Axes and scoring
Five axes, fifty statements, one distance. Everything here is in the repository and can be changed by pull request.
Overview
The chart has five axes. Representation (x) and architecture (y) place you on the 13 × 12 grid, paradigm picks the shade of the cell, and openness and speed decide the titles on the ring. The quick test's 10 statements measure the two main axes; the full test's 50 measure all five.
| Axis | −1 | +1 | Statements | Used for |
|---|---|---|---|---|
| x · Representation | Language | World | 48 | place on the chart |
| y · Architecture | Harness | Monolith | 57 | place on the chart |
| p · Paradigm | Mainstream | Paradigm shift | 41 | shade of the cell |
| o · Openness | Frontier | Open | 16 | titles on the ring |
| a · Speed | Caution | Acceleration | 18 | titles on the ring |
Representation x
What is intelligence built on? On the Language side, language is the medium of thought and text pretraining is where knowledge comes from; other modalities are senses that feed a linguistic core. On the World side, intelligence is grounded in perception, video, physics and a body; language is a late, lossy label on what the senses already know.
| ID | Statement | Agreeing moves to |
|---|---|---|
x1 quick | Language is the vehicle of thought; almost anything a mind can think, it could in principle work through in words. | Language |
x2 quick | A model that has only read text will never truly know what “heavy” or “slippery” means. | World |
x3 | Internet text already holds most of human knowledge. | Language |
x4 | A four-year-old has seen more data than any language model has read, and that matters. | World |
x5 | The main job of a vision module is to translate images for the language model. | Language |
x6 quick | The next generation of foundation models should train mainly on video, not text. | World |
x7 | Reasoning is fundamentally linguistic: the hard part of solving a problem happens in words and symbols, not in wordless intuition. | Language |
x8 | A system that cannot predict what the physical world does next is not intelligent. | World |
x9 quick | Robots and embodied data are a necessary road to general intelligence. | World |
x10 quick | Most progress in multimodal ability comes from a stronger language backbone. | Language |
x11 | Language is a low-bandwidth, lossy compression of the world, so models trained mainly on text will hit a hard capability ceiling. | World |
x12 | A model that writes good code and proves theorems is closer to general intelligence than one that understands video. | Language |
x13 | Once high-quality web text runs out, the next big source of new knowledge must be sensory data such as video, not synthetic text. | World |
x14 | Model-generated, verifier-checked synthetic text in math, code and reasoning traces can keep feeding pretraining well past the human-text data wall. | Language |
x15 | Most of the physics a video model learns is already written down somewhere in text; video adds perceptual detail more than new knowledge. | Language |
x16 | Interaction data, where a model observes the consequences of its own actions, teaches causal knowledge that passively scraped text cannot. | World |
x17 | For general reasoning ability, a compute budget spent on code and formal mathematics data buys more than the same budget spent on video. | Language |
x18 | For long-horizon control, planning in a learned latent world model, as in Dreamer or MuZero, scales better than planning in natural-language steps. | World |
x19 | Hard geometry and spatial puzzles are best solved by writing code or symbolic descriptions, not by reasoning over pixels or latent images. | Language |
x20 | Reasoning in continuous hidden states, without emitting chain-of-thought tokens, will beat explicit token-level chains of thought at equal compute. | World |
x21 | Visual chain-of-thought, generating intermediate images or sketches while reasoning, adds little over a good textual chain of thought on multimodal tasks. | Language |
x22 | Predicting in pixel or latent space, as JEPA-style or Genie-style world models do, is a better substrate for intuitive physics than text tokens. | World |
x23 | Object common sense, such as that cups spill and cloth folds, is learned mainly by acting on objects, not by watching or reading about them. | World |
x24 | A disembodied model could solve open problems in mathematics and science without ever controlling a body or a sensor. | Language |
x25 | Most of the reasoning a household robot needs can be done by a language model that has never controlled a body. | Language |
x26 | Any serious benchmark for general intelligence should include physical tasks in the real world, not only text, code and puzzles like ARC. | World |
x27 | Learning to act in physically rich simulators such as MuJoCo is needed for general intelligence, not only for robotics; text pretraining cannot substitute for it. | World |
x28 | Digital agents that operate browsers, terminals and codebases are a more direct path to general intelligence than physical robots. | Language |
x29 | Touch and proprioception matter as much as vision for a general world model, and today's multimodal models badly neglect them. | World |
x30 | Embodied AI can wait: once disembodied models are capable enough, controlling robots will follow cheaply as a downstream application. | Language |
d1 | A capable household robot is best built as a modular stack of perception, mapping, planning and control, not one end-to-end policy. | World |
d2 | A large enough language model can serve as a general agent end to end, with screens and sensor readings simply encoded as tokens. | Language |
d3 | One action-conditioned video world model, trained end to end, will replace hand-built simulators and planners in robotics. | World |
d4 | Most of what an AI scientist needs is a language model that calls literature search, code execution and symbolic solvers as tools. | Language |
d5 | Robots must keep learning from their own physical interactions after deployment; no fixed pretraining corpus can anticipate the real world. | World |
d6 | More text, code and synthetic reasoning traces under the same next-token objective will keep producing better reasoners. | Language |
d7 | Video models trained with plain next-frame prediction will learn intuitive physics at scale; no new learning objective is needed. | World |
k2 | Computer-use agents that see raw screenshots and act through mouse and keyboard will eventually beat agents given structured APIs and accessibility trees. | World |
k3 | A world model should predict future states in a learned latent space, not generate future video frames pixel by pixel. | World |
k4 | Turning images into detailed captions and letting a text-only model reason over them loses little on most real-world tasks. | Language |
k5 | General-purpose robot manipulation mainly needs far more real robot data under today's Transformer recipe, not new learning algorithms. | World |
k14 | Vision-language-action models should emit actions as discrete tokens in the language model's vocabulary, not through a separate continuous action head. | Language |
k18 | Curated knowledge graphs still answer factual questions more reliably than any language model's parametric memory, and will keep doing so. | Language |
k21 | Spatial reasoning in current models will stay brittle until they are trained on data from agents moving through and acting in 3D environments. | World |
k22 | Most of what a multimodal model knows about images comes from their paired captions, so caption quality matters more than image volume. | Language |
k27 | Letting the model decide how long to think in its own chain of thought beats imposing external search like tree-of-thought over its outputs. | Language |
k29 | Physical reasoning needs explicit object-centric slot representations; dense patch embeddings will not get there by scaling alone. | World |
k30 | Next-token prediction on enough text already yields a usable world model inside the network; no separate world-model objective is needed. | Language |
Architecture y
Where does intelligence live? The Monolith side bets on one end-to-end model; memory, planning and tool use should emerge inside it, and today's scaffolding will be absorbed by the next generation. The Harness side bets on the system around the model — tools, memory, workflows, explicit structure, many agents — and holds that the harness can matter more than the model.
| ID | Statement | Agreeing moves to |
|---|---|---|
y1 quick | Today’s agent frameworks and harnesses will be absorbed by stronger models within a few years. | Monolith |
y2 quick | The same model with a better harness can gain more than a whole model generation. | Harness |
y3 quick | End-to-end training will always beat hand-designed modular pipelines. | Monolith |
y4 quick | Intelligence lives in the system (memory, tools, workflow), not just in the weights. | Harness |
y5 quick | Tool use is a stopgap; models should eventually do everything internally. | Monolith |
y6 | Several specialized models working together are more reliable than one general model. | Harness |
y7 | Making one general model stronger is a better use of research effort than engineering systems that combine several models. | Monolith |
y8 | Explicit, readable, editable structure (behavior trees, programs, rules) is the key to reliable agents. | Harness |
y9 | Memory, planning and reflection should emerge inside the model, not be bolted on. | Monolith |
y10 | Context engineering — what the model gets to see — matters more than a bigger model. | Harness |
y11 | Multi-agent systems are a temporary patch for weak single models; better models will make most of them unnecessary. | Monolith |
y12 | An agent should distill what it learns into code or rules instead of calling the model every time. | Harness |
y13 | Routing each request to the cheapest adequate model in a pool will keep beating sending everything to one frontier model on cost-adjusted quality. | Harness |
y14 | Natively multimodal models trained jointly on all modalities will outperform systems that bolt a vision encoder or speech module onto an LLM. | Monolith |
y15 | Pairing a generator with a separately trained verifier or reward model at inference time beats spending the same compute on a bigger generator. | Harness |
y16 | Mixture-of-experts inside one network is the right way to specialise; separately deployed expert models are an engineering detour. | Monolith |
y17 | Keeping ASR, OCR and TTS as separate components around an LLM is the better design, because each can be fixed without retraining everything. | Harness |
y18 | One model trained to handle chat, coding, search and computer use beats maintaining a separate fine-tuned model for each product. | Monolith |
y19 | On hard agentic tasks such as SWE-bench, combining several distinct models is a better investment than squeezing more performance out of any single model. | Harness |
y20 | With million-token context windows, most RAG pipelines become unnecessary; putting the documents straight into context is simpler and works better. | Monolith |
y21 | Long-term agent memory should live in an external, inspectable store such as files or a database, not be written into the weights. | Harness |
y22 | Domain knowledge belongs in the weights through continued pretraining or fine-tuning, rather than being retrieved from a document store at inference time. | Monolith |
y23 | A mid-sized model with a calculator, code interpreter and web search beats a much larger model without tools on most real-world tasks. | Harness |
y24 | Most hallucinations are better fixed by grounding answers in retrieved sources with citations than by further training the model. | Harness |
y25 | External memory files and skill libraries add more failure modes than they remove; an agent's working state belongs in context and its skills in the weights. | Monolith |
y26 | End-to-end reinforcement learning on agentic tasks will produce better planners than any hand-written workflow or prompt chain. | Monolith |
y27 | For production agents, a fixed workflow with LLM calls at chosen steps is more reliable than letting the model pick its next action freely. | Harness |
y28 | Detailed step-by-step scaffolds and long system prompts increasingly hurt frontier models, which plan better when given only the goal. | Monolith |
y29 | An explicit state machine or orchestration graph, not the model itself, should decide when an agent may call tools, retry or stop. | Harness |
y30 | Task decomposition should be learned by the model, not specified by developers; hand-designed subtask graphs cap what an agent can achieve. | Monolith |
d1 | A capable household robot is best built as a modular stack of perception, mapping, planning and control, not one end-to-end policy. | Harness |
d2 | A large enough language model can serve as a general agent end to end, with screens and sensor readings simply encoded as tokens. | Monolith |
d3 | One action-conditioned video world model, trained end to end, will replace hand-built simulators and planners in robotics. | Monolith |
d4 | Most of what an AI scientist needs is a language model that calls literature search, code execution and symbolic solvers as tools. | Harness |
d8 | Long-term memory should live in weights that update during inference, replacing external vector databases and retrieval pipelines. | Monolith |
d9 | Reliable reasoning will come from coupling neural networks with explicit symbolic solvers and program search, not from scaling a single network. | Harness |
d10 | A plain decoder-only Transformer, scaled up, will learn retrieval and planning internally without any new architectural components. | Monolith |
k1 | A vision-language model should fully retrain its vision encoder on its own data, rather than keep a frozen contrastive encoder and train only the projector. | Monolith |
k2 | Computer-use agents that see raw screenshots and act through mouse and keyboard will eventually beat agents given structured APIs and accessibility trees. | Monolith |
k4 | Turning images into detailed captions and letting a text-only model reason over them loses little on most real-world tasks. | Harness |
k7 | Models should learn to compress and manage their own long-running context, instead of relying on hand-written summarization and memory-management code. | Monolith |
k9 | A model that keeps updating its own weights from each user's interactions is worth the cost in reproducibility and auditability. | Monolith |
k10 | Agents should keep what they learn on the job as editable text notes and skill files rather than fine-tuning it into weights. | Harness |
k12 | Using separate visual tokenizers for image understanding and image generation is a mistake; one shared visual representation should serve both. | Monolith |
k13 | Classical motion planning and model-predictive control should stay in the loop underneath any learned robot policy that is actually deployed. | Harness |
k14 | Vision-language-action models should emit actions as discrete tokens in the language model's vocabulary, not through a separate continuous action head. | Monolith |
k15 | One policy trained on data from many different robot bodies will beat policies trained for a single specific robot, even on that robot. | Monolith |
k18 | Curated knowledge graphs still answer factual questions more reliably than any language model's parametric memory, and will keep doing so. | Harness |
k19 | Multi-agent debate improves answers mainly by sampling more; a single model with majority voting at equal compute matches it. | Monolith |
k20 | A world model written as executable code, such as a physics program or game engine, generalizes better than a neural video model. | Harness |
k24 | For coding agents, improving the test-and-execution scaffolding around the model now matters more for reliability than improving the model itself. | Harness |
k25 | Separate detection, tracking, mapping and planning modules remain the right foundation for deployed self-driving, rather than one end-to-end driving network. | Harness |
k26 | Speech should be modeled end to end as audio tokens; cascading ASR, a text model and TTS throws away what matters. | Monolith |
k27 | Letting the model decide how long to think in its own chain of thought beats imposing external search like tree-of-thought over its outputs. | Monolith |
k28 | Small task-specific models fine-tuned with standard methods will match a frontier general model on most narrow tasks a team actually needs. | Harness |
k29 | Physical reasoning needs explicit object-centric slot representations; dense patch embeddings will not get there by scaling alone. | Harness |
k30 | Next-token prediction on enough text already yields a usable world model inside the network; no separate world-model objective is needed. | Monolith |
Paradigm p
Is today's mainstream recipe the road? The Mainstream side holds that the Transformer, pretraining on next-token or next-pixel prediction, post-training and scale will keep winning, improved by engineering. The Paradigm-shift side holds that a different paradigm is needed, whether a new idea or an older one the mainstream left behind, such as symbolic methods or evolution. On the chart this axis is the shade of each cell.
| ID | Statement | Agreeing moves to |
|---|---|---|
p1 | The Transformer will be replaced as the default architecture for frontier models within five years. | Paradigm shift |
p2 | The current recipe — pretraining, post-training, scale — is enough to reach general intelligence. | Mainstream |
p3 | The next breakthrough is more likely a small group’s new idea than a bigger cluster. | Paradigm shift |
p4 | Backpropagation will eventually be replaced; future frontier systems will learn with a fundamentally different algorithm. | Paradigm shift |
p5 | Most “new architecture” papers end up losing to a scaled-up Transformer. | Mainstream |
p6 | Continual learning after deployment is what current models lack most. | Paradigm shift |
p7 | Today's winning recipe is mostly decades-old neural-network ideas finally given enough data and compute; the hunt for a new paradigm is overrated. | Mainstream |
p8 | We need new learning objectives, not next-token or next-pixel prediction. | Paradigm shift |
p9 | Incremental engineering moves the field more than paradigm shifts do. | Mainstream |
p10 | In ten years, today’s mainstream methods will look as dated as SVMs look today. | Paradigm shift |
p11 | State-space models or other recurrent designs will replace attention as the default architecture for long-context modeling. | Paradigm shift |
p12 | Hybrid layers, mixture-of-experts and new attention variants are refinements; the core Transformer block will remain the backbone for at least a decade. | Mainstream |
p13 | ConvNets and LSTMs trained with modern recipes, data and compute close most of their gap to Transformers; scale matters far more than architecture. | Mainstream |
p14 | Predicting in a learned latent space, as joint-embedding methods do, is a better pretraining signal than reconstructing tokens or pixels. | Paradigm shift |
p15 | Reinforcement learning from environment reward, not imitation of human data, should become the main source of a model's capabilities. | Paradigm shift |
p16 | Next-token prediction on diverse enough data already forces a model to learn whatever world model the task requires. | Mainstream |
p17 | Test-time training, where weights adapt to each new input or task, will become a standard part of inference. | Paradigm shift |
p18 | Retrieval over an external memory gives deployed models most of what continual learning promises, without touching the weights. | Mainstream |
p19 | A deployed model should be a fixed, versioned checkpoint; weights that drift with user interactions cannot be properly evaluated. | Mainstream |
p20 | Catastrophic forgetting is still unsolved at scale, so periodic retraining on refreshed data beats online weight updates. | Mainstream |
d5 | Robots must keep learning from their own physical interactions after deployment; no fixed pretraining corpus can anticipate the real world. | Paradigm shift |
d6 | More text, code and synthetic reasoning traces under the same next-token objective will keep producing better reasoners. | Mainstream |
d7 | Video models trained with plain next-frame prediction will learn intuitive physics at scale; no new learning objective is needed. | Mainstream |
d8 | Long-term memory should live in weights that update during inference, replacing external vector databases and retrieval pipelines. | Paradigm shift |
d9 | Reliable reasoning will come from coupling neural networks with explicit symbolic solvers and program search, not from scaling a single network. | Paradigm shift |
d10 | A plain decoder-only Transformer, scaled up, will learn retrieval and planning internally without any new architectural components. | Mainstream |
k3 | A world model should predict future states in a learned latent space, not generate future video frames pixel by pixel. | Paradigm shift |
k5 | General-purpose robot manipulation mainly needs far more real robot data under today's Transformer recipe, not new learning algorithms. | Mainstream |
k8 | RL-trained chains of thought have opened a scaling axis within the current recipe that will keep delivering large gains for years. | Mainstream |
k9 | A model that keeps updating its own weights from each user's interactions is worth the cost in reproducibility and auditability. | Paradigm shift |
k10 | Agents should keep what they learn on the job as editable text notes and skill files rather than fine-tuning it into weights. | Mainstream |
k11 | Within five years, diffusion rather than left-to-right autoregression will be the main way frontier models generate text. | Paradigm shift |
k13 | Classical motion planning and model-predictive control should stay in the loop underneath any learned robot policy that is actually deployed. | Paradigm shift |
k14 | Vision-language-action models should emit actions as discrete tokens in the language model's vocabulary, not through a separate continuous action head. | Mainstream |
k16 | Tokenizer-free models that read raw bytes will replace subword tokenization in frontier language models. | Paradigm shift |
k17 | Agents that rewrite their own scaffolding code and prompts from experience will improve faster than any harness designed by humans. | Paradigm shift |
k18 | Curated knowledge graphs still answer factual questions more reliably than any language model's parametric memory, and will keep doing so. | Paradigm shift |
k20 | A world model written as executable code, such as a physics program or game engine, generalizes better than a neural video model. | Paradigm shift |
k23 | Claims that scaling has hit a wall keep being proven wrong; scaling the current recipe will continue to unlock new capabilities. | Mainstream |
k29 | Physical reasoning needs explicit object-centric slot representations; dense patch embeddings will not get there by scaling alone. | Paradigm shift |
k30 | Next-token prediction on enough text already yields a usable world model inside the network; no separate world-model objective is needed. | Mainstream |
Openness o
Who should hold the strongest models? The Open side wants open weights, open data and reproducible science. The Frontier side prefers capability concentrated in a few labs that ship it through products and APIs. It earns titles on the ring rather than a place on the chart.
| ID | Statement | Agreeing moves to |
|---|---|---|
o1 | The weights of the strongest models should be open. | Open |
o2 | Frontier capability is better concentrated in a few well-resourced developers than spread across many open-weight releases. | Frontier |
o3 | Results on closed models that no one can reproduce should not count as scientific contributions. | Open |
o4 | Publicly funded AI research should be required to release its models, training data and code. | Open |
o5 | Releasing full training data and recipes is too risky. | Frontier |
o6 | Being able to run, inspect and modify a model locally is worth accepting a noticeable capability gap to the best closed model. | Open |
o7 | A commercial API is the best way to put AI in everyone’s hands. | Frontier |
o8 | Training data should be fully public and auditable. | Open |
o9 | Public institutions should only adopt AI models whose weights they can download, inspect and host on their own infrastructure. | Open |
o10 | Once a model has been superseded by its developer's newer release, its weights should be made public. | Open |
o11 | Open-weight releases should stop above a capability threshold, since safety tuning can be fine-tuned away in a few hours. | Frontier |
o12 | For most developers, a well-priced API with fine-tuning endpoints makes direct access to the weights largely unimportant. | Frontier |
o13 | A system card with thorough evaluations is sufficient transparency for a frontier model; its architecture and training code can stay private. | Frontier |
o14 | AI developers are justified in withholding key training innovations from publication for a year or more to protect their lead. | Frontier |
o15 | Developers who pay for data curation have a legitimate right to keep their pretraining data mixture a trade secret. | Frontier |
k28 | Small task-specific models fine-tuned with standard methods will match a frontier general model on most narrow tasks a team actually needs. | Open |
Speed a
How fast should we go? The Acceleration side wants progress as fast as possible and sees regulation and human sign-off as bottlenecks. The Caution side takes catastrophic risk seriously and wants alignment, interpretability and independent testing to keep pace. It earns titles on the ring rather than a place on the chart.
| ID | Statement | Agreeing moves to |
|---|---|---|
a1 | The faster AI progresses, the better. | Acceleration |
a2 | The probability that advanced AI causes a global catastrophe within this century is above ten percent. | Caution |
a3 | Regulation does AI more harm than good. | Acceleration |
a4 | Frontier training should slow down until alignment is solved. | Caution |
a5 | The more autonomy agents get the better; human approval is the bottleneck. | Acceleration |
a6 | Interpretability research should get more resources than capability research. | Caution |
a7 | AI doom talk is mostly marketing or panic. | Acceleration |
a8 | Independent bodies should safety-test frontier models before deployment, even if this delays releases by months. | Caution |
a9 | The cost of delaying AI applications in medicine and science outweighs the risk of deploying them sooner. | Acceleration |
a10 | Each frontier release teaches more about safety than years of analysis beforehand, so releases should not be held back for theoretical risks. | Acceleration |
a11 | Developers should commit in advance to pausing training when evaluations reveal dangerous capabilities that current safeguards cannot contain. | Caution |
a12 | An international cap on training compute for the largest runs would be worth the slower progress it causes. | Caution |
a13 | Coding agents should be allowed to merge and deploy their own changes once tests pass, without a human review step. | Acceleration |
a14 | Mandatory pre-deployment licensing of frontier models would mainly entrench incumbents rather than make models safer. | Acceleration |
a15 | An agent that can spend money or send messages for a user should get explicit confirmation before every such action, even small routine ones, despite the friction. | Caution |
k6 | Developers should accept some loss of accuracy to keep reasoning traces human-readable, rather than let RL drift them into unreadable sequences. | Caution |
k9 | A model that keeps updating its own weights from each user's interactions is worth the cost in reproducibility and auditability. | Acceleration |
k17 | Agents that rewrite their own scaffolding code and prompts from experience will improve faster than any harness designed by humans. | Acceleration |
Scoring
Each statement has five answers: strongly agree +2, agree +1, unsure (not counted), disagree −1, strongly disagree −2. The answer is multiplied by the statement’s weight on each axis (which way agreeing moves you), summed per axis, and divided by the most the answered statements on that axis could score, giving a number between −1 and +1. An “unsure” neither moves the score nor pulls it toward the middle.
score[axis] = Σ answer × weight[axis] / Σ 2 × |weight[axis]| (sums over statements not answered “unsure”)
Paradigm levels and their p score:
| Level | Name | p |
|---|---|---|
| 1 | Beyond Revolutionary | 1.00 |
| 2 | Ultra-Revolutionary | 0.75 |
| 3 | Revolutionary | 0.50 |
| 4 | Exploratory | 0.25 |
| 5 | Neutral | 0.00 |
| 6 | Steady | -0.25 |
| 7 | Mainstream | -0.50 |
| 8 | Orthodox | -0.75 |
| 9 | Ultra-Orthodox | -1.00 |
Matching
Your ideology is the cell nearest to you, measured in cells: one column is 2/13 of x, one row 2/12 of y, and one paradigm level of mismatch costs about 0.7 of a cell. The centre block is drawn three cells wide and two tall, but it is matched at its centre point like any other cell, so its size does not pull extra people in. The quick test ignores paradigm and splits x and y into thirds, giving one of nine camps.
d² = (Δx / (2/13))² + (Δy / (2/12))² + 0.5 × (Δp / 0.25)²
Each ring title states its condition (for example “Speed > 0.6”); meet it and you earn the title, and you can earn several.
Calibration
Where each ideology sits is decided by the test itself, not by its name. For each ideology, its wiki entry alone is given to someone who answers every statement as a committed adherent, blind to the axes and their directions; its coordinates are then scored exactly as anyone’s would be.
People rarely answer every statement at the extreme, so raw scores crowd the middle. Each chart axis has a stretch factor that brings 90% of the ideologies inside the chart’s edges; every test-taker’s scores are stretched by the same factors before being placed, so a person and an ideology with the same answers land on the same cell. Ideologies are then assigned to cells by the Hungarian algorithm, with the least total displacement, and each paradigm level comes from the stretched paradigm score.
scale x 1.4 y 1.39 p 1.52
The answers are in data/calibration/personas/ in the repository, and node scripts/calibrate.mjs redoes the whole calibration. After statements or entries change, they should be answered again and the chart recalibrated.
Pro test
The pro test has 120 core statements and up to 8 follow-ups. Each statement belongs to a sub-dimension of its axis as well as to the axis, and the result scores every sub-dimension. Its question bank is versioned separately, now v2.
| Axis | Sub-dimension | Meaning |
|---|---|---|
| Representation | Knowledge source | Where knowledge comes from. Language end, the text and code humanity has written; World end, perception, video and interaction with the world. |
| Representation | Reasoning medium | What thinking is done in. Language end, words, chains of thought and code; World end, latent states, images, space and simulation. |
| Representation | Embodiment | Whether intelligence needs a body. Language end, a mind without a body is enough; World end, intelligence comes from acting in the physical world. |
| Architecture | Architecture | One model or many parts. Monolith end, a single end-to-end network; Harness end, modules, pipelines and specialised models working together. |
| Architecture | Memory and tools | Where memory and abilities live. Monolith end, in the weights and the context; Harness end, in external memory, retrieval and tools. |
| Architecture | Control | Who decides what happens next. Monolith end, the model's own planning; Harness end, explicit structure such as workflows, behaviour trees and orchestrated agents. |
| Paradigm | Architecture | Whether today's architecture is the road. Paradigm-shift end, something else will replace the Transformer, new or revived; Mainstream end, the Transformer only needs scaling and refinement. |
| Paradigm | Learning objective | What models should learn to do. Paradigm-shift end, objectives beyond predicting the next token or pixel; Mainstream end, prediction plus post-training is enough. |
| Paradigm | Continual learning | Whether learning stops at deployment. Paradigm-shift end, models must keep learning from experience; Mainstream end, train once, then deploy and update by retraining. |
| Openness | Weights and products | Who holds the models. Open end, open weights for everyone; Frontier end, capability delivered through closed products and APIs. |
| Openness | Data and science | How transparent research should be. Open end, open data, recipes and reproducible results; Frontier end, some of this is too risky or too valuable to release. |
| Speed | Pace | How fast progress should go. Acceleration end, as fast as possible; Caution end, slow enough to understand the risks. |
| Speed | Oversight | How much checking AI needs. Acceleration end, regulation and human sign-off are bottlenecks; Caution end, independent testing, interpretability and human control first. |
Follow-ups
After the core statements, the test takes the four cells nearest to you and, from a pool of 30 follow-ups, asks the one whose direction best separates them; it recomputes after each answer and asks at most 8. Each follow-up says which two ideologies it is telling apart. Follow-up answers count toward the score like the core ones.
Consistency
Over the x, y and paradigm axes, consistency = the sum of each axis’s absolute net pull ÷ the sum of every answer’s absolute pull. Someone who agrees with both ends of an axis scores low, and below 40% the result says the position is less certain.
consistency = Σ_axis |Σ answer × weight| / Σ_axis Σ |answer × weight|
Principles
- No real researcher is labelled. Entries cite papers, models, systems and concepts only.
- Every ideology gets its strongest arguments and its fairest critiques.
- Statements are balanced: on each axis about as many push each way when agreed with (validate checks this).
- The question bank is versioned; changing a statement bumps the version, and old links keep the scores they were made with.
Contributing
Everything is a plain file in the repository:
data/ideologies/<id>.yaml place, level, names, motto, relations, works content/wiki/en/<id>.md English entry content/wiki/zh/<id>.md Chinese entry data/questions.yaml every statement and its weights data/tests.yaml the two tests and the version
Before a pull request, run npm run validate: it checks that no two ideologies share a cell, every relation points to a real entry, and both languages have every entry. CONTRIBUTING.md ↗
Sources
The form follows the political tests and charts found online: the two axes of the Political Compass; the multi-axis scoring, open question banks and result bars of 8values and the tests built on it (9Axes, SapplyValues, LeftValues, PolitiScales and others); the entry structure and friends / frenemies / enemies of the Polcompball wiki; and the big ideology charts.