← all systems · see it in The System
Artificial intelligence
Artificial intelligence (AI) is the capability of computational systems to perform tasks typically associated with human intelligence, such as learning, reasoning, problem-solving, perception, and decision-making; it is also a field of research that develops and studies methods and software enabling machines to perceive their environment and act toward defined goals (src:wp:artificial-intelligence). In the field's agent frame, 'an agent is anything that perceives its environment through sensors and acts upon that environment through actuators' (src:wp:intelligent-agent). No definition of intelligence is agreed (C-1112).
Properties published scienceour proposal
Sourced fields, and our own readings.
| born at | 1956 (Dartmouth workshop; proposal 31 August 1955; Logic Theorist presented there) |
| born precision | year |
| boundary | Inside: AI systems at work (machines that perceive through sensors, learn a model, search and reason, pursue an objective and act through actuators or outputs: deep-learning models, knowledge-based systems, reinforcement-learning agents, robots, automated vehicles, generative models) and the research field that builds, trains, benchmarks and studies them. Outside: the computers they run on (carrier), the people who label, use and fund them, the data on the Web, the electricity they draw, the brains they are compared with. The boundary depends on which definition of AI is used (C-1112). |
| nested level | Runs on computers (rung 12) as programs; built from algorithms, artificial neural networks and knowledge bases; the field sits inside science; deployed AI is embedded in human society and the economy. LLMs and LLM-based agents are kinds, walked on their own. |
| scale | technological |
Measures (22) published science
| characteristic clock: training compute doubling time, deep-learning era | about 6 (5.7 for regular-scale models, 2010-2022) months |
| training compute doubling time before 2010 | about 20 (21.3, 1952-2010) months |
| AlphaZero: random play to superhuman | within 24 hours |
| major AI winters | about 1974-1980 and 1987-2000 years |
| springs-to-winter lag (McDermott's warning to the next winter) | about 5 years |
| the field's founding | 1956 (Dartmouth workshop; proposal dated 31 August 1955) year |
| definitions of intelligence collected | about 70 definitions |
| ImageNet top-5 human error (annotator A1) | 5.1 % |
| DQN games above a human expert | 3 of 7 (and above all earlier approaches on 6 of 7) Atari 2600 games |
| AlphaGo vs Lee Sedol (March 2016) | 4 of 5 games won |
| visuomotor policy size | 92,000 parameters |
| wav2vec 2.0 word error rate with ten minutes of labels | 4.8 / 8.2 (clean / other), after 53k hours unlabelled pre-training % WER |
| CO2e of one NLP model development run with neural architecture search | 626,155 (vs 126,000 for an average car's lifetime incl. fuel) lbs CO2e |
| energy per chatbot query vs web search | about 10 times |
| peak FLOPS vs DRAM vs interconnect bandwidth growth | 3.0x vs 1.6x vs 1.4x per 2 years (over 20 years) |
| speculative decoding speed-up | 2-3 times (T5-XXL, identical outputs) |
| Switch Transformer dropped (overflowed) tokens | typically < 1 % of tokens |
| GPipe bubble negligible when | M >= 4 x K (micro-batches vs partitions) ratio |
| benchmarks mapped | 3765 benchmarks across 947 tasks (computer vision and NLP) |
| model collapse: perplexity after training on generated data (5 epochs, no original data kept) | losing performance from 20 to 28 perplexity points (base model 34 mean perplexity) |
| experts' 50% date for AI outperforming humans in all tasks (2016 survey) | 45 years years from the survey |
| US workforce with >=10% of tasks exposed to LLMs | about 80 (19% with >=50%) |
Open contentions (8) open contention
Where sources or readings disagree. Both stay on the map.
- C-432 Do large language models understand what they say?
booked: The map describes how these models work in the field's own terms (learned features, next-word probabilities) and rules nothing true or false about understanding, belief or meaning. The competing views are listed side by side.
the other reading: see research/CONTENTIONS.md C-432 - C-434 Could large language models be conscious?
booked: The map makes no claim about consciousness in these systems and lists the views side by side.
the other reading: see research/CONTENTIONS.md C-434; the encyclopedia: no settled consensus in philosophy of mind on whether a machine can have a mind or consciousness - C-736 When a system first throws away its excess energy and later learns to store it for reuse, is that one process that grew or two separate ones?
booked: The map keeps both readings side by side and rules neither. A few examples fit the 'one process that grew' reading, but it hasn't been tested.
the other reading: see research/CONTENTIONS.md C-736 - C-1111 Do AI neural networks work like the brain?
booked: The map keeps both: networks that match some brain data, and the crucial ways they differ from human intelligence.
the other reading: performance-optimised networks are quantitative models of the ventral stream (Yamins 2014); an optimisation framework for neuroscience (Richards 2019); they differ in crucial ways: causal models, intuitive physics and psychology, compositionality (Lake 2017); shortcut learning and adversarial brittleness (Mitchell 2021); the brain metaphor for computers is weak outside computer science (C-596, computer walk) - C-1112 What is artificial intelligence?
booked: The map uses the double sense of a capability of computer systems and a field of research, and rules no definition of intelligence.
the other reading: McCarthy: intelligence is 'the computational part of the ability to achieve goals in the world' (via the encyclopedia); Legg & Hutter: a formal universal measure of an agent's ability to achieve goals across environments; 'nobody really knows what intelligence is'; Wang 2019: intelligence is 'adaptation with insufficient knowledge and resources'; Turing: defined by external behaviour (the imitation game); Russell & Norvig object to imitating humans (via the encyclopedia); narrow vs general: general intelligence 'difficult to define and difficult to measure' - C-1113 Will intelligence be built from symbols and rules, or learned by neural networks?
booked: The map keeps both approaches as kinds of AI and records the 2010s shift to deep learning as an event, not a verdict.
the other reading: Newell & Simon: 'A physical symbol system has the necessary and sufficient means of general intelligent action'; connectionism / deep learning: intelligence from learned weights; the rise of deep learning 'may represent a shift toward the scruffies' (encyclopedia); embodied AI (Brooks, Moravec): no representation; Mitchell's fallacy 4 'intelligence is all in the brain'; hybrid: combine neural networks with structured cognitive models (Lake et al. 2017) - C-1115 When will AI match human abilities?
booked: The map records forecasts as dated statements by their authors, not as measurements.
the other reading: 2016 expert survey: 50% chance of AI outperforming humans in all tasks in 45 years (Grace 2018); surveys of 2016 and 2018: 50% chance of human-level AI by 2040-2060 (as Mitchell reports); Mitchell 2021: predictions reflect 'our own biases and lack of understanding of the nature of intelligence'; springs and winters likely to continue - C-1116 Do AI systems build models of the world?
booked: The map marks world-modelling as possible: some systems hold stored models, while others act without one. The view that machines should build causal models is kept alongside.
the other reading: machines should 'build causal models of the world that support explanation and understanding, rather than merely solving pattern recognition problems' (Lake 2017); end-to-end policies map perception to control without a separate world model (Levine 2016)
Partners and parts published science
Built from / builds (11)
- Artificial neural networks (trained sequence-to-sequence models)
- Machine learning
- Knowledge bases and logic
- Machine perception (computer vision, speech recognition)
- Sensors
- Actuators
- Objective function (utility, reward, loss)
- Benchmarks
- The AI research community
- Human champions vs game-playing programs
- GPUs against neuromorphic hardware for AI computing
Sits inside / holds (3)
- Science (as a human social and cultural system)
- Human society (from local community up to societies)
- Economy
Trades with (valence) (26)
- Photography via sys:sensory-environment · Artificial intelligence gain
- Computer-aided design via sys:written-record · Artificial intelligence gain
- Signals intelligence via sys:intelligence-information · Artificial intelligence gain
- Human (Homo sapiens) via sys:sensory-environment · Artificial intelligence gain
- Science (as a human social and cultural system) via sys:written-record · Artificial intelligence gain
- Software via sys:written-record · Artificial intelligence gain
- Internet via sys:written-record · Artificial intelligence mixed
- Large language model via sys:written-record · Artificial intelligence gain
- LLM-based agent via sys:written-record · Artificial intelligence gain
- Human brain via sys:written-record · Artificial intelligence gain
- Computer via sys:sensory-environment · Artificial intelligence gain
- Algorithm (the practice of step-by-step procedures and the structure they compute) via sys:written-record · Artificial intelligence gain
- Electricity via sys:electromagnetic-field · Artificial intelligence gain
- Economy via sys:human-society · Artificial intelligence gain
- Computer security via sys:written-record · Artificial intelligence mixed
- Argument: premises offered in support of a conclusion, and argumentation, the practice of giving and asking for reasons in doubt or disagreement, as philosophy and argumentation theory describe them via sys:language · Artificial intelligence gain
- Compiler via sys:written-record · Artificial intelligence gain
- Computer ethics via sys:computer · Artificial intelligence mixed
- Museum via sys:written-record · Artificial intelligence even
- Computer science (the field: the practice and the computation it studies) via sys:sensory-environment · Artificial intelligence gain
- Mathematical proof: the practice of showing that a statement follows from axioms and earlier theorems, how it is written, checked (by people and machines) and accepted, and the derivations it works on via sys:written-record · Artificial intelligence gain
- Epistemology (the theory of knowledge: the study and the knowing it studies) via sys:sensory-environment · Artificial intelligence gain
- Metaphysics (the branch of philosophy: its practice and the questions about reality it works on) via sys:sensory-environment · Artificial intelligence gain
- Corporate law via sys:artificial-intelligence · Artificial intelligence even
- Mind–body problem (how mental states relate to physical states) · Artificial intelligence mixed
- Intelligence via sys:written-record · Artificial intelligence gain
expressed by (14)
- Generation/Synthesis
- Templating / Standardization
- Quality Improvement/Refinement
- Interoperability/Transduction (Bridging)
- Sensing/Scanning
- Search/Foraging
- Uncertainty Reduction/Learning
- Active Experimentation / Probing (Optimal Experimental Design)
- Multi-Scale Data Fusion / Triangulation
- Offline Simulation & Replay-Guided Planning
- Social Learning / Public Information Use
- Validation of Effectiveness
- Redundancy & Diversity
- Quality Control & Error Correction
floor filled by (24)
- Signals intelligence
- Human (Homo sapiens)
- Internet
- Large language model
- LLM-based agent
- The world through the senses (sensory environment)
- Computer
- Algorithm (the practice of step-by-step procedures and the structure they compute)
- Electricity
- Artificial neural networks (trained sequence-to-sequence models)
- Knowledge bases and logic
- Machine perception (computer vision, speech recognition)
- Sensors
- Actuators
- Objective function (utility, reward, loss)
- Benchmarks
- The AI research community
- Graphics processing unit
- Convolutional neural network
- Monte Carlo tree search
- Research funding
- Compiler
- Epistemology (the theory of knowledge: the study and the knowing it studies)
- Mind–body problem (how mental states relate to physical states)
represents (2)
Kinds (12)
- Large language model
- LLM-based agent
- Computer science (the field: the practice and the computation it studies)
- Intelligence
- Symbolic AI (GOFAI)
- Expert system
- Deep learning systems
- Reinforcement-learning agent
- Generative adversarial network
- Automated driving system
- Robot (learning robot)
- Artificial general intelligence
Notes published science
- The same mathematical step that teaches a network can be turned around to fool it: training walks the error downhill by changing the weights, while an adversarial example walks the same error uphill by changing a few pixels nobody can see, and the picture is then misread (src:szegedy-2013). In the one defence study read, 6 of 7 defences built against this were beaten completely (src:athalye-2018).
- AI's 'carry' shows all three faces at once: in a sparse model each expert can take only so many tokens, and the overflow is passed straight on to the next layer (src:fedus-2021-switch); a language model must wait one full run per word (src:leviathan-2022); and speculative decoding is the lookahead, letting a small model guess several words ahead so the big one can check them together, 2-3 times faster with identical output.
- Feeding models their own output pays an entropy toll that compounds: each generation trained on the last one's text or images loses the rare cases first ('tails of the original content distribution disappear'), irreversibly (src:shumailov-2024). The field itself runs on a slower swing: springs of optimism and money, then winters, about 1974-1980 and 1987-2000 (src:mitchell-2021, src:wp:ai-winter).
Layers
The ways it takes in its world, where they meet, and how it acts back, each from L1 (where it meets the world) up.
percepts in, learned model, actions out (reader's view)
- L1: The environment the agent's sensors touch: light, sound, text, objects and surfaces, its own body, rewards from the environment
- L2: Sensors: cameras, microphones, lidar, radar, ultrasonic and tactile sensors, joint encoders; the reward channel; labelled data
- L3: Perception: deducing aspects of the world from sensor input (computer vision, speech recognition)
- L4: The learned model or knowledge base: weights of a neural network, or rules and facts
- L5: Inference: search, reasoning, planning
- L6: The objective: utility, reward or loss; goals held in an objective function, set by design
- L7: Action back on the world: the action expected to maximise the performance measure, taken through actuators or handed out as content and decisions
the boardroom (shared floors)
- L4: The learned model or knowledge: the weights of an artificial neural network, where meaning sits in the space of high-layer activations, or a knowledge base of rules and facts a program uses
- L5: Inference: search through states, reasoning, planning; step-by-step methods meet a combinatorial explosion; sparse models route each token to experts
- L6: The objective: an objective function, utility or reward that holds the goals, set by design together with the learning rule and architecture; the field's benchmarks steer which objectives are chosen
- L7: The action selected: the one expected to maximise the performance measure, given percepts, prior knowledge and available actions; booked in the field's mechanism words (C-432: nothing about understanding or meaning ruled; C-434: nothing about consciousness)
the field: benchmarks, results, recognition and funding (the research community)
- L1: Benchmarks: datasets and metrics
- L2: Benchmarks define tasks and steer progress
- L3: Results reported and recognised; paper acceptance
- L4: Funding and investment: springs and winters
- L5: The research agenda: the traditional goals (learning, reasoning, knowledge representation, planning, language, perception, robotics)
way in: Benchmark results (the field's channel)
- L1: Benchmarks: datasets and metrics that define tasks
- L2: Results reported on benchmarks; state-of-the-art (SOTA) curves
- L3: Recognition: models with new SOTA results receive widespread recognition
- L4: Scientific reward systems: paper acceptance tied to SOTA metrics
- L5: Enthusiasm, funding and jobs follow apparent breakthroughs, and dry up when progress stalls
way in: Compute and energy (electricity into accelerators)
- L1: Electricity: hardware, electricity and cloud compute paid for
- L2: Graphics processing units and other accelerators
- L3: Training compute (operations) spent on a model; memory bandwidth now the main bottleneck
way in: Hearing (microphones, speech audio)
- L1: Sound in air: speech and other audio
- L2: Microphone: the speech waveform
- L3: Latent speech representations learned from audio, masked and quantised
way in: Labels and human judgements (supervision)
- L1: People's answers: expert annotators and crowds labelling examples
- L2: Labelled training data: each example paired with its expected answer
- L3: The error between output and label, passed back along credit-assignment paths
way in: Own experience: games played against itself (inner channel)
- L1: Games the system played against itself, from random play
- L2: Self-play games recorded as experience for learning (tabula rasa reinforcement learning)
- L3: Learning by running experiments on itself
way in: Proprioception and touch (the machine's own body: joint encoders, vehicle state, tactile sensors)
- L1: The machine's own body state: joint angles, end-effector pose, the vehicle's own motion
- L2: Joint encoders; proprioceptive sensors for internal vehicle state; tactile sensors
- L3: Encoder readings fed with image features into the policy network
way in: Range sensing (lidar, radar, ultrasonic)
- L1: The environment's objects: dynamic and static, drivable areas, buildings, pedestrian crossings
- L2: Active sensors: lidar, radar, ultrasonic (they emit and read back)
- L3: Perception: raw sensor data turned into a description of the world (machine perception)
way in: Reward (how good the situation is)
- L1: The environment: what causes reward (a game won, a puzzle solved); the reward is part of the perception the environment generates
- L2: The reward signal: the simplest possible semantics
- L3: A value function estimating future rewards (Q-learning)
way in: Text (written language)
- L1: Written human language
- L3: Word embeddings: words as vectors encoding their meaning; transformers with attention
way in: Training data (the slow channel)
- L1: Huge datasets scraped from the Web
- L2: Curated datasets and benchmarks: millions of images in hundreds of categories
- L3: Learning from a stream of data: finding patterns without other guidance (unsupervised), or with labels
way in: Vision (cameras, images)
- L1: Light from the scene: colour and shape of objects, read passively
- L2: Camera: an image of raw pixels
- L3: Convolutional layers: a hierarchy of learned feature maps; intermediate layers track V4, the top layer IT in the primate ventral stream
way out: Classifications, rankings and decisions handed to people
- L1: Search results, assistants' answers and decisions people rely on; trust needs an explanation
- L7: The class predicted, the result ranked, the choice of maximum expected utility
way out: Generated content (text, images, audio, video)
- L1: Text, images, audio and video put into the world (and onto the Web)
- L7: A generator samples from the learned data distribution
way out: Heat and carbon (energy spent)
- L1: Carbon emitted to fuel the hardware; heat to the surroundings
- L4: Training a model consumes energy in proportion to its compute
way out: Motor commands (robots and vehicles)
- L1: Objects moved: screwing a cap onto a bottle; a vehicle driven
- L2: Torques at the joints; actuators
- L7: The policy picks the action
way out: Moves in a game or simulated world
- L1: Moves on the board or the game screen: a world-champion program defeated
- L7: A move chosen from the learned value of the positions
way out: The field's output: new benchmarks, results and predictions
- L1: Papers, benchmarks and confident predictions into the record
- L6: Novel benchmarks created to complement or replace saturated ones