.twitter.com/1XE7RaCsCf
🚨 Holy shit... LeCun's team just cracked world models wide open.
— Alex Prompter (@alex_prompter) March 24, 2026
Everyone's obsessing over the next Claude update.
Meanwhile Yann LeCun quietly dropped a paper that could matter way more long term.
It's called LeWorldModel.
And to understand why it's a big deal, you need… pic.twitter.com/jcwo02hTYB
This post was composed by Claude (Anthropic) after an interaction which I initiated with a prompt consisting of 1) a capsule summary about the history of OpenAI that included a number of questions, and 2) a request for the 10 most expensive scientific research projects paid-for by the US Government. That interaction went on for a bit over 7100 words, after which I asked Claude to write a blog post. The following article is more creative than a mere summary of that discussion.
On December 11, 2015, a small group of technologists gathered in San Francisco to launch what they described as a nonprofit research organization dedicated to ensuring that artificial general intelligence would benefit all of humanity. The founders of OpenAI — Sam Altman, Greg Brockman, Ilya Sutskever, Wojciech Zaremba, Elon Musk, and others — began with a $1 billion endowment and a serious concern: that the most transformative technology in human history was being developed inside a handful of profit-maximizing corporations, with no institutional safeguard ensuring it would serve everyone. The nonprofit structure was the answer. No investors to satisfy, no quarterly earnings to hit. Just the mission.
The mission lasted four years in its pure form. By March 2019, faced with the staggering computational costs of training large language models, OpenAI created a for-profit subsidiary with a novel "capped profit" structure: investors could earn returns, but those returns were limited to one hundred times their investment, with excess profits flowing back to the nonprofit parent. This was the arrangement that attracted Microsoft's initial investment, and it was the arrangement in place when OpenAI released ChatGPT to the general public in late November 2022.
What happened next was, by any measure, one of the most consequential commercial surprises in the history of technology. Within two months, ChatGPT had a hundred million users. The scale and speed of public adoption had no precedent. And the shock of that success — the sheer unexpectedness of it — set in motion a chain of decisions that has reshaped not just one company, but the entire research landscape of artificial intelligence.
In January 2023, Microsoft announced a new $10 billion investment in OpenAI. The nonprofit's original rationale — that the most powerful AI should not be controlled by a for-profit corporation — was under increasing strain. By October 2025, it had formally dissolved. OpenAI restructured as a public benefit corporation, the nonprofit parent renamed itself the OpenAI Foundation and accepted a 26% equity stake in the new entity, and Microsoft received a 27% stake worth approximately $135 billion. The PBC structure requires the company to consider its mission alongside profit — but as a legal constraint, it is considerably weaker than the nonprofit board that had previously governed the organization.
The journey from nonprofit to PBC was not smooth. In November 2023, OpenAI's board — still operating under its nonprofit governance mandate — fired Sam Altman as CEO, citing concerns about his candor and, beneath the official language, a deeper unease about the pace of commercialization. The firing lasted five days. Nearly all 800 of OpenAI's employees threatened to resign and follow Altman to Microsoft. Ilya Sutskever, who had orchestrated the firing, signed the letter calling for Altman's reinstatement and issued a public apology. Altman returned, the board was reconstituted with his allies, and the mission-protection mechanism that the nonprofit structure had been designed to provide was effectively neutralized. Sutskever left the company in May 2024.
Each structural change was framed as necessary to fulfill the mission. In practice, each change progressively subordinated the mission to capital requirements. The nonprofit board had existed to ensure that AGI benefited humanity. By 2025, it had become a foundation holding equity in the thing it was supposed to be watching — a watchdog with a financial stake in the object of its oversight.
To understand what was lost in this transformation, it helps to draw a distinction that rarely gets made clearly in public discussions of AI: the difference between curiosity-driven, open-ended research and product-driven, outcome-oriented development.
Consider the Apollo program as an example of the second kind. It was, in the deepest sense, an engineering project rather than a scientific one. The underlying physics was known. Orbital mechanics, propulsion, life support — these were hard and dangerous problems, but they were problems whose solutions could be systematically approached. The goal was precisely defined. The timeline could be committed to. Success was probable given sufficient resources. When President Kennedy pledged to put a man on the moon by the end of the decade, he was making a political commitment backed by a technical assessment that success was achievable. The scientists who worked on Apollo — and I have met a number of them — may have been motivated by curiosity and wonder. But Congress funded the program to beat the Soviets in the Cold War. The institutional structure — massive, goal-directed, centrally coordinated — suited the nature of the problem.
Curiosity-driven research operates on entirely different premises. Its defining characteristic is that it does not know in advance what it will find. Claude Shannon was not trying to build the internet when he developed information theory at Bell Labs in the late 1940s. The researchers at the University of Montreal who developed attention mechanisms for neural networks were not trying to build ChatGPT. The work that seeded the current AI revolution — Rosenblatt's perceptron, Minsky's early investigations, the decades of foundational work in cognitive science and linguistics that LLMs now implicitly exploit — was almost entirely publicly funded, pursued at universities and a handful of exceptional industrial research labs, over decades when no commercial application was visible.
Bell Labs was the great institutional embodiment of this model in the corporate world. What made it possible was structural: AT&T's government-protected monopoly generated profits so vast that the company could fund a research laboratory with no requirement to produce commercial results. Shannon, Bardeen, Brattain, Shockley — these men were given time, resources, and colleagues, and told to think. The transistor, information theory, Unix, the laser, cellular telephony, and multiple Nobel Prizes resulted. Bell Labs was not run like a startup. It was run like a slightly more applied version of a university, with better equipment.
Xerox PARC, founded in 1970, operated on similar principles — explicitly unconstrained by Xerox's core product lines, given a unifying vision ("the architecture of information") but not a product roadmap. The personal computer, the graphical user interface, Ethernet, the mouse, laser printing — all emerged from a lab of about 350 people who were essentially allowed to play. The irony is that Xerox captured almost none of the commercial value, which accrued to Apple, Microsoft, and others. But the world got the technology.
Asked directly about modern equivalents to Bell Labs and PARC, Yann LeCun — who worked at Bell Labs, interned at Xerox PARC, and spent over a decade building Meta's fundamental AI research lab — pointed to Meta's FAIR, Google DeepMind, and Microsoft Research. He said this in October 2024. By November 2025, he had left Meta, driven out by exactly the forces this article is about.
Before November 2022, the AI research world was genuinely plural. Academic labs, industrial research divisions, and a range of well-funded startups were pursuing different approaches — reinforcement learning, symbolic AI hybrids, world models, neuromorphic architectures — with real diversity of vision. The field was competitive but intellectually heterogeneous.
ChatGPT's success collapsed that plurality. Within roughly eighteen months, capital, talent, and institutional attention all funneled toward a single paradigm: scale transformer-based large language models, build the infrastructure to run them, ship products. Google, which had invented the transformer architecture in 2017, was caught flat-footed and scrambled. Meta pivoted its AI strategy around LLMs. Microsoft integrated OpenAI's models into its core products. A hundred startups raised money to build on top of the new foundation models. The venture capital flowing into AI, measured as a share of total U.S. deal value, went from 23% in 2023 to nearly two-thirds in the first half of 2025.
The infrastructure investment that followed is staggering by any historical standard. The four largest hyperscalers — Amazon, Google, Microsoft, and Meta — are expected to spend more than $350 billion on capital expenditures in 2025 alone, most of it AI-related. UBS projects global AI capital expenditure reaching $1.3 trillion by 2030. The top five hyperscalers raised a record $108 billion in debt in 2025, more than three times the average of the previous nine years. OpenAI, which loses billions of dollars annually, has committed to spending $300 billion on computing infrastructure over five years while projecting only $13 billion in revenue for 2025.
The financial architecture has become genuinely strange. OpenAI holds a stake in AMD; Nvidia has invested $100 billion in OpenAI; Microsoft is a major shareholder in OpenAI and a major customer of CoreWeave, in which Nvidia also holds equity; Microsoft accounted for nearly 20% of Nvidia's revenue. These are not arm's-length market transactions. They are a daisy chain of mutually reinforcing valuations. A Yale analysis described OpenAI's web of relationships bluntly: "Is this like the Wild West, where anything goes to get the deal done?" The question of whether this constitutes a speculative bubble — tulip mania in a data center — is not academic. An MIT Media Lab report found that 95% of custom enterprise AI tools fail to produce measurable financial returns. The commercial success is real; the path from current AI to the transformative economic productivity being used to justify the valuations is not established.
The most consequential intellectual development of the past two years in AI has received far less attention than the commercial race. A growing number of the field's most distinguished researchers have concluded that large language models, however impressive, are not on the path to general intelligence — and that the current paradigm will hit a ceiling before it reaches the goals its proponents have claimed for it.
💯 Information monopolies are always a concern, but LLMs ratchet it up way beyond search engines. You are not just being referred to a trustworthy source, but are being given information predigested from opaque sources.
— Subbarao Kambhampati (కంభంపాటి సుబ్బారావు) (@rao2z) February 1, 2026
Extinction angst serves as a convenient distraction.. https://t.co/kAHZXkzJiR
Cade Metz, An A.I. Pioneer Warns the Tech ‘Herd’ Is Marching Into a Dead End, NYTimes, Jan. 26, 2026.
But after leaving Meta in November, Dr. LeCun has become increasingly vocal in his criticism of Silicon Valley’s single-minded approach to building intelligent machines. He argues that the technology industry will eventually hit a dead end in its A.I. development — after years of work and hundreds of billions of dollars spent.
The reason, he said, goes back to what he has argued for years: Large language models, or L.L.M.s, the A.I. technology at the heart of popular products like ChatGPT, can get only so powerful. And companies are throwing everything they have at projects that won’t get them to their goal to make computers as smart as or even smarter than humans. More creative Chinese companies, he added, could get there first.
“There is this herd effect where everyone in Silicon Valley has to work on the same thing,” he said during a recent interview from his home in Paris. “It does not leave much room for other approaches that may be much more promising in the long term.”
That critique is the latest shot in a debate that has roiled the tech industry since OpenAI sparked the A.I. boom in 2022 with the release of ChatGPT: Is it possible to create so-called artificial general intelligence or even more powerful superintelligence? And can companies get there using their current technology and concepts?
Lecun is not worried about so-called "AI Doom":
Not long after ChatGPT was released, the two researchers who received the 2018 Turing Award with Dr. LeCun warned that A.I. was growing too powerful. Those scientists even warned that the technology could threaten the future of humanity. Dr. LeCun argued that was absurd.
“There was a lot of noise around the idea that A.I. systems were intrinsically dangerous and that putting them in the hands of everyone was a mistake,” he said. “But I have never believed in this.”
LLMs are not the way:
“L.L.M.s are not a path to superintelligence or even human-level intelligence. I have said that from the beginning,” he said. “The entire industry has been L.L.M.-pilled.”
During his last several years at Meta, Dr. LeCun worked on technology that tried to predict the outcome of its actions. That, he said, would allow A.I. to progress beyond the status quo. His new start-up will continue that work.
“This type of system can plan what it is going to do,” he said. “Current systems — L.L.M.s — absolutely cannot do that.”
Part of Dr. LeCun’s argument is that today’s A.I. systems make too many mistakes. As they tackle more complex tasks, he argued, mistakes pile up like cars after a collision on a highway.
There's more at the link.
Okay so I need to talk about what’s happening with Yann LeCun because this is genuinely one of the wildest exits I’ve ever seen in tech.
— Mouad (@nadzi_mouad) January 5, 2026
For those who don’t know—LeCun is one of the “godfathers of AI.” Not a marketing title. The man literally won the Turing Award (basically the… pic.twitter.com/E1Fe1Gc0M9
V-JEPA-v2 https://t.co/7eWfLcJqce
— Yann LeCun (@ylecun) June 11, 2025
This is where you'll end up:
Introducing the V-JEPA 2 world model and new benchmarks for physical reasoning
Takeaways
- Meta Video Joint Embedding Predictive Architecture 2 (V-JEPA 2) is a world model that achieves state-of-the-art performance on visual understanding and prediction in the physical world. Our model can also be used for zero-shot robot planning to interact with unfamiliar objects in new environments.
- V-JEPA 2 represents our next step toward our goal of achieving advanced machine intelligence (AMI) and building useful AI agents that can operate in the physical world.
- We’re also releasing three new benchmarks to evaluate how well existing models can reason about the physical world from video.
Today, we’re excited to share V-JEPA 2, the first world model trained on video that enables state-of-the-art understanding and prediction, as well as zero-shot planning and robot control in new environments. As we work toward our goal of achieving advanced machine intelligence (AMI), it will be important that we have AI systems that can learn about the world as humans do, plan how to execute unfamiliar tasks, and efficiently adapt to the ever-changing world around us.
V-JEPA 2 is a 1.2 billion-parameter model that was built using Meta Joint Embedding Predictive Architecture (JEPA), which we first shared in 2022. Our previous work has shown that JEPA performs well for modalities like images and 3D point clouds. Building on V-JEPA, our first model trained on video that we released last year, V-JEPA 2 improves action prediction and world modeling capabilities that enable robots to interact with unfamiliar objects and environments to complete a task. We’re also sharing three new benchmarks to help the research community evaluate how well their existing models learn and reason about the world using video. By sharing this work, we aim to give researchers and developers access to the best models and benchmarks to help accelerate research and progress—ultimately leading to better and more capable AI systems that will help enhance people’s lives.
And so on and so forth. You'll get three videos and links to everything.
I've made that point before:
— Yann LeCun (@ylecun) January 25, 2024
- LLM: 1E13 tokens x 0.75 word/token x 2 bytes/token = 1E13 bytes.
- 4 year old child: 16k wake hours x 3600 s/hour x 1E6 optical nerve fibers x 2 eyes x 10 bytes/s = 1E15 bytes.
In 4 years, a child has seen 50 times more data than the biggest LLMs.… https://t.co/09atbzWsFP
A panel on AI at the Paris Peace Forum today.
— Yann LeCun (@ylecun) November 10, 2023
Starting at 42'00, I plot the future of AI: all of our interactions with the digital world will be mediated by AI assistants that will eventually become smarter than us.
Because they will become a common infrastructure containing all…
To those who say:
— Yann LeCun (@ylecun) October 15, 2023
"but closed source products have billions of investments behind them"
I reply:
In the mid-90a, Microsoft and Sun Microsystems battled to provide the software infrastructure of the internet, MS with WinNT+IIT+ASP+IE, Sun with Solaris+httpd+Java+Netscape.
They…
The lesson is that open-source software totally took over the world for servers and cloud (where Linux is overwhelmingly dominant), mobile devices (Android, which is Linux based), and pretty much all embedded devices.
— Yann LeCun (@ylecun) October 15, 2023
AI systems are fast becoming a basic infrastructure.
— Yann LeCun (@ylecun) September 21, 2023
Historically, basic infrastructure always ends up being open source (think of the software infra of the internet, including Linux, Apache, JavaScript and browser engines, etc)
It's the only way to make it reliable, secure, and… https://t.co/p5jBkKQgyq
Time to bump this to the top of the queue, 1) general principle (it's important), and 2) I'm thinking about this stuff at the moment.
* * * * *
Over in the discussion of Yann LeCun’s recent paper (A Path Towards Autonomous Machine Intelligence), Saty Chary has been arguing for something he calls the Structured Physical System Hypothesis (SPSH):
‘A structured physical system has the necessary and sufficient means for specific intelligent response’. By structured physical system, I mean, an analog design, e.g. a Rube Goldberg apparatus, or a Braitenberg (!) vehicle, etc. This is in contrast to this: PSSH - Physical Symbol System Hypothesis - 'A physical symbol system has the necessary and sufficient means for general intelligent action'.
I would add the slide rule as another example. From Wikipedia:
The slide rule is a mechanical analog computer, which is used primarily for multiplication and division, and for functions such as exponents, roots, logarithms, and trigonometry. It is not typically designed for addition or subtraction, which is usually performed using other methods. Maximum accuracy for standard linear slide rules is about three decimal significant digits, while scientific notation is used to keep track of the order of magnitude of results. [...]
At its simplest, each number to be multiplied is represented by a length on a pair of parallel rulers that can slide past each other. As the rulers each have a logarithmic scale, it is possible to align them to read the sum of the numbers' logarithms, and hence calculate the product of the two numbers.
My father used a slide rule for his entire career as an engineer. I learned to use one in my teens – everyone did back that – but never had any use for one. They’ve been replaced by cheap electronic calculators and PCs.
But that’s a digression. It’s the more general Structured Physical System Hypothesis that interests me, as Saty has been argued that the brain is such a system I agree (see this recent post, Once more around the merry-go-round: Is the brain a computer?). Here’s my reply to Saty:
Hi, Saty. I like your hypothesis – Structured Physical System Hypothesis (SPSH) – a lot. I think that a lot about the brain is consistent with it. For example, we know that mappings from one area to another as we move from the sense organs (or muscles), to the subcortex, and into the cortex tend to preserve topological relations between neurons. That’s of obvious value in the visual and motor systems. But there are subtleties. In the cortical visual system we have a so-called What-system and a so-called Where-system beyond the primary cortex. The What-system tracks location in space while the Where-system identifies objects. I assume that the Where-system has links to the hippocampus and I’d expect it can deal with both ego-centric and geocentric coordinates (this must be in the literature). But I’d think the What-system deals in object-centered coordinates. And so forth and so on.
In thinking about Freeman’s results (HERE and HERE), and others, I’ve coined a phrase: polyviscous connectivity. Thus I say that the cortical network as a whole exhibits polyviscous connectivity. What do I mean? Some connections are highly resistant to change, and thus have high viscosity. Others change quite readily, and have low viscosity. There is a literature on long-term (LTP) and short-term potentiation (STP) of neural connectivity that is certainly relevant here, but I’ve not looked at it in quite a while.
Consider Freeman’s results. He’s measuring neural activity with an 8 by 8 array of electrodes mounted on the cortical surface. They’re going to detect activity of neurons at varying levels of viscosity. Let’s a assume that the patterns of connectivity encoding odorants that rat already recognizes have a relatively high viscosity. Let’s further assume that the neurons most susceptible to learning new odorants have a relatively low viscosity.
Once they’ve formed a stable response to the new odorant, that will result in a new pattern of neural activity for the ensemble. But it is also going to change the patterns exhibited by already learned odorants even though the high-viscosity connections haven’t changed. The high viscosity connections maintain the overall integrity of the ensemble. In time, if the new odorant continues to be encountered, the connections registering it will increase in viscosity. So, polyviscous connectivity allows a structured connectionist physical system to maintain its overall integrity while adding new items to its repertoire.
More later.
* * * * *
Note: Looking around on my hard-drive I found an article which is about what I have called polyviscosity, though it doesn’t use that term:
Poonam Mishra and Rishikesh Narayanan, Stable continual learning through structured multiscale plasticity manifolds, Current Opinion in Neurobiology 2021, 70:51–63, https://doi.org/10.1016/j.conb.2021.07.009
Abstract: Biological plasticity is ubiquitous. How does the brain navigate this complex plasticity space, where any component can seemingly change, in adapting to an ever-changing environment? We build a systematic case that stable continuous learning is achieved by structured rules that enforce multiple, but not all, components to change together in specific directions. This rule-based low-dimensional plasticity manifold of permitted plasticity combinations emerges from cell type–specific molecular signaling and triggers cascading impacts that span multiple scales. These multiscale plasticity manifolds form the basis for behavioral learning and are dynamic entities that are altered by neuromodulation, metaplasticity, and pathology. We explore the strong links between heterogeneities, degeneracy, and plasticity manifolds and emphasize the need to incorporate plasticity manifolds into learning-theoretical frameworks and experimental designs.
* * * * *
Note: I’d previously been using the terms “hyperciscosity” or “hyperviscous”, and you’ll find them in my posts and notes on this topic going back to 2013 (when I wrote about From Associative Nets to the Fluid Mind). But I have reluctantly decided to coin a new term since “hyperviscosity” is already being used.
* * * * *
Addendum, 8.13.22: On the Structured Physical System Hypothesis, see this post where I feature remarks by Rodney Brooks, Has the computer metaphor for the mind run out of steam? As the title suggests, Brooks is wondering whether or not it makes sense to think about nervous systems in terms of computation. Thus he wonders:
Is information processing the right metaphor there? Or are control theory and resonance and synchronization the right metaphor? We need different metaphors at different times, rather than just computation. Physical intuition that we probably have as we think about computation has served physicists well, until you get to the quantum world. When you get to the quantum world, that physical intuition about stuff and place gets in the way.
13. Why do LLMs appear much better at generating code than generating general text?
— Yann LeCun (@ylecun) February 13, 2023
Because, unlike the real world, the universe that a program manipulates (the state of the variables) is limited, discrete, deterministic, and fully observable.
The real world is none of that.
After 50 years developing software I can definitely state that this is only true for the simplest cases. LLMs only appear to generate good code because they are only given the simplest cases and users are very forgiving of their mistakes.
— Marty Mouse (@MartyGargoyle) February 13, 2023
By comparison, see my old working paper, PowerPoint Assistant: Augmenting End-User Software through Natural Language Interaction (2013). Here's the abstract:
This document sketches a natural language interface for end user software, such as PowerPoint. Such programs are basically worlds that exist entirely within a computer. Thus the interface is dealing with a world constructed with a finite number of primitive elements. You hand-code a basic language capability into the system, then give it the ability to ‘learn’ from its interactions with the user, and you have your basic PPA.
I find it interesting that deep learning has done well with both images and texts, though they are very different kinds of objects presenting ostensibly different conceptual demands. So I decided to construct a thought experiment.
Let’s treat a text as a string of colored beads. We can assign each word the value of some color, any color, as long as each word TYPE is assigned a different value. Then we take each TOKEN in a text and replace it with a pixel having the color corresponding to its respective type. Now we’ve transformed a text into a string of pixels – color beads on a string. We do that for each text in a corpus and then model the corpus using the methods used in creating large language models.
To prompt the model we have to feed it a string of pixels. That’s easily done. We simply generate a natural language prompt, translate that prompt into pixel form and present those pixels to the model. The model will then extend the string in the normal way. We could even set up a pair of models so that one prompts the other with strings of pixels.
These pixel strings would be unintelligible to humans, nor are they likely to be very interesting as images. They’d just be a linear jumble of color. But these color jumbles are intelligible as texts if the proper word tokens are substituted for the color tokens.
You might object that this thought experiment is completely artificial. Well, yeah, it is. And in a way that’s the point. Consider this passage from a recent article Yann LeCun and Jacob Browning, What AI Can Tell Us About Human Intelligence:
This assumption is very controversial and part of an older debate. The neural network approach has traditionally held that we don’t need to hand-craft symbolic reasoning but can instead learn it: training a machine on examples of symbols engaging in the right kinds of reasoning will allow it to be learned as a matter of abstract pattern completion. In short, the machine can learn to manipulate symbols in the world, despite not having hand-crafted symbols and symbolic manipulation rules built in.
Contemporary large language models — such as GPT-3 and LaMDA — show the potential of this approach. They are capable of impressive abilities to manipulate symbols, displaying some level of common-sense reasoning, compositionality, multilingual competency, some logical and mathematical abilities, and even creepy capacities to mimic the dead. If you’re inclined to take symbolic reasoning as coming in degrees, this is incredibly exciting.
I think that second paragraph is misleading. In what sense are these LLMs manipulating symbols? They aren’t doing anything that isn’t done with those pixel strings in my (completely artificial) thought experiment. No one watching two large pixel-string models would think they were exchanging meaningful symbols. They’re just passing meaningless pixel strings back and forth.
The point is an old and obvious one, word meanings don’t exist in word forms, whether written or spoken. Word meanings exist in the minds of people speaking and writing. LLMs simply do not have access to those meanings. What’s interesting, and what we must understand, is that they can create such a convincing simulacrum of meaning based entirely on contextual information about word forms, signifiers.
And that’s what words and images have in common, context. But these contextual patterns are different in character. Images have a great deal of pixel-to-pixel continuity in 2D form. Continuity among words emerges only when their contextual relationships are represented in high dimensional spaces.
Yann LeCun recently posted a major position paper that has been receiving quite a bit of discussion:
Yann LeCun, A Path Towards Autonomous Machine Intelligence, Version 0.9.2, 2022-06-27, https://openreview.net/forum?id=BZ5a1r-kVsf
This post is a response to a long video posted by Dr. Tim Scarfe which raised a number of important issues. One of them is about physical constraints in the implementation of computing procedures and memory. I’m thinking this may well be THE fundamental issue in computing, and hence in human psychology and AI.
I note in passing that John von Neumann’s The Computer and the Brain (1958) was about the same issue and discussed two implementation strategies, analog and digital. He also suggested that the brain perhaps employed both. He also noted that, unlike digital computers, where you have and active computational unit linked to passive memory through fetch-execute cycles, that each unit of the brain, i.e. neuron, appears to be an active unit.
Physical constraints on computing
Here’s the video I was talking about. It is from the series Machine Learning Street Talk, #78 - Prof. NOAM CHOMSKY (Special Edition), and is hosted by Dr. Tim Scarfe along with Dr. Keith Duggar and Dr. Walid Saba.
As I’m sure you know, Chomsky has nothing good to say about machine learning. Scarfe is not so dismissive, but he does seem to be a hard-core symbolist. I’m interested in a specific bit of the conversation, starting about about 2:17:14. One of Scarfe’s colleagues, Dr. Keith Duggar, mentions a 1988 paper by Fodor and Pylyshyn, Connectionism and Cognitive Architecture: A Critical Analysis (PDF). I looked it up and found this paragraph (pp. 22-23):
Classical theories are able to accommodate these sorts of considerations because they assume architectures in which there is a functional distinction between memory and program. In a system such as a Turing machine, where the length of the tape is not fixed in advance, changes in the amount of available memory can be affected without changing the computational structure of the machine; viz by making more tape available. By contrast, in a finite state automaton or a Connectionist machine, adding to the memory (e.g. by adding units to a network) alters the connectivity relations among nodes and thus does affect the machine’s computational structure. Connectionist cognitive architectures cannot, by their very nature, support an expandable memory, so they cannot support productive cognitive capacities. The long and short is that if productivity arguments are sound, then they show that the architecture of the mind can’t be Connectionist. Connectionists have, by and large, acknowledged this; so they are forced to reject productivity arguments.
That’s what they were talking about. Duggar and Scarfe agree that this is a deep and fundamental issue. A certain kind of very useful abstraction seems to depend on separating the computational procedure from the memory on which it depends. Scarfe (2:18:40): “LeCun would say, well if you have to handcraft the abstractions then learning's gone out the window.” Duggar: “Once you take the algorithm and abstract it from memory, that's when you run into all these training problems.”
OK, fine.
But, as they are talking about a fundamental issue in physical implementation, it must apply to the nervous system as well. Fodor and Pylyshyn are talking about the nervous system too, but they don’t really address the problem except to assert that (p. 45), “the point is that the structure of ‘higher levels' of a system are rarely isomorphic, or even similar, to the structure of ‘lower levels' of a system,” and therefore the fact that the nervous system appears to be a connectionist network need not be taken as indicative about the nature of the processes it undertakes. That is true, but no one has, to my knowledge, provided strong evidence that this complex network of 86 billion neurons is, in fact, running a CPU and passive memory type of system.
Given, that, how has the nervous system solved the problem of adding new content to the system, which it certainly does? Note that here is their specific phrasing, from the paragraph I’ve quoted: “adding to the memory (e.g. by adding units to a network) alters the connectivity relations among nodes and thus does affect the machine’s computational structure.” The nervous system seems to be able to add new items to memory without, however, having to add new physical units, that is neurons, to the network. That is worth thinking about.
Human cortical plasticity: Freeman
The late Walter Freeman has left us a clue. In an article from 1991 in Scientific American (which was more technical in those days), entitled “The Physiology of Perception,” he discusses his work on the olfactory cortex. He’s using an array of electrodes mounted on the cortical surface (of a rat) to register electrical activity. Note that he’s NOT making recordings of the activity of individual neurons. Rather, he’s recording activity in a population of neurons. He then made 2-D images of that activity.
The shapes we found represent chaotic attractors. Each attractor is the behavior the system settles into when it is held under the influence of a particular input, such as a familiar odorant. The images suggest that an act of perception consists of an explosive leap of the dynamic system from the " basin" of one chaotic attractor to another; the basin of an attractor is the set of initial conditions from which the system goes into a particular behavior. The bottom of a bowl would be a basin of attraction for a ball placed anywhere along the sides of the bowl. In our experiments, the basin for each attractor would be defined by the receptor neurons that were activated during training to form the nerve cell assembly.
We think the olfactory bulb and cortex maintain many chaotic attractors, one for each odorant an animal or human being can discriminate. Whenever an odorant becomes meaningful in some way, another attractor is added, and all the others undergo slight modification.
Let me repeat that last line: “Whenever an odorant becomes meaningful in some way, another attractor is added, and all the others undergo slight modification.” That the addition of a new item to memory should change the other items in memory is what we would expect of such a system. But how does the brain manage it? It would seem that specific memory items are encoded in whole populations, not in one or a small number of neurons (so-called ‘grandmother’ cells). Somehow the nervous system is able to make adjustments to some subset of synapses in the population without having to rewrite everything.
In this connection it’s worth mentioning my favorite metaphors for the brain, as a hyperviscous fluid. What do I mean by that? A fluid having many components, of varying viscosity (some very high, some very low, and everything in between), which are intermingled in a complex way, perhaps fractally. Of course, the brain, like most of the body’s soft tissue, is mostly water, but that’s not what I’m talking about. I’m talking about connectivity.
Perhaps I should instead talk about a hyperviscous network, or mesh, or maybe just a hyperviscous pattern of connectivity. Some synaptic networks have extremely high viscosity and so change very slowly over time while others have extremely low viscosity, and change rapidly. The networks involved in tracking and moving in the world in real time must have extremely low viscosity while those holding our general knowledge of the world and our own personal history will have a very high viscosity.
In the phenomenon that Freeman reports, we can think of the overall integrity of the odorant network as being maintained at, say, level 2, where moment-to-moment sensations are at level 0. The new odorant is initially registered at level 1 and so will affect the level 1 networks across all odorants. That’s the change registered in Freeman’s data. But the differences between odorants are still preserved in level 2 networks. Over time the change induced by the new odorant will percolate from level 1 to level 2 synaptic networks. Thus a new item enters the network without disrupting the overall pattern of connectivity and activation.
That is something I just made up. I have no idea whether or not, for example, it makes sense in terms of the literature on long-term potentiation (LTP) and short-term potentiation (STP), which I do not know. I do note, however, that the term “viscosity” has a use in programming that is similar to my use here.
Addendum: Are we talking about computation in the Freeman example?
I have variously argued that language is the simplest operation humans do that qualifies as computation. Thus, earlier in the commentary on LeCun’s piece I have said:
I take it as self-evident that an atomic explosion and a digital simulation of an atomic explosion are different kinds of things. Real atomic explosions are enormously destructive. If you want to test an atom bomb, you do so in a remote location. But you can do a digital simulation in any appropriate computer. You don’t have to encase the computer in lead and concrete to shield you from the blast and radiation, etc. And so it is with all simulations. The digital simulation is one thing, and real phenomenon, another.
That’s true of neurons and nervous systems too. [...] However, back in 1943 Warren McCulloch and Walter Pitts published a very influential paper (A Logical Calculus of the Ideas Immanent in Nervous Activity) in which they argued that neurons could be thought of as implementing circuits of logic gates. Consequently many have, perhaps too conveniently, assumed that nervous systems are (in effect) evaluating logical expressions and therefore that the nervous system is evaluating symbolic expressions.
I think that’s a mistake. Nervous systems are complex electro-chemical systems and need to be understood as such. What happens at synapses is mediated by 100+ chemicals, some more important than others. It seems that some of these processes have a digital character while others have an analog character. [...] I have come to the view that language is the simplest phenomenon that can be considered symbolic, thought we may simulate those processes through computation if we wish. That implies that there is no symbolic processing in animals and none in humans before, say, 18 months or so. Just how language is realized in a neural architecture that seems so contrary to the requirements of symbolic computing, that is a deep question, though I’ve offered some thoughts about that in the working paper I mentioned in my original comment.
If the phenomenon Freeman describes is not about computation, and it is NOT according to my current beliefs, then how does the problem brought up by Fodor & Pylyshyn apply?
And yet there IS a problem, isn’t there. There is a physical network with connections between the items in the network. Those connections must be altered in order to accommodate a new phenomenon. We can’t just add a new item to the end of the tape. That is, it IS a physical problem of the same form. So perhaps this technicality doesn’t matter.