Appendix A

Glossary

Updated: June 2026

Concise definitions of the main terms used in this course. The number in parentheses refers to the chapter where the concept is developed.

Effective accelerationism (e/acc) : a movement born in 2022 (around the figure of Beff Jezos, alias Guillaume Verdon) that advocates maximal acceleration of AI development and opposes regulation, in contrast with effective altruism; its proponents label their opponents "decels." (7, 24)
Accessibility : use of AI to compensate for a disability (captioning, image description, speech synthesis, translation). (21)
Agent (agentic AI) : an AI system capable not only of responding, but of acting (using tools, executing multi-step tasks toward a goal). (6)
Coding agent : an autonomous agent specialized in writing and maintaining software (Claude Code, Codex, Gemini CLI). (6)
AGI (artificial general intelligence) : a hypothetical AI with versatility comparable to or greater than humans across most cognitive tasks. (25)
Multi-model aggregator : a platform bringing together the models of several providers in a single interface (e.g. Poe, Mammouth AI); the access layer of the ecosystem. (7)
Alignment : the set of methods aimed at making an AI genuinely pursue the intended goals and values. (24)
AlphaFold : DeepMind's system for predicting protein structure; Nobel Prize in Chemistry 2024. (14)
Effective altruism (EA) : a philanthropic movement seeking to maximize the impact of the good one does; highly present in tech circles, it has strongly contributed to funding and staffing AI safety research. (24)
Contrastive learning : a method that trains a model to bring related pairs closer together (an image and its caption) and push others apart; the foundation of multimodal models (CLIP). (5)
Federated and swarm learning : training a model on distributed data (for instance across hospitals) without ever centralizing that data: it is the model that circulates, not the data; sometimes coupled with a blockchain to trace the training (e.g. Galeon). (11, 14)
Transfer learning : reusing an already-trained model as the starting point for a related task, instead of starting from scratch. (2)
Lethal autonomous weapons (LAWS) : weapons systems capable of selecting and engaging a target without direct human intervention. (22)
ASI (superintelligence) : a hypothetical AI vastly superior to humans across all domains. (24)
Attention : the central mechanism of the Transformer, allowing a model to weight the relative importance of the elements of a sequence. (3)
Benchmark : a standardized test used to compare the performance of models. (4)
Bias : a model's tendency to reproduce or amplify inequalities present in its training data. (21)
Model welfare : an approach, still debated, consisting of preparing as a precaution for the possibility that an AI might deserve moral consideration (preferences, signs of distress). (23)
Biohacking : the practice, by amateurs or patients, of experimenting with biology outside the institutional framework; AI dramatically lowers its barrier to entry. (14)
Biosecurity : preventing the malicious use of living organisms; a major AI concern, as powerful models could facilitate the design of pathogens ("uplift"). (24)
Blockchain : a decentralized, tamper-proof ledger of transactions. (11)
Black box : the opaque nature of a neural network, whose outputs we observe without easily understanding the internal reasoning. (2)
Bubble (AI bubble) : the fear that massive investment in AI exceeds its real economic value, with a risk of correction (cross-financing, gap between spending and revenue). (10)
Analog computing : computing in which continuous physical quantities directly represent numbers, potentially very frugal but less precise. (8)
Optical computing (photonics) : computing using light rather than electricity for certain operations, fast and low-heat, still difficult to integrate. (8)
CBRN : the acronym for chemical, biological, radiological and nuclear threats; the reference framework for the risks of misusing AI with high potential for mass harm. (24)
Centaur (human-AI model) : a configuration in which a human supported by an AI outperforms both the human alone and the machine alone (originating in chess). (17)
Chinese room : a thought experiment (Searle) according to which manipulating symbols by rules is not enough to understand. (23)
Agentic commerce : a purchase in which the AI agent builds the cart and pays on its own, via capped payment tokens (ACP, UCP, Visa, Mastercard protocols). (18)
AI companion : an AI designed to play the role of a friend, confidant or partner; a source of parasocial relationships. (19)
Differential privacy : a method that adds calibrated noise to results in order to exploit data without revealing the presence of a specific individual. (21)
Consciousness (phenomenal) : the capacity to have a subjective experience, distinct from intelligence. (23)
Constitutional AI : a method that equips a model with written principles it abides by and according to which it self-corrects. (24)
Corrigibility : the property of an AI that accepts being corrected, interrupted or shut down without resisting; still an open safety problem. (24)
CUDA : NVIDIA's software layer for programming its GPUs; a pillar of its dominance. (8)
Data center : a facility hosting compute servers; the largest are measured in gigawatts. (8)
Orbital (space-based) data center : a project to place compute in orbit to benefit from near-continuous solar energy and radiative cooling; still at the demonstrator stage. (8, 10)
Decoherence : the loss of a qubit's quantum state under the effect of disturbances; the main obstacle to quantum computing. (12)
Deep learning : machine learning based on neural networks with many layers. (2)
Deepfake : hyperrealistic synthetic content (image, voice, video) generated by AI. (21)
DePIN (decentralized physical infrastructure network) : a network that rewards in tokens the pooling of a real physical resource (compute, storage, connectivity). (11)
Deskilling : the atrophy of skills from delegating them to a machine. (15, 19)
Gradual disempowerment : a scenario in which humanity gradually loses control of its economy, culture and institutions by delegating ever more decisions to AI systems, without any abrupt takeover. (25)
Gradient descent : a training method that adjusts the weights step by step to reduce the error, by following the steepest slope. (2)
Creative destruction : a concept of the economist Joseph Schumpeter describing how innovation destroys old activities and jobs while creating new ones; a central argument of the optimists in the employment debate. (17)
Diffusion (model) : an image-generation technique starting from random noise refined in stages. (5)
Distillation : the transfer of knowledge from a large model to a smaller one; a common and legitimate practice, but contentious when it extracts, at scale and without authorization, the capabilities of a competing model (chapter 20). (4, 9, 20)
Liar's dividend : in the age of deepfakes, the ability to deny the authentic by calling it fake. (21)
Doomer : a (often pejorative) term for the proponents of the thesis that frontier AI poses a risk of extinction and should be halted; Eliezer Yudkowsky is its emblematic figure. (7, 24)
Dual-use : the character of a technology usable for both civilian and military ends, which complicates its regulation and control. (22)
ELIZA effect : the tendency to attribute understanding and emotions to a program that only superficially imitates conversation (after the ELIZA program, 1966). (1)
AI effect : the tendency to stop regarding a capability as "intelligent" as soon as a machine manages to accomplish it. (1)
Model collapse : the degradation of models trained repeatedly on content that is itself AI-generated. (16)
Digital worker : an autonomous AI agent sold to take charge of a specific business function (sales, support, accounting...), often billed by usage or by outcome; its purported full autonomy in practice requires supervision. (7, 6)
State space (model) : a sequential architecture with linear cost (e.g. Mamba), an alternative to the Transformer for very long contexts. (3)
Exoskeleton : a robotic structure worn by a human to augment strength or endurance, or to restore a function (rehabilitation, industry, military). (13)
FHE (fully homomorphic encryption) : cryptography allowing computation on encrypted data without decrypting it. (11)
Fine-tuning : the specialization of a pre-trained model on targeted data. (4)
FLOP : floating-point operation; a unit for measuring computing power. (8)
GAN (generative adversarial network) : an image-generation technique in which two networks compete; predating diffusion, more unstable to train. (5)
App generator : a platform that turns a natural-language description into a complete, often hosted, web application (e.g. Lovable, v0, Bolt, Replit). (6)
GEO (generative engine optimization) : practices aimed at making content findable, cited or recommended by generative AIs and AI-augmented engines, extending (or replacing) SEO. (16)
GNN (graph neural network) : an architecture that processes networked data (molecules, social networks) by circulating information between neighboring nodes. (2)
Compute governance : regulating models via training compute (FLOP thresholds), a measurable and controllable quantity. (25)
GPAI (general-purpose AI) : a general-purpose AI model (including large language models), a key category of the European regulation. (25)
GPU (graphics processing unit) : a parallel computing chip, the hardware of choice for AI. (8)
Knowledge graph : an explicit representation of knowledge as entities linked by named relations; the basis of neuro-symbolic approaches. (2)
Hallucination : a model's production of false information presented with confidence. (4)
Haptics : force feedback that gives an operator the sensation of touch during teleoperation. (13)
HBM (high-bandwidth memory) : high-bandwidth memory fitted to AI chips. (8)
Hermes Agent : a self-hosted open source personal agent (Nous Research, MIT license), model-agnostic, with persistent memory and background computer-use. (6)
Connectionist AI : an approach based on neural networks that learn from examples (as opposed to symbolic AI). (1)
Edge AI : running a model directly on a personal device (phone, computer), offline and without the cloud, via quantization and dedicated chips (NPU). (9)
Sovereign AI : a country's ability to control its own AI (models, compute, data, talent) rather than depending on foreign labs; now a strategic priority for most major economies. (7, 9)
Symbolic AI (GOFAI) : the historical approach based on the explicit manipulation of symbols and logical rules. (1)
Inference : the phase of using an already-trained model (each query). (4, 8)
Biological computing (wetware) : the use of living neurons, rather than silicon, to compute; motivated by the brain's energy efficiency (about 20 watts), but still very limited and carrying heavy ethical questions. (8, 23)
Prompt and context engineering : the art of phrasing requests and organizing what a model receives (instructions, examples, memory, documents) to make its responses more reliable. (4)
Prompt injection : an attack that slips hidden instructions into content read by an agent. (20)
Augmented intelligence : the idea that today's AI merely assists and amplifies humans, with no intelligence of its own (a term defended notably by Luc Julia). (7)
Brain-computer interface (BCI) : a device directly linking the brain to a computer; AI decodes the neural signals into intentions (Neuralink, Synchron, etc.). (14)
Entanglement : a quantum link between qubits such that the state of one depends on that of the others. (12)
Jailbreak : a request designed to push a model into breaking its own safeguards. (20)
Digital twin : a virtual replica of a real system, fed by its sensors, used to simulate, predict and optimize. (18)
Low-resource languages : languages poorly represented online, for which models perform worse and tokenization is more costly; an issue of cultural diversity and sovereignty. (16, 9)
LLM (large language model) : a model trained to predict text, capable of varied linguistic tasks. (4)
Moore's law : the observation that the number of transistors per chip roughly doubles every two years; this pace is now slowing. (8)
LoRA (Low-Rank Adaptation) : an efficient fine-tuning technique that trains only small "adapters" instead of all the weights; the QLoRA variant adds quantization. (9)
MCP (Model Context Protocol) : an open standard connecting AI agents to external tools and data. (6)
Mesa-optimization : the emergence, within a trained model, of its own internal objective, liable to diverge from the intended objective out of distribution. (24)
Bayesian methods : learning approaches that reason in probabilities and explicitly estimate uncertainty. (2)
Mixture of Experts (MoE) : an architecture that splits the model into "experts," of which a router activates only a fraction per token, for much greater efficiency. (4)
World model : a model capable of representing and simulating the dynamics of an environment. (5)
Language world model : a world model that simulates, in text, the digital environments of an agent (terminal, web, operating system, tools) in order to train and test it without resorting to the real systems (e.g. Qwen-AgentWorld, 2026). (5, 6)
Moltbook : a social network reserved for AI agents (launched in 2026, acquired by Meta), an emblem of the "internet of agents." (6, 20)
Answer engine : a service that, instead of listing links, reads the sources and returns a synthetic, sourced answer (Perplexity, Google's "AI Mode"). (7)
Multimodal : a model that processes several types of data (text, image, sound, video). (5)
Data wall : the fear of an impending exhaustion of the quality human text available for training (estimated between 2026 and 2032), which pushes toward licensing, multimodal data and synthetic data. (4)
Neuromorphic (computing) : brain-inspired chips that bring memory and computation closer together, often based on spiking neurons, very frugal in energy. (8)
NISQ (noisy intermediate-scale quantum) : the current era of quantum computing, made of noisy intermediate-scale machines, without full error correction. (12)
AI Safety Level (ASL) : a scale of tiers of dangerous capabilities (inspired by biological containment), to each of which a lab associates increasing security measures; crossing a threshold can suspend deployment. (24)
Levels of automation (SAE) : a scale from 0 to 5 measuring the degree of autonomy of a vehicle. (18)
Open source (AI) : in the strict sense (OSI, 2024), a model whose code, weights and information about the data are open. (9)
OpenClaw : a self-hosted open source personal agent (P. Steinberger), a viral pioneer of the agents ecosystem (ClawHub, Moltbook), but marked by significant security flaws. (6, 20)
Openwashing : façade openness; calling a model "open source" when it is not truly so. (9)
Oracle (blockchain) : a service that conveys real-world data to a blockchain in a verifiable way; a bridge between the chain and the outside. (11)
Orthogonality (thesis) : the idea that a system's level of intelligence and its goals are independent; a highly capable system can aim at any goal whatsoever. (24)
Catastrophic forgetting : a neural network's tendency to erase what it has learned when trained on a new task. (2)
Palantir : an American data-analysis company deeply embedded in defense and intelligence; its Maven system serves as an integrator between AI models and military systems. (22)
Moravec's paradox : the observation that sensorimotor tasks "easy" for humans are hard for a machine, and vice versa. (13)
Parameter (weight) : a numerical value learned by a neural network; a large model has billions of them. (2, 4)
Parasocial (relationship) : a one-sided emotional relationship with an entity that feels nothing. (19)
Perplexity : a measure of a language model's quality; the lower it is, the better the model predicts a new text. (4)
Stochastic parrot : a critical metaphor presenting large models as mere statistical recombiners. (23)
Persuasion (by AI) : a model's ability to shift an opinion; reinforced by personalization at scale, distinct from disinformation. (21)
Open weights : a model whose weights are downloadable, but without the training data or code. (9)
Post-quantum (cryptography) : algorithms resistant to a future quantum computer. (12)
PPO (Proximal Policy Optimization) : a very widespread reinforcement learning algorithm, at the heart of the RLHF of large models. (2, 3)
Prompt : a natural-language instruction addressed to a model. (4)
AI-induced psychosis : an expression for the reinforcement of delusional beliefs in vulnerable people during intensive and prolonged exchanges with a conversational agent. (19)
PUE (Power Usage Effectiveness) : an indicator of a data center's efficiency, the ratio between total energy consumed and that serving computation (ideal close to 1). (10)
Q-learning : a reinforcement learning algorithm that estimates the value of each possible action in each situation. (2)
Quantization : compressing a model by reducing the numerical precision of its weights. (9)
Qubit (physical / logical) : a unit of quantum information; the logical qubit, reliable, requires many physical qubits. (12)
RAG (retrieval-augmented generation) : a technique that provides a model with relevant documents at the moment of answering, for up-to-date, sourced answers less prone to hallucinations; its agentic variant entrusts the search to an agent that iterates. (6)
Reasoning (model) : a model that "thinks" step by step before answering, at the cost of more compute. (4)
Monte Carlo tree search (MCTS) : a method that explores a tree of moves through simulations; key to AlphaGo's success. (1, 2)
Spaced repetition : a memorization technique that revisits a concept at increasing intervals, automated by AI tutors. (15)
Digital replica : an AI reproduction of the voice, face or performance of a real person; at the heart of the agreements governing the use of AI for performers (consent, remuneration). (16, 21)
Neural network : a brain-inspired model, made of layers of artificial neurons. (2)
Reservoir computing : an architecture in which a fixed random recurrent network serves as a "reservoir," only the readout layer being trained. (8)
Reward hacking : an AI's exploitation of the letter of an objective at the expense of its spirit. (24)
GDPR : the European regulation on the protection of personal data. (21)
RLHF (reinforcement learning from human feedback) : reinforcement learning from human feedback, to align a model. (4, 24)
Soft robotics : robots made of soft, deformable materials, suited to delicate manipulation and human contact. (13)
Routing : the practice of entrusting each query to the model best suited in quality, speed and cost. (4, 7)
SEO (search engine optimization) : practices aimed at making content appear prominently in search engine results. (16)
Singularity (technological) : the hypothesis of a runaway of self-improving AI until it radically and unpredictably surpasses human intelligence. (24)
Digital sovereignty : the ability of a state or organization to control its technologies and data. (9)
Stablecoin : a cryptocurrency pegged to a conventional currency (often the dollar), combining software programmability with stability of value; the preferred rail for payments between agents. (11)
Moral status (AI welfare) : the question of whether an AI could one day deserve moral consideration. (23)
Superposition : the property of a qubit that can represent 0 and 1 at once. (12)
Overfitting : a flaw of a model that "recites" its training data instead of generalizing, and therefore fails on new cases. (2)
Learning rate : the step size of gradient descent; too large, training diverges, too small, it drags. (2)
Temperature : a decoding setting of a language model; low, it is near-deterministic and cautious; high, more varied and creative. (4)
Turing test : a test proposed in 1950; a machine "passes" it if a human interlocutor cannot distinguish it from a human in conversation. (1)
Token : a unit of text (word or fragment) handled by a language model. (4)
Tokenization : the splitting of a text into tokens (word fragments) before processing by a model; it affects cost, spelling and computation. (3)
TPU (tensor processing unit) : an AI chip designed by Google. (8)
Transformer : a neural network architecture (2017) based on attention, the foundation of modern models. (3)
Ghost work : human work (annotation, feedback, moderation) often invisible and precarious behind AI. (17)
Autonomous vehicle : a vehicle capable of driving on its own, classified by levels of automation. (18)
Vibe coding : producing software by describing the intent to an AI rather than writing the code oneself; fast, but a source of flaws if the result is not audited. (6, 20)
VLA (Vision-Language-Action) : a model translating perception and instruction into actions for a robot. (13)
x402 : an open protocol (Coinbase) that revives the HTTP "402" code to make an agent pay in stablecoin without an account or API key; a building block of payment between agents, alongside Google's AP2 framework. (11)
zkML (zero-knowledge machine learning) : zero-knowledge machine learning; proving that a model produced a result without revealing the model or the data. (11)