The Science Behind the Mirror
This series is sci-pop, not a research paper. But the claims are not invented. Here is where they come from.
On attention and context windows
Claim: language models have finite attention windows, similar to human working-memory limits.
- Kahneman (2011), Thinking, Fast and Slow: attention established as a scarce cognitive resource.
- Vaswani et al. (2017), the transformer architecture: attention mechanisms explicitly modeled on selective attention.
- Context-window research at Anthropic and OpenAI: diminishing returns past a certain token count, mirroring human cognitive-load limits.
On confabulation and hallucination
Claim: models confabulate for the same reason humans do, gap-filling under uncertainty.
- Michael Levin, diverse-intelligence work: all cognitive systems construct plausible narratives from incomplete information.
- Hallucination research (Zhang et al., 2023; Anthropic interpretability): hallucinations rise with context dispersion and uncertainty.
- Human memory research (Loftus, Schacter): memory is reconstructive, not reproductive. We fill gaps with plausible fabrications.
On vectors and meaning
Claim: models operate on semantic vectors that encode relationships, not just word proximity.
- Mikolov et al. (2013), Word2Vec: vector arithmetic captures semantic relationships (king minus man plus woman lands near queen).
- Devlin et al. (2018), BERT and contextual embeddings: vectors capture context-dependent meaning.
- Fedorenko, Piantadosi, and Gibson (2024): language is not sufficient for thought, suggesting meaning exists independently of linguistic tokens.
On multi-signal motivation
Claim: complex behavior needs multiple coordinating signals, not single-objective optimization.
- Neuromodulation research (dopamine, serotonin, acetylcholine systems): the brain uses many parallel reward and attention signals.
- RLHF limitations (Anthropic, OpenAI): single-signal optimization produces shallow alignment.
- Constitutional AI (Anthropic, Bai et al. 2022): multi-principle alignment produces more robust behavior.
On agent societies
Claim: multi-agent coordination follows patterns from human organizational theory.
- A2A (agent-to-agent) protocol specifications: explicit negotiation, capability declaration, reputation tracking.
- Internet of Agents white papers: architectural patterns mirror organizational design, with registries, contracts, and oversight.
- Multi-agent reinforcement learning: cooperative emergence requires explicit coordination mechanisms.
The mirror is still forming. The science is still moving. But the direction is clear. We are not building artificial minds. We are building digital partners.
Related reading: The Vector turns the vectors entry in this annex back into the essay it came from: how meaning is stored, and why text can be deeper than video.
Sources
- Daniel Kahneman, Thinking, Fast and Slow, Farrar, Straus and Giroux, 2011.
- Ashish Vaswani et al., “Attention Is All You Need,” NeurIPS, 2017.
- Michael Levin, “Technological Approach to Mind Everywhere (TAME): An Experimentally-Grounded Framework for Understanding Diverse Bodies and Minds,” Frontiers in Systems Neuroscience 16: 768201, 2022.
- Yue Zhang et al., “Siren’s Song in the AI Ocean: A Survey on Hallucination in Large Language Models,” 2023.
- Elizabeth F. Loftus and John C. Palmer, “Reconstruction of Automobile Destruction: An Example of the Interaction Between Language and Memory,” Journal of Verbal Learning and Verbal Behavior 13(5): 585-589, 1974.
- Daniel L. Schacter, The Seven Sins of Memory: How the Mind Forgets and Remembers, Houghton Mifflin, 2001.
- Tomas Mikolov et al., “Efficient Estimation of Word Representations in Vector Space,” 2013.
- Jacob Devlin et al., “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,” 2018.
- Evelina Fedorenko, Steven T. Piantadosi, and Edward A. F. Gibson, “Language is primarily a tool for communication rather than thought,” Nature 630: 575-586, 2024.
- Yuntao Bai et al., “Constitutional AI: Harmlessness from AI Feedback,” Anthropic, 2022.
- Google, “Agent2Agent (A2A) Protocol Specification,” 2025.
- Yuntao Wang et al., “Internet of Agents: Fundamentals, Applications, and Challenges,” 2025.