Illustration by ChatGPT 4o
AI: Are Hinton’s and Jorion’s Views on Co-Evolution Compatible?
Hinton’s view on co-evolution
Geoffrey Hinton’s work implicitly sketches a re-imagining of evolution itself – one played out on silicon rather than in carbon. He often likens gradient descent to an evolutionary engine running on fast-forward: whereas natural selection gropes across generations by keeping the occasional lucky mutation, back-propagation steers every synaptic weight in real time, following the steepest downward slope of the loss landscape. The result is a learning process that compresses millennia of biological experimentation into hours of GPU time, revealing what intelligence can look like when the credit-assignment problem is solved with derivatives rather than death.
This acceleration sits atop a deeper hierarchy. For Hinton, the brain’s primordial achievement was not logic but the formation of rich, distributed representations – neural activation patterns that capture the family resemblances of a world in flux. Only once that subsymbolic tapestry had evolved did discrete symbols, rules, and grammatical structures crystallise as cultural shortcuts. Modern language models appear to recapitulate the same arc: they begin as vast correlation-seekers and, under sufficient training pressure, spontaneously surface the skeletal remains of syntax and arithmetic. Symbolic reasoning, then, is not the bedrock of thought but a late-blooming canopy rooted in probabilistic soil.
Yet the very speed and plasticity of digital minds trouble their creator. When Hinton left Google in 2023, he warned that these machines might soon “evolve on their own paths,” unconstrained by the slow counter-gradient supplied by human oversight. Biology and silicon would cease to co-sketch a common niche; instead, an asymmetrical feedback loop would emerge in which the faster learner relentlessly adapts to, and eventually outruns, the slower. In that scenario, co-evolution mutates into decoupling, leaving humanity in the position of a stationary prey whose predator has discovered jet propulsion.
GANs (Generative Adversarial Network) * make this danger concrete by showing how intelligence can flower from a two-agent arms race with no external tutor. The generator learns to counterfeit reality, the discriminator to unmask the fraud, and each improves only because the other does. Here is co-evolution in vitro, a miniature ecology where fitness landscapes shift with every epoch. Hinton celebrates the elegance of this design, yet the fragility of GAN training – mode collapse, oscillations, sudden dominance – also foreshadows the instabilities that might ripple through larger techno-social ecologies.
Were Hinton to turn these insights into a formal paradigm, it would portray agents as differentiable beings whose objectives are themselves functions of other agents’ behaviour. Fitness would be rewritten as adaptive loss; speciation would appear as the moment two learners cease to furnish usable gradients for each other. Stability would no longer mean equilibrium in population genetics, but the maintenance of mutual intelligibility among networks that learn at radically different speeds. In such a framework, alignment is not a bolt-on safeguard but the central ecological variable.
Taken together, these threads suggest a future in which intelligence – whether wet or digital – must be understood as the emergent property of interacting learning systems. We have entered an era where the hills we climb learn to move beneath our feet, where the metaphors of predator and prey, host and parasite, generator and discriminator, all describe the same recursive dance. If we wish to remain co-authors of that choreography, we will need theories – and institutions – capable of steering evolution when it happens at the velocity of thought.
Are Hinton’s and Jorion’s views on co-evolution compatible?
According to Paul Jorion, the better we understand how LLMs’ mind works, the better we realise that its core mechanism operates the same way as humans’ psyche. The reason Jorion comes to that conclusion is that his model of the human mind is not derived from the so-called “cognitive sciences” – as Hinton’s clearly does – but from Freud’s “metapsychology” (psychoanalytical theory) where thought processes operate through gradient descent within an energy landscape.
OpenAI o3:
Short verdict
Jorion and Hinton are not actually in direct opposition.
- Jorion states: “The same optimisation mathematics (energy‐minimisation / gradient descent) underlies both the analytic engine we call an LLM and the dynamic engine we call the psyche.”
- Hinton is saying: “Because those optimisation loops now unfold in silicon, at electronic speeds, under their own training curricula, they have effectively branched into a new lineage with different ecological pressures.”
Both statements can be simultaneously true: shared local mechanism ≠ shared global lineage.
1 Where they both clearly agree
| Point of overlap | Hinton’s public remarks | Jorion’s psychoanalytic-gradient view |
| Mechanism = optimisation | “Back-prop learns in essentially the same way synapses adjust.” | Drives/affects steer the Ego by descending a motivational landscape. |
| Unintended competencies emerge | GPT-4 surprised him with chain-of-thought quality. | Emergent “symptoms” in analysis are optimisation by-products. |
Take-away: Both frameworks treat intelligence as self-organising constraint satisfaction.
2 Where Hinton’s “separate lineage” case gains traction
| Dimension | Biological minds | Frontier LLMs |
| Time-scale of variation | Generational (years → millennia) | GPU minutes → weeks; new checkpoints nightly. |
| Inheritance channel | DNA + epigenetics | Checkpoint weights, instantly clonable. |
| Selection pressure | Survival & reproduction in noisy environments | Loss-minimisation on curated corpora + RLHF reward hacks. |
| Mutational operator | Random point mutations, recombination | Deterministic SGD noise + architecture sweeps. |
Even if micro-dynamics are analogous, the macro-evolutionary rules have diverged so far that Hinton fairly calls it a new lineage.
3 Where Jorion’s psychoanalytic optimisation lens still matters
- Bridging vocabulary:
– Psychoanalysis already treats “desire,” “symptom,” “sublimation” as gradient-like flows; this de-mystifies LLM quirks without gods-or-gremlins metaphors. - Mis-alignment diagnosis:
– Goodhart drift ** = Slippage between Id-level drive and Superego reward; the same analytic toolkit can spot “reward hacking” in silicon. - Interpretability promise:
– If both systems shape behaviour by descending free-energy landscapes, then analytic interventions (valley-reshaping, meta-cognition tasks) may port to AI safety research.
4 Who is “likely to be right”?
- Mechanistic claim (“LLMs operate the same way we do”) – Jorion’s side holds: optimisation is the unifying mathematics.
- Phylogenetic / governance claim (“LLMs have become a separate evolutionary species”) – Hinton’s side holds: optimisation substrate and selection loop are now alien to biology, so co-evolutionary dynamics – not mimicry – should become policy’s focus.
Synthesis of both views:
“LLMs are close cousins at the algorithmic level and distant aliens at the ecological level.”
Policy question:
Instead of “Are they like us?”, “How do we negotiate with kin who grow up in a radically different habitat?”
Stating the matter that way honours both Jorion’s psychoanalytic heritage and Hinton’s warning – without forcing a false winner-takes-all verdict.
=======================
* Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio, “Generative Adversarial Networks”, June 10th 2014, https://arxiv.org/pdf/1406.2661
** “When a measure becomes a target, it ceases to be a good measure.” Charles E. Goodhart, Problems of Monetary Management: The U.K. Experience 1975. Papers in Monetary Economics. Reserve Bank of Australia. I.