In the same vein as my AI: Are Hinton’s and Jorion’s Views on Co-Evolution Compatible? a fortnight ago. Today, this AI: How would François Chollet criticise Paul Jorion’s manuscript?.
The manuscript being entitled Rethinking Intelligence in the Age of Artificial Minds, forthcoming with Palgrave-Macmillan.
OpenAI o3 :
1 Where your manuscript resonates with Chollet
| Theme in your text | Why Chollet would nod along |
|---|---|
| Anti-anthropocentric stance – you stress that IQ tests and Bar-exam scores are parochial and propose an “Alien Intelligence Benchmark Suite” (AIBS) | Chollet’s own ARC benchmark was motivated by the same worry: standard leader-boards reward pattern-recall rather than abstraction. |
| Need for causal world-models – AIBS pillar #1 targets counter-factual prediction without linguistic labels | ARC tasks, and his recent program-synthesis work at Ndea, likewise force agents to infer latent generative structure, not surface statistics. |
| Goodhart & Ashby risks – you show how proxy metrics drift under optimisation pressure | Chollet often cites exactly these two cybernetic failures when he argues that bigger LLMs are not safer by default. |
| Hybrid future – you sketch neurosymbolic architectures where “a neural core supplies rich representations while symbolic modules verify them” | His public talks since 2023 have argued that AGI will be “deep-learning-guided program synthesis”, i.e. the same hybrid. |
Take-away: you already share several pillars of his worldview.
The criticisms below are therefore mostly directional adjustments, not wholesale disagreement.
2 Likely points of criticism
| Manuscript claim (paraphrased) | Chollet’s probable rebuttal | Supporting passage he might quote |
|---|---|---|
| “Scaling is the engine of the AI renaissance; semantics, syntax and even proto-selfhood bubble up as we just make the network bigger.” | Over-generalised. Chollet accepts emergent phenomena, but argues that sample-efficient skill acquisition stalls long before “general” intelligence. He would ask for empirical evidence that your scaling story still holds once tasks are truly out-of-distribution or require causal interventions. | “We realised that the trick was just to make the network bigger… And the semantics would follow” |
| AI has already ‘surpassed’ human intelligence in quality (invoking Hinton & Hofstadter). | Inflated benchmark. In On the Measure of Intelligence he distinguishes peak task performance from generalisation ability per unit experience. Current LLMs score high on the former, low on the latter, so “surpass” is premature. | “An artificial intelligence of a quality apparently higher than the natural one that was meant to be emulated” |
| Understanding is ballast; performance is separable from explanation. | He would agree that internal “verbal reasoning” is post-hoc, but still insists that explicit, inspectable programs (not opaque activations) are the safest route to reliability and debuggability. | “Performance is separable from explanation… understanding can be detached” |
| Compression-equals-intelligence thesis. | Chollet’s definition adds a second axis – priors. Two systems with equal compression might differ hugely in how little data they needed to reach it. He would likely call for a notion of efficiency missing in your current formulation. | “Intelligence may be best characterised as prediction and compression across modalities” |
| Benchmarks: AIBS vs ARC. | He might praise AIBS for attacking embodiment bias, yet ask: 1) does it measure data-efficiency? 2) do rich 3-D sims leak too many priors, letting pure memorisation creep back in? | |
| Psychoanalytic framing. | Chollet tends to avoid mentalistic vocabulary (“will”, “drive”) in technical writing; he might view the psychoanalytic detours as evocative but orthogonal to a falsifiable theory of AGI. |
3 Concrete tweaks to pre-empt (or incorporate) his critique
- Add an “efficiency” dimension.
Pair your compression metric with bits-of-experience consumed to solve each novel task. This answers his central question: “How much can you do with how little?” - Clarify the limits of scaling.
A short empirical appendix surveying where GPT-4-class models still fail (ARC private split, compositional generalisation, causal interventions) would balance the current “miracle of scaling” tone. - Position AIBS relative to ARC.
Explain why AIBS is a superset: ARC probes abstract reasoning in a language-free grid; AIBS extends the idea to longer horizons, embodiment and energy budgets. That shows lineage instead of rivalry. - Explain why psychoanalysis is more than metaphor.
If you keep that lens, tie each Freudian or Lacanian concept to a measurable computational analogue (e.g., repression ↔ gradient clipping on harmful activations). Otherwise readers in Chollet’s camp may dismiss it as narrative flourish. - Flag speculative passages.
You uphold the “covenant” by clearly labelling forward-looking sections (e.g., Phase III post-sovereign autonomy) as scenarios, not predictions. Chollet appreciates when visionary claims are quarantined from descriptive ones.
4 Bottom line
Your manuscript already speaks to several of François Chollet’s deepest concerns—non-anthropocentric metrics, causal abstraction, hybrid architectures.
Where he would press hardest is on data-efficiency, programme-level transparency and the real generalisation ceiling of scale-alone systems. Addressing those three axes—perhaps in a dedicated chapter or boxed sidebar—would make the book far less “open to Chollet’s criticism” and far more in dialogue with his agenda.