#115 – Dileep George: Brain-Inspired AI

Theme 1: Simulating the Brain Is Not Understanding It — The Fundamental Flaw of the Blue Brain Project

George believes that attempting to "reproduce" the brain through detailed simulation of individual neurons and large-scale interconnection, as the Blue Brain project does, is a dead end.

• Historical Context: The Blue Brain project initially attempted to simulate a cat brain, using the Hodgkin-Huxley model (simulating signal propagation in cables) to build ultra-detailed biophysical neuron models, then interconnecting them according to connectivity statistics from real neuroscience, and finally "powering them on to see what happens."
• Mechanism Deconstruction: George used a precise analogy — "Making a single neuron model 99% correct is like making a transistor model correct, then trying to assemble them into a microprocessor — if you don't understand Boolean logic and how gate circuits are implemented from transistors, you will never build a working processor." He cited a classic paper, Can a neuroscientist understand a microprocessor?, pointing out that the paper's conclusion is: even using all the tools of neuroscience, one cannot infer key functions like Boolean logic, clock signals, or architecture from the physical level of a chip.
• Data Chain: George emphasized that the experimental data from neuroscientists themselves are valuable, but "those data cannot simply be assembled and simulated; you must understand, from an information processing perspective, what the computational basis of these findings is and how they combine into a whole — this requires someone to painstakingly piece them together and form hypotheses."
• Deduction and Falsification: George advocates an iterative cycle — extract computational principles from neuroscience → build functional models that can run → use the models to fill missing pieces in biology → feed model predictions back to neuroscience experiments. Falsification condition: If one day neuroscience experiments prove that a key computational principle in the brain (e.g., the role of feedback connections) is completely different from the model's assumptions, the cycle must be revised.

---

Theme 2: The Core Secret of the Visual Cortex — Feedback Connections and Inference

George notes that the visual cortex is not merely a feedforward cascade; feedback connections far outnumber feedforward ones, and the brain is essentially a generative model that continuously performs inference.

• Mechanism breakdown: The visual cortex is roughly divided into four levels: V1 → V2 → V4 → IT (object recognition pathway). Each level contains about six layers of microcircuits, with extensive feedback connections both within and between layers. George illustrates with an example: When a triangle on a textured background is presented, the dynamics of neural activity are "first convergence from edges (contours), then filling in the surface inward from the edges" — this process is encoded in different layers of V1: which layer encodes edges, which encodes surfaces, and which handles feedback can all be tracked experimentally. He also mentions the famous Kanitsa triangle (where only three "Pac-Man" corners are shown, yet the brain "hallucinates" the complete triangle edges). Experiments show that when real edges are present, neurons respond immediately; whereas the response to illusory edges lags by tens of milliseconds, proving that feedback connections drive the "hallucination" — the brain actively projects an internal model onto the input.
• Data chain: George emphasizes that neuroscience experiments have become very precise: "You can ask, 'If I stimulate a neuron in the IT layer, what effect does it have in the V1 layer?' — the answer is that such experiments have already been done, and very systematically." He mentions a key figure: "In V1 itself, there are about 6 layers" (connections between layers have a specific structure).
• Inference: The brain does not passively receive input; it continuously performs "inference" — using an internal world model to interpret the current input, and through iterations of feedback and feedforward, arrives at the "best explanation." George likens this process to "explaining away" in probabilistic graphical models — for example, a home alarm might be triggered by either a burglar or an earthquake: when you hear the alarm and drive home, if the radio happens to report that an earthquake occurred nearby, your belief that a burglar broke in weakens because the "earthquake" cause "explains away" the evidence of the alarm. "The brain has the same mechanism, allowing multiple competing hypotheses to explain each other away, and the one that wins is your perception."

---

Theme 3: Recursive Cortical Network (RCN) — Breaking Deep Learning's Sample Efficiency Bottleneck with Reasoning

George introduced the RCN model developed by Vicarious, which he leads. The core breakthrough of this model is replacing deep learning's "massive training data + single feedforward" paradigm with dynamic reasoning, thereby achieving strong generalization with very few samples.

• Historical Context: RCN was published in a 2017 Science paper, used to crack the most difficult text CAPTCHAs at the time. George emphasized, "CAPTCHAs are far from solved — you can take a 5-year-old child who has just learned the alphabet, show them any CAPTCHA from a new website, and they can solve it without any training. Current mainstream deep learning systems cannot do this." He quoted Douglas Hofstadter: "The core problem of AI is 'What is the letter A?' — If you can reliably detect all variants of the letter A, you don't even need to touch B and C."
• Mechanism Breakdown: The key differences between RCN and convolutional neural networks (CNNs) are threefold:

1. Feedback Connections and Recursive Reasoning: CNNs are single feedforward; RCN has backward propagation in addition to forward propagation, forming iterative reasoning that dynamically aligns local evidence with global context.

2. Lateral Connections: CNN feature detectors have no coordination across different positions; RCN's lateral connections enforce consistency of feature transformations across different positions, preventing "jaggedness" and incoherence during generation (or inference).

3. Generativity: RCN is not a simple classifier; it outputs a complete interpretation of the entire scene — "This is an A, these pixels are parts of the A, these are parts I 'hallucinated', this is the contour, this is the interior, it is in front of another object." The model also incorporates priors such as foreground/background separation, shape and texture separation, which are not hand-coded but derived from the continuity properties of natural signals. "Natural signals" refer to signals from the real world (e.g., cats, dogs, trees) rather than artificial signals (e.g., QR codes). George emphasized: "Our brains are tuned for natural signals."

• Data Efficiency: On MNIST, RCN achieved 95% accuracy with only tens to hundreds of training samples, far surpassing other systems at the time. George explicitly stated, "RCN's entire training set is much smaller than the test set — this is exactly what we want to do: reverse deep learning's pattern of having a training set hundreds of times larger than the test set."
• Self-Defense and Risk Warning: George acknowledged, "Brain-inspired AI is often overhyped, much like quantum computing is used as a marketing term." However, he insisted his approach is serious — "I don't use 'brain-inspired' as a marketing term; I study the details of biology every day, racking my brain over those problems." He also shared an ironic anecdote: "After our paper was accepted by Science, their press office wrote a press release titled 'A New Deep Learning Model Cracks CAPTCHA' — even Science's press department used the term 'deep learning' for hype." This narrative reminds readers that even top-tier journals may deviate from the paper's original intent during dissemination.

---

Theme 4: Language is Simulation Control; GPT-3 Lacks a World Model

George argues that the breakthroughs of current large language models (e.g., GPT-3) are real, but their fundamental limitation is the absence of a world model, rendering them incapable of genuine causal reasoning and simulation.

• Mechanism Deconstruction: George uses a classic example: "Sally hammered a nail into the ceiling. Question: Is the nail horizontal or vertical?" You can instantly answer "vertical" because you simulate that scene in your visual-motor system—you imagine Sally standing, arm raised with a hammer, the nail driven vertically into the ceiling. All your everyday experience (not linguistic text) provides the answer. George points out: "Language is just simulation control; your perception + motor system is the real world simulator."
• Assessment of GPT-3: George's evaluation of GPT-3 is cautious—"It looks very interesting, and its text generation is more coherent than GPT-2, but its weakness is obvious: it does not carry world state, and sometimes generates self-contradictory sentences like 'I climbed the hill to reach the valley.' It does not run a simulation in its head to verify whether something holds in the real world." He refutes the view that simply scaling Transformers can approach AGI: "The information is simply not there—unless someone writes down all the common sense, such as 'walking forward is easier than walking backward,' 'the action of opening a door is to push then walk,' 'doctors also wear underwear,' the model cannot learn them from text."
• Conflict with Market Consensus: George's stance is clearly at odds with the mainstream view (which holds that larger models will approach AGI). He further points out that Transformers remain feedforward architectures, lacking recursion, feedback, causal intervention, and counterfactual reasoning capabilities. He offers a witty falsification condition: "If future GPT-10 can design experiments like a scientist, propose hypotheses, and revise hypotheses based on experimental results—rather than just solving 'simple puzzles'—then that would be AGI. But for now, watching it generate HTML code for a red button, while cool, is far from true understanding."
• Open Questions: George acknowledges, "We humans also hope that AGI is not that 'dumb'—but what if it is fundamentally 'dumb'? After all, all computation is connection and information transfer. Recursion and feedback loops seem to be more fundamental."

---

Mentioned Targets

Target	Guest Attitude	Key Data
Vicarious (company)	Founder background, no explicit position or investment advice	Founded around 2014, RCN model published in Science, achieved breakthrough on CAPTCHA
Numenta (company)	Founder background, no explicit position or investment advice	Early work centered on Hierarchical Temporal Memory (HTM), George co-founded with Jeff Hawkins
Neuralink (company)	Objective evaluation (optimistic about the application prospects of brain-computer interfaces, but questions the difficulty and safety of surgery)	Mentions competitor Paradromics, believes brain-computer interfaces may in the future enable "direct control of external limbs from the brain"
OpenAI / GPT-3 (model)	Risk warning (lacks world model, unable to perform causal reasoning and simulation)	175 billion parameters; George believes "scaling alone will not achieve AGI"

---

Judgments Worth Remembering

1. George believes that simulating the brain (e.g., the Blue Brain Project) is unreliable — without a theoretical framework, you cannot debug it; when it fails to behave as expected, you can only keep adding details without knowing what went wrong. Support: Analogy of "transistor model correct → microprocessor fails to run"; citing the conclusion of the paper Can a neuroscientist understand a microprocessor?

2. George points out that the number of feedback connections in the visual cortex far exceeds feedforward connections, and the brain is essentially a generative model that continuously performs inference — you use an internal world model to "explain" inputs, rather than passively receiving them. Support: The Kanizsa triangle experiment shows that neural responses to illusory edges occur tens of milliseconds later than those to real edges, driven by feedback connections; the probabilistic graphical model analogy of "alarm triggered by thief vs. earthquake" explains the explaining away mechanism.

3. George proposes that the core breakthrough of the RCN model is "replacing deep learning's heavy reliance on large training data and single-pass feedforward patterns with dynamic inference," thereby achieving strong generalization with very few samples. Support: On MNIST, it achieves 95% accuracy with just dozens to hundreds of samples; on CAPTCHA, it can solve any new CAPTCHA style (without training samples from that style), which current deep learning cannot do.

4. George argues that GPT-3's fundamental limitation is its lack of a world model, making it unable to perform causal reasoning and simulation — language is merely a simulation control, and true "understanding" is rooted in the sensorimotor system. Support: Using the example of "Sally hammering a nail into the ceiling" to show that the answer comes from visual-motor simulation, not linguistic statistics; noting that common sense like "walking forward is easier than walking backward" is never fully recorded in text, so the model cannot learn it.

5. George proposes the framework of "language as simulation control" — you use language to call simulators in the visual-motor system to answer questions; answering is not a natural language processing task, but the output of simulation. Support: This framework is the origin of the name Vicarious ("Vicarious" means "to experience indirectly" — running a model in the mind instead of physically interacting with the world).

6. George believes the core principle of brain-inspired AI is "understand the principles of the brain, but know when to deviate from it" — convolutional layers are a successful example of engineering deviation. Support: The brain does not have translation invariance (weight sharing); it relies on eye movements and fixation points to gain focus. Convolution is an efficient engineering trick, but it may be replaced in the future when robots have movable "eyes and heads."

7. George points out that memory requires an architecture where two types coexist: statistical (learning from regularities) and episodic (a unique timeline). The latter is stored as a pointer sequence in the hippocampus and then replayed through the cortex to generate abstractions. Support: When a new situation triggers a sense of déjà vu, the hippocampus indexes the past timeline, driving the cortex to replay that memory and inject it into the current perception.

8. George believes that a preliminary explanation of consciousness is that when you apply the same modeling system used to model the external world to your own modeling system itself, self-awareness arises — but whether "qualia" and the "urgency of death" are essential attributes of intelligence remains unknown. Support: He points out that "if AI can be replicated without loss, it would not have the urgency that comes with concurrent death — and humans may be goal-driven entities precisely because of this urgency." This is an open question.