Reference
Marcus's claims are anchored to his own posts; the architecture and Mythos claims to public lab statements; the emotional-intuitive judgment is labeled as the author's.
The strongest anchors are Gary Marcus's own public writing on broad-but-shallow intelligence, hallucination, and scaling skepticism — used so the essay argues with his documented positions, not a caricature.
The architecture claims are anchored to public lab material from OpenAI, Google DeepMind, and Meta, kept at the level of what those companies describe in public, not their proprietary internals.
The Fable/Mythos case rests on Anthropic's own restriction statement and Dario Amodei's scaling-law policy essay, bounded to what the public record establishes. The reception reading about Marcus's ability to change adoption-committed readers' minds and the author's emotional-intuitive judgement are labeled author opinion, not sourced fact.
More on the references
A source counts here only if it could have changed, narrowed, or killed a claim. This is not a reading list.
The job was to distinguish three things that are easy to blur: Marcus's documented positions versus the author's synthesis; public architecture descriptions versus proprietary internals; and the Fable/Mythos dispute versus proof about either side's larger theory of AGI.
What the references did
Changed
- An external review removed the unsupported attribution that Marcus treats continued AI use as a moral failure.
- The Meta / LeCun world-model framing was corrected to match the public record.
- The Fable/Mythos case was narrowed to a split public-record verdict instead of evidence for one side.
Narrowed
- Marcus's ability to change adoption-committed readers' minds was held as a reception hypothesis and the author's reading, not a sourced fact.
- The lab-path claim became "no public evidence shows labs claim LLMs alone are the sole route to AGI," not a statement about private roadmaps.
- Architecture labels were tied to specific public programs rather than used as free-floating capability words.
Not imported
- No claim that Marcus is dishonest, corrupt, or arguing in bad faith.
- No claim that public architecture labels disclose proprietary internals, reliability, or readiness for authority.
- No claim that Fable/Mythos proves either side's larger theory about AGI.
- No claim that the editorial process itself proves the essay is true.
Source ledger
Open any entry to see the outside constraint and what it could have changed.
Marcus's own claims, used instead of a caricature.
The essay argues with positions Marcus has published himself — broad-but-shallow intelligence, hallucination as a structural property, and skepticism that scaling alone reaches AGI — so the disagreement is with the record, not the reflex.
Public sourcesAGI versus broad, shallow intelligence; why LLMs hallucinate; scaling skepticism.
How it could have changed the claimIf his documented positions did not match the essay's summary of them, the audit would be auditing a strawman.
What actually happenedConfirmed; the strongest sourced claims (limited usefulness, unreliability, not-AGI, current-AI risk distinct from AGI risk, market hype) were kept, and unsourced ones were cut or softened.
Remaining limitA reader should still check the posts; paraphrase compresses, and Marcus's positions can move over time.
Related claimSL-016: "Marcus is not simply anti-AI."
The Fable/Mythos restriction is a constraint, not a verdict for either side.
A frontier model restricted under a national-security directive is hard to square with a "merely autocomplete" reading — but a model being capable enough to worry a government is not the same as being dependable enough for institutional authority.
Public sourceAnthropic on Fable/Mythos access.
How it could have changed the claimIf the restriction were about something other than capability, the "weakens low-ceiling claims" reading would lose support.
What actually happenedConfirmed as a restriction under a national-security directive; used to split Marcus's claims into weakened (low-ceiling capability) and strengthened (governance, reliability, authority).
Remaining limitThe statement does not disclose Mythos's architecture or validated capability; the verdict is bounded by the public record.
Related claimSL-016: "Capable enough to be dangerous is not dependable enough for authority."
Scaling-law confidence is real, but it is not a bare-LLM-alone claim.
The essay needed to be fair about what lab leaders actually argue in public.
Public sourceDario Amodei, Policy on the AI Exponential.
How it could have changed the claimIf a leading lab publicly bet on bare next-token scaling as the sole route to AGI, Marcus's strongest target would shift from hype to lab doctrine.
What actually happenedThe public argument leans on scaling and "general cognitive capabilities" but couples them with policy, national strategy, and surrounding machinery — not a bare-LLM-alone claim.
Remaining limitPublic essays are not internal roadmaps; this bounds the claim to what is publicly stated.
Related claimSL-016: "No public evidence shows labs claim LLMs alone are the sole path to AGI."
LLM-centered reasoning and scaling: the first layer of the roadmap.
Used to describe the LLM-centered layer in plain terms — bigger models, better data, reinforcement learning, and test-time reasoning — without claiming it is the whole field.
Public sourcesOpenAI, learning to reason with LLMs; GPT-5.
How it could have changed the claimIf "reasoning" here were architecturally distinct from LLMs, the taxonomy's first layer would be mislabeled.
What actually happenedTreated as LLM-centered work, so the essay does not pass it off as evidence against Marcus's LLM critique.
Remaining limitPublic pages describe behavior and direction, not the proprietary internals.
Related claimSL-016: "The roadmap splits three ways."
Agentic and robotics systems: scaffolding around the model.
Used to describe the middle layer — agents, tools, and embodiment built around a core model — where capability grows but the core intelligence may still be LLM inference.
Public sourceGoogle DeepMind, Gemini Robotics 1.5.
How it could have changed the claimIf agentic/robotics work were clearly non-LLM at its core, it would belong in the "architecturally distinct" layer, not scaffolding.
What actually happenedTreated as scaffolding around a core model, so "agentic" is not used to imply a new architecture by itself.
Remaining limitThe public description does not settle how much of the core reasoning is LLM-based.
Related claimSL-016: "The roadmap splits three ways."
World models and modular architectures: the layer closest to Marcus.
Used to show that architecturally distinct work — world models and modular or neurosymbolic systems — is being pursued in public, which is closer to what Marcus advocates than "just scale the chatbot."
Public sourcesMeta, V-JEPA 2 world model; Meta / LeCun research program.
How it could have changed the claimIf no leading lab pursued architecturally distinct directions, the claim that the field is broader than LLM scaling would weaken.
What actually happenedConfirmed that world-model and modular programs exist in public; the Meta / LeCun framing was corrected during external review to match the record.
Remaining limitPublic research programs are not the same as deployed, validated capability.
Related claimSL-016: "The roadmap splits three ways."
Remaining risks
About this Reference record
A source counts here only if it could have changed, narrowed, or killed a claim. This is not a bibliography.
Remaining risks, plainly: public lab descriptions do not reveal proprietary internals, so the architecture taxonomy is a public-record reading, not a verified blueprint; the Fable/Mythos verdict is bounded by a restriction statement and a policy essay, not by audited capability; the reception reading and the author's emotional-intuitive judgement are explicitly the author's; and the disciplined audit constrains the claims but does not make them true.