Artificial Atlas

Who is betting on what in AI, and against whom.

GlossaryThe words you hit on the first page, in plain terms.

Each entry says which camps the term matters for. Definitions are short on purpose; the linked paper or essay is where to go next.

Transformer

The network architecture behind every large language model since 2017. It reads all tokens at once and uses attention to decide which ones matter for predicting the next. Cheap to scale, which is why the LLM camp exists.

Read Attention Is All You Need arXiv 2017-06-12

Token

The unit a language model reads and writes, usually a word fragment. "Next-token prediction" means guessing the next fragment given everything before it. Camp 1's entire bet is that this objective, at scale, is enough.

Scaling laws

Empirical curves showing that a model's error falls predictably as parameters, data and compute grow. They let a lab forecast a training run before spending on it, and they are why capital flows to compute.

Pretraining

The first, largest phase of training, where a model learns from a huge unlabelled corpus (text for LLMs, video for world models). Everything after it, from RLHF to fine-tuning, is called post-training.

Post-training

Everything done to a pretrained model to make it useful: instruction tuning, reinforcement learning from human feedback, reinforcement learning on verifiable rewards. Whether the next gains come from here or from bigger pretraining is the main fight inside camp 1.

RLHF

Reinforcement learning from human feedback. People rank model answers, a reward model learns their taste, and the language model is optimized against it. It is what turned raw GPT-3 into an assistant.

RL with verifiable rewards

Reinforcement learning where the reward comes from a checker (a unit test, a maths answer) rather than a human rater. DeepSeek-R1 showed that long reasoning chains emerge from this alone.

Read DeepSeek-R1 arXiv 2025-01-22

Test-time compute

Spending more computation when answering, by reasoning longer or searching over candidates, rather than only when training. The substrate camp did it first for ARC; the LLM camp productised it with the o-series.

Reinforcement learning

Learning by acting and receiving a reward, rather than from labeled examples. Every camp uses some of it; the experience camp says it is the whole thing.

Reward

A scalar signal that tells an agent how well it is doing. "Reward is enough" (2021) argues that maximising one such signal in a rich environment produces every ability we call intelligence.

Read Reward is enough Artificial Intelligence 2021-10-01

Continual learning

Learning new things after deployment without forgetting old ones. Current deep networks lose the ability (loss of plasticity) the longer they train, which is the central unsolved problem of the experience camp.

Plasticity

A network's remaining capacity to learn. A 2024 Nature paper showed standard deep learning loses it under continual training and proposed reinitialising dormant units to recover it.

World model

A model that predicts what happens next in an environment, from video or sensors, so an agent can imagine outcomes before acting. The camp splits on whether to predict pixels or abstract latent features.

JEPA

Joint-Embedding Predictive Architecture, LeCun's design for world models. Encode the input, predict the encoding of the future, never decode back to pixels. AMI Labs raised $1.03B to scale it.

Latent space

The compressed internal representation a model uses instead of raw pixels or words. "Predict in latent space" means predicting the compressed future, on the theory that most pixel detail is noise.

Learned simulator

A neural network trained to stand in for a physics simulation or the real world. The physics camp wants one accurate enough to replace experiments; the robotics camp uses them to manufacture training data.

Neural operator

A network that learns a map between whole functions, such as an initial weather state and the forecast, at any resolution. The Fourier neural operator made them fast, and FourCastNet made them beat physics-based forecasts.

Autonomous lab

Robots that synthesize and test what a model proposes, closing the design loop without a person at the bench. Periodic, Lila and Radical AI are building them.

Vision-language-action model

A vision-language model with an added output head that emits robot actions. RT-2 introduced the idea; pi0 added flow-matching heads for smooth continuous motion. The default architecture of the robotics camp.

Teleoperation

A person controls the robot remotely while its sensors record; the recordings become demonstrations for training. Most humanoid demos you have seen are teleoperated, which is why "teleop-heavy" is a criticism.

Physical AI See also: vla

Cross-embodiment

One policy trained on data from many different robot bodies, so it transfers to a new one. The bet behind Skild's "robot brain" and the Open X-Embodiment dataset.

Physical AI See also: vla
Read Open X-Embodiment arXiv 2023-10-13

Sim-to-real

Training in simulation with randomised appearance and physics so the policy still works on a real robot. Nvidia's Isaac is the standard toolkit.

Program synthesis

Finding a short program that reproduces the examples, instead of a network that approximates them. The program generalizes by construction. Ndea's whole bet, and how the ARC leaders work.

ARC-AGI

Grid puzzles most people solve from two examples, designed so that memorized knowledge does not help. The substrate camp's scoreboard. Best verified 2025 result on ARC-AGI-2 was 54%; ARC-AGI-3 (2026) is interactive and started near 1%.

Read On the Measure of Intelligence arXiv 2019-11-05

Active inference

Friston's theory that brains act to minimize surprise about their inputs, unifying perception and action. Verses builds agents on it.

Read The free-energy principle: a unified brain theory? Nature Reviews Neuroscience 2010-01-13

Thermodynamic computing

Chips that use the physical noise of transistors to sample probabilities directly, instead of simulating randomness with digital arithmetic. Extropic and Normal Computing have taped out first chips.

Mortal computation

Hinton's 2022 proposal for analog hardware whose learned weights are tied to the physical device and die with it, trading copyability for energy efficiency.

Read The Forward-Forward Algorithm arXiv 2022-12-27

Open weights

Publishing a trained model's parameters so anyone can run and fine-tune it. Not the same as open source (training data and code usually stay private). China's labs mostly do it; the American frontier mostly does not.

Valuation (pre-money, post-money)

The price a funding round implies for the whole company. Pre-money excludes the new cash, post-money includes it. This atlas records which one the source reported; "reported" means the figure came from press reports rather than the company.

Seed round

Traditionally the first small check. In this field the word has lost its meaning: AMI Labs' "seed" was $1.03B and Thinking Machines' was $2B.