Yoshua Bengio

Bengio's central early claim is that the performance of a learning system depends overwhelmingly on the representation of the data it is given, and that deep, hierarchical, distributed representations can be *learned* rather than hand-engineered, disentangling the underlying factors of variation (Bengio, Courville & Vincent 2013, "Representation Learning: A Review and New Perspectives," IEEE TPAMI; 12,969 citations). Key work: the TPAMI review; the *Deep Learning* textbook (Goodfellow, Bengio & Courville 2016). Implication: a model's competence is bounded by what its learned features actually capture, and brittleness lives in the gap between the training distribution and the deployment world.

Bengio's group introduced the attention mechanism for neural machine translation, letting a model learn *where to look* rather than compressing all context into a single fixed vector (Bahdanau, Cho & Bengio 2014, "Neural Machine Translation by Jointly Learning to Align and Translate," arXiv:1409.0473). This is the conceptual seed of the transformer era. Key work: the 2014 alignment paper. Implication: modern autonomous space agents inherit attention-based architectures; the same mechanism that gives flexibility also makes behavior hard to audit.

Bengio reframes the frontier as moving from fast, intuitive, in-distribution pattern recognition (System 1) toward deliberate, compositional, causal reasoning that generalizes out of distribution (System 2). Key work: his "From System 1 Deep Learning to System 2 Deep Learning" program and the consciousness-prior line of research. Implication: space environments are the definitional out-of-distribution case (novel debris geometries, unmodeled maneuvers, rare conjunctions), so a candidate must show their AI reasons rather than interpolates.

GFlowNets sample diverse candidate structures in proportion to a reward, enabling a model to maintain a *distribution* over plausible explanations rather than collapsing to a single point estimate (Bengio et al. 2021/2023, "GFlowNet Foundations," JMLR/arXiv:2111.09266). Key work: GFlowNet Foundations. Implication: directly relevant to maintaining calibrated uncertainty over orbital states and intent hypotheses instead of overconfident single predictions.

Bengio's flagship safety statement argues that capability and autonomy gains may "massively amplify AI's impact," risking large-scale harm, malicious use, and "an irreversible loss of human control over autonomous AI systems," and that society's response is "incommensurate" with the pace; the remedy pairs safety R&D with anticipatory, adaptive governance drawing on lessons from other safety-critical technologies (Bengio, Hinton, Yao, Song, Abbeel et al. 2024, "Managing extreme AI risks amid rapid progress," *Science*, DOI 10.1126/science.adn0117). Key work: the *Science* consensus paper and the longer preprint (arXiv:2310.17688). Implication: this is the master template a candidate's space-AI governance argument is measured against.

Bengio's recent constructive proposal is to build trustworthy, *non-agentic* AI: systems designed to model the world and explain it (a "Scientist AI") with probabilistic guardrails, rather than goal-pursuing agents that can lose alignment. Key work: his "Scientist AI" / safe-by-design agenda. Implication: maps onto a real design choice in space autonomy - decision-support that surfaces calibrated risk to a human controller versus closed-loop autonomous actuation.

As chair of the first *International AI Safety Report* (2025), Bengio institutionalized the idea that AI risk governance needs a shared, evidence-graded scientific baseline, separate from any single developer's claims, analogous to climate assessment bodies. Key work: the International AI Safety Report. Implication: a space candidate proposing AI for safety-critical functions (collision avoidance, custody, traffic management) must show an independent evidence base, not vendor benchmarks.