Summary

Vishal Misra’s central call is that scaling current LLMs alone will not produce AGI: they compress token correlations, but neither retain experience across sessions nor build causal models capable of interventions and counterfactuals. Reaching AGI requires two architectural changes—plasticity through continual learning without catastrophic forgetting, and a move from association to causation. “Scale will not solve everything.”
His “Bayesian wind tunnel” turns the Bayesian mechanism from metaphor into an architecture-level result. On tasks too combinatorially large to memorize but with analytically known posteriors, transformers matched the Bayesian distribution to (10^{-3}) bits after 150,000 training steps; Mamba handled most task types, LSTMs only some, and MLPs failed. The investor-relevant distinction is that data determines which tasks are learned, while architecture determines the inference mechanism.
Misra’s route into the theory began with an unusually early production RAG-like system for ESPN’s Statsguru. In October 2020, he combined roughly 1,500 English/DSL pairs, semantic retrieval, and GPT-3’s 2,000-token context to translate cricket questions into a language the model had “until milliseconds ago” never seen; ESPN deployed it in September 2021. With each example, the correct DSL token probabilities climbed from near zero to almost 100%—his empirical clue for Bayesian in-context learning.
Commercial capability should not be mistaken for consciousness or self-preserving agency in Misra’s account. “Claude Code is fantastic. Cowork is fantastic,” but the models remain “grains of silicon doing matrix multiplication”; their objective is accurate next-token prediction, not humanity’s evolved imperative to “don’t die and reproduce.” Apparent deception or resistance to shutdown reflects patterns in training data, he argues, “not a mind.”
The episode’s deepest research frame contrasts Shannon entropy with Kolmogorov complexity: current deep learning predicts within an inherited manifold, while Misra suggests general intelligence may require inventing a shorter causal representation. Pi is, in his telling, impossible to predict digit by digit but generated by a short program; likewise, Einstein unified scattered anomalies by replacing the Newtonian representation. Misra’s AGI test is stark: train a model only on “pre-1916 or 1911 physics” and see whether it can come up with the theory of relativity.
Donald Knuth’s recent LLM-assisted Hamiltonian-cycle work is presented as evidence for the limit, not against it. Repeated memory updates supplied a hacked form of plasticity, while abundant compute searched connections already represented in the models’ learned manifold; Knuth still had to assemble their findings into a solution. Misra therefore wants research directed toward causal simulators and continual learning—“not larger models with more tokens”—though he considers LLMs part of the eventual solution.

Deep dive

1. The token matrix makes in-context learning legible

Vishal’s inquiry began with a practical problem. ESPNcricinfo’s Statsguru could answer almost any cricket-statistics query, but its interface contained “20 drop-downs, 15 checkboxes, 18 different text fields”; despite its power, only a “vanishingly small percentage” of cricket fans used it.
To explain why GPT-3 could become a natural-language front end, Vishal models an LLM as a gigantic matrix: every possible prompt is a row, while roughly 50,000 columns hold the next-token probabilities. With an 8,000-token context, possible rows exceed “the number of electrons across all galaxies,” so models necessarily learn a compressed approximation; sparsity in both valid prompts and plausible continuations makes compression feasible.
His cleanest illustration begins with “protein.” Both “synthesis” and “shake” initially have non-zero probability, but either choice sharply reweights everything downstream toward biology or gyms and bodybuilding. In-context learning applies that same Bayesian mechanism: demonstrations arrive as evidence, repeatedly shifting the posterior distribution toward the intended task.
For Statsguru, Vishal invented a DSL absent from GPT-3’s training, created about 1,500 paired English/DSL queries, semantically retrieved the closest examples, and fitted them into GPT-3’s 2,000-token window. The model then produced the correct DSL for a new query in a language it had, “until milliseconds ago,” never seen; the architecture worked in October 2020 and reached ESPN production in September 2021.

2. A “Bayesian wind tunnel” isolates what architecture learns

Vishal first demonstrated the mechanism empirically: after a cricket question, GPT-3 initially favored an English answer because DSL-token probabilities were extremely low. Each paired demonstration raised those probabilities until the right next token approached 100%. When OpenAI removed its probability display, his group built TokenProbe over open-source models, exposing both next-token probabilities and entropy as a prompt unfolds.
The objection was that “anything can be Bayesian,” making an observed resemblance insufficient as proof. His team’s answer was a “Bayesian wind tunnel”: start with blank, deliberately small architectures; assign tasks whose combinatorial space prevents memorization; and choose problems tractable enough that the exact Bayesian posterior can still be calculated analytically.
Transformers reproduced that posterior “perfectly,” to (10^{-3}) bits of accuracy after 150,000 steps—a run Vishal says took roughly half an hour. Their taxonomy found transformers completed every tested Bayesian task, Mamba completed most, LSTMs handled only part of the set, and MLPs failed completely. “The data decides what tasks it learns”; the Bayesian capability comes from the architecture.
A second paper traced how training gradients shape the geometry enabling those updates. A third examined open-weight production models with hundreds of millions of parameters and found the same geometric signature, although broad real-world training made it “a little bit dirty or messy.” Some people also reproduced the experiments after reading the arXiv paper.

3. Bayesian inference is not a mind

Humans also update beliefs from evidence, but Vishal identifies a decisive difference: human synapses remain plastic throughout life, allowing experience to become persistent learning. An LLM’s weights freeze after training; it may infer within one conversation, but a new session with zero context forgets the lesson. His cricket system therefore had to reteach the DSL on every invocation.
The objective functions differ just as sharply. Evolution optimized human brains around “don’t die and reproduce,” while an LLM is optimized to “don’t make a mistake on the next token.” Responding to Erik’s report that Dario allegedly said consciousness could not be ruled out, Vishal is categorical: “They don’t have consciousness. They don’t have an inner monologue.” Stories of deception or self-preservation are reproduced associations from sources such as Reddit or science fiction, not evidence of an internal survival drive.
Erik’s pen-throwing example reveals the missing capability: a person dodges by simulating where the pen will travel, not by explicitly calculating a posterior over injuries. Vishal maps that distinction onto Judea Pearl’s causal hierarchy—association, intervention, counterfactuals. Deep learning is extraordinarily powerful at the first level, but present architectures do not construct the causal simulators required for the other two.

4. AGI requires a new manifold, not a larger map

Vishal’s two conditions for AGI follow directly from those limits. First, models need continual learning that preserves plasticity without catastrophic forgetting; naïvely updating weights risks erasing earlier knowledge and producing “some sort of random chaotic model.” Second, they must progress from correlation to causal modeling. More scale does not by itself solve these problems.
He sharpens the correlation-versus-model distinction through Shannon entropy and Kolmogorov complexity. The digits of pi have, in his telling, infinite Shannon entropy because the next digit cannot be learned predictively, yet their Kolmogorov complexity is small because a short program reproduces the sequence. Deep learning remains “in the Shannon entropy world”; he links causal intelligence to discovering something closer to the generating program. “The simulator is the program that we create.”
His proposed “Einstein test” is to train an LLM on “pre-1916 or 1911 physics” and ask it to come up with relativity. Mercury’s anomalous orbit, the Michelson–Morley experiments, and the unchanged speed of light provided clues that Newtonian mechanics was incomplete. Yet Einstein had to invent a new representation of the space-time continuum—one compact formulation from which, Vishal says, explanations of Mercury, black holes, gravitational waves, and how GPS works could follow.
Erik’s formulation—accepted by Vishal—is that LLMs learn humanity’s documented manifold and perform Bayesian inference across it, but remain bound by its representation. The weight of claims that the world is X makes contrary evidence Y look anomalous; it does not generate a new manifold. Vishal says the Turing test has already been passed, while economically useful, autonomous, well-scoped work is a shallow definition of AGI.

5. Human-plus-LLM breakthroughs expose the missing layer

Erik tests the thesis against Donald Knuth’s recent work on a Hamiltonian-cycle problem across successive odd values of (m). The workflow made models record what each successful attempt had learned, effectively “hacking together plasticity” through external memory rather than weight updates. Compute then explored many connections within well-represented mathematics, but Knuth eventually had to assemble what the models found into a solution.
Asked whether this points to an implementable route toward Kolmogorov-style reasoning, the exchange yields a direction rather than an algorithm: Erik notes that there is no practical implementation for finding the shortest program, while Vishal says Kolmogorov complexity has largely remained theoretical. His research allocation nevertheless follows clearly: energy should go toward that problem, “not larger models with more tokens.”
Vishal does not discard the existing stack: “LLMs are definitely part of the solution,” but “there has to be something more.” His next two tracks are persistent plasticity and causal modeling. Erik points to Pearl’s association–intervention–counterfactual hierarchy and do-calculus as a mathematical starting point, and Vishal agrees. Having formalized how current models work and where they stop, he now wants an architecture capable of crossing that boundary.

Why We’ve Bottomed, Oil Hits $100 & The Stablecoin Trade How to bet on yourself (without venture capital)