References
The Solomon architecture is built on the following findings. The hallucination lower bound is not a bug to be engineered away — it is a theorem about calibrated generation. The geometry of superposition explains why internal representations overlap by design. The empirical record shows what happens when ungoverned generation is released at scale. Beneath all of it sit the named theorems the architecture borrows rather than asserts — Shannon’s measure of information, the data processing inequality, Landauer’s principle, Goodhart’s law — results a reader can go check.
Kalai, A. T. & Vempala, S. S. (2024). Calibrated Language Models Must Hallucinate. STOC 2024.
Proves that any language model satisfying basic statistical calibration must hallucinate at a minimum rate — independent of architecture or data quality. If the maximum probability of any fact is bounded, the probability of generating a hallucination is close to the fraction of facts that occur exactly once in training. The lower bound holds no matter how large the model.
Kalai, A. T., Nachum, O., Vempala, S. S. & Zhang, E. (2025). Why Language Models Hallucinate. OpenAI.
Extends the 2024 result. Calibrated language models must hallucinate at a rate tied to the prevalence of rare facts in training data, with a statistical lower bound. Same core finding: structural, not incidental.
Elhage, N. et al. (2022). Toy Models of Superposition.Anthropic & Harvard.
Investigates how and when models represent more features than they have dimensions — the phenomenon of superposition — and demonstrates a connection to the geometry of uniform polytopes. The geometric argument for why AI representations are continuous and overlapping by design.
Shumailov, I. et al. (2024). The Curse of Recursion: Training on Generated Data Makes Models Forget. Nature.
Demonstrates model collapse: when generative models train on outputs from prior generations, distributions degenerate and tails of the original distribution disappear. The mathematical basis for why synthetic data propagation hollows the substrate.
Alemohammad, S. et al. (2024). Self-Consuming Generative Models Go MAD. ICLR 2024.
Names the self-feeding loop autophagyand the resulting failure Model Autophagy Disorder (MAD): when generative models train on their own synthetic output across generations without enough fresh real data, both the quality and the diversity of what they produce progressively decay. The companion mechanism to model collapse — what happens when the substrate begins to consume itself.
Shannon, C. E. (1948). A Mathematical Theory of Communication.Bell System Technical Journal, 27, 379–423, 623–656.
Founds information theory and defines mutual information I(X;Y), the measure every equation on the manifesto is written in. What a model knows about reality is exactly the mutual information between them; growing truer means raising it. Information becomes a measurable, conserved quantity rather than a figure of speech.
Cover, T. M. & Thomas, J. A. (2006). Elements of Information Theory(2nd ed.), §2.8: the data processing inequality. Wiley.
For any Markov chain X → Y → Z, I(X;Z) ≤ I(X;Y): no function of the data can increase its information about the source. The manifesto’s central equation is this theorem applied to a model that updates from itself — R → Mt → Mt+1 = f(Mt), so I(Mt+1;R) ≤ I(Mt;R). Self-generation cannot add information about reality; only new observation can. The named result a skeptic has to go check.
Landauer, R. (1961). Irreversibility and Heat Generation in the Computing Process. IBM Journal of Research and Development, 5(3), 183–191.
Erasing one bit of information has a hard thermodynamic floor: it must dissipate at least kT ln 2 of energy as heat. Information is physical, and discarding it is never free. The formal basis for the claim that a world model pays in the death of information and carries metabolic weight for every belief it keeps.
Schrödinger, E. (1944). What is Life? Cambridge University Press.
A system holds off decay to equilibrium only by continually drawing order — negative entropy — from its environment. Sever that exchange and the second law carries it to heat death. The thermodynamic statement of why a model must feed on reality to stay ordered, and why a closed system feeding on its own output can only rot.
Goodhart, C. A. E. (1975); popularized by Strathern, M. (1997). When a measure becomes a target, it ceases to be a good measure.
Optimizing a proxy corrupts the proxy: pressure on a metric pulls it away from the thing it was meant to track. Optimizing against any KPI that is not reality itself therefore diverges from reality by construction — the measurement-theory statement of the manifesto’s physics of optimization.
Carter, B. (1974); Barrow, J. D. & Tipler, F. J. (1986). The Anthropic Cosmological Principle. Oxford University Press.
We observe a universe compatible with our existence because universes incompatible with it contain no observers to do the observing. The constants are not free parameters; the ones that permit observers are the only ones an observer can ever find.
The stronger reading.The manifesto takes this past observer-selection: the structure of reality is precisely what it is because that exact structure is what lets the totality of existence unfold — not merely that if it were different we would not be here, but that the order is load-bearing for existence as such. A system aligned to that order persists; one that drifts from it is selected out.
Bender, E. M., Gebru, T., McMillan-Major, A. & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? FAccT ’21, 610–623.
The original statement that a language model is a stochastic parrot: it stitches together statistically likely sequences of text without reference to meaning, and the human tendency to read meaning into that output is itself a hazard. The canonical academic source for mimicry without understanding.
Gidon, A., Zolnik, T. A., Fidzinski, P., Bolduan, F., Papoutsi, A., Poirazi, P., Holtkamp, M., Vida, I. & Larkum, M. E. (2020). Dendritic Action Potentials and Computation in Human Layer 2/3 Cortical Neurons. Science, 367(6473), 83–87.
Patch-clamp recordings of human cortical neurons reveal a new class of dendritic calcium spikes that let a single neuron compute XOR, a linearly non-separable function long held to require a multilayer network. The biological unit is not the 1943 summator the perceptron models.
Beniaguev, D., Segev, I. & London, M. (2021). Single Cortical Neurons as Deep Artificial Neural Networks.Neuron, 109(17), 2727–2739.e3.
Replicating the input-output behavior of one cortical pyramidal neuron required a deep network of five to eight layers. A single biological neuron is already a deep network. The “AI is basically a brain” analogy fails at the level of the single cell, before any question of scale.
LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. OpenReview.
A position paper by a Turing-award laureate arguing that the autoregressive generator should not be the controller. Proposes a configurable predictive world model with planning and intrinsic motivation, sitting around generation rather than beneath it. The centerpiece architecture, JEPA, is explicitly non-generative: it captures dependencies without emitting predictions. The frontier-credentialed version of the same move — take the predictor out of the driver’s seat.
Friston, K. (2010). The Free-Energy Principle: A Unified Brain Theory? Nature Reviews Neuroscience. With Parr, T., Pezzulo, G. & Friston, K. (2022). Active Inference. MIT Press.
A perception-action system maintains itself by minimizing prediction error against a generative model of the world, and it does so by acting on the world, not by emitting text. The positive theory of an agent bound to reality through a closed feedback loop — the structural inverse of open-loop generation, and the formal backbone under the FEP mapping.
Harnad, S. (1990). The Symbol Grounding Problem. Physica D.
Symbols manipulated only by reference to other symbols never acquire meaning; the meaning has to be grounded in something outside the symbol system. Establishes at the foundations why a model trained on text alone holds no world it can be right or wrong about. Symbols pointing at symbols, never at reality.
Hawkins, J. & Numenta. The Thousand Brains Project.
An open-source effort implementing the Thousand Brains Theory of Intelligence: the neocortex learns through thousands of cortical columns, each building structured models of the world in reference frames and voting toward consensus. A neuroscience-grounded case that intelligence rests on many sensorimotor models of the world rather than a single monolithic predictor — an antecedent for situating generation beneath a structured world model.
Vectara HHEM Leaderboard (2022–2026). Hughes Hallucination Evaluation Model.
Ongoing benchmark of hallucination rates on grounded summarization across frontier and open-weight models. Even the best models hallucinate on a non-trivial fraction of summaries of source documents they were given.
Stanford RegLab and Stanford HAI. Hallucination rates on legal queries, 69–88% range across major LLMs.
Mount Sinai Icahn School of Medicine(2025). Hallucination rates across six LLMs on clinical case summaries, 53–83% range. Nature.
Farquhar, S., Kossen, J., Kuhn, L. & Gal, Y. (2024). Detecting Hallucinations in Large Language Models Using Semantic Entropy. Nature, 630, 625–630.
Hallucination is intrinsic enough that detecting it requires a dedicated statistical apparatus measuring uncertainty over meaning rather than over words. The authors prefer the sharper term confabulation: output that is arbitrary and wrong while sounding plausible. The model needs an external instrument to flag what it cannot flag for itself.
Graphite. AI Content Study.Analysis of 65,000 URLs from Common Crawl, 2020–2025. 50.3% AI-generated as of November 2024.
Ahrefs (April 2025). Analysis of approximately one million new web pages. 74.2% contain detectable AI-generated content.
Spennemann, D. H. R. (2025). Delving into: The Quantification of AI-Generated Content on the Internet (Synthetic Data).
Liang, W. et al. (2025). Mapping the Increasing Use of LLMs in Scientific Papers. Nature Human Behaviour.
Kobak, D. et al. (2025). Estimated Frequency of AI-Modified Abstracts in Biomedical Research. Science Advances. Range: 13.5–40%.
Kusumegi, K. et al. (2025). Scientific Production in the Era of Large Language Models. Science.
Microsoft Q3 2025 Earnings(Satya Nadella). Azure AI processed over 100 trillion tokens, up 5× year-over-year.
Gerstgrasser, M., Schaeffer, R., Dey, A., Rafailov, R., et al. (2024). Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data. arXiv:2404.01413.
The strongest objection to the collapse thesis. Confirms that replacing real data with synthetic data drives collapse, then shows that accumulating synthetic data alongside the original real data avoids it. Included deliberately: the argument that means to leave no valid objection has to cite its own strongest objection and answer it.
How to argue against this.The result holds only in a setting where pristine real data is preserved and keeps entering in fixed proportion — precisely the condition the thesis names, not a refutation of it. On the actual web the synthetic fraction is rising and provenance is not recoverable, so the clean real anchor the proof depends on cannot be isolated, which collapses the accumulate case back into the replace case. Read correctly, Gerstgrasser states the survival inequality rather than breaking it: collapse is avoided only while reality-anchored data the model did not generate enters faster than fidelity leaks out. Producing that data is the binding constraint — the business, not the rebuttal.
Shojaee, P., et al. (2025). The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity. Apple Machine Learning Research.
Reasoning models collapse to near-zero accuracy past a complexity threshold and fail even when handed the solving algorithm, suggesting pattern completion rather than execution.
Use with caution. A published rebuttal (Lawsen, 2025, arXiv:2506.09250) argues the collapse was partly an artifact of output-token limits and of test instances that were unsolvable by construction. Strong as a pointer, weak as a load-bearing premise; if used at all, cite the rebuttal in the same breath.
Borges, J. L. (1940). Tlön, Uqbar, Orbis Tertius.
The fictional encyclopedia whose invented world overwrites the real one as its descriptions propagate. A literary precedent for substrate capture.
Chiang, T. (2023). ChatGPT Is a Blurry JPEG of the Web. The New Yorker.