{"atlas":{"title":"Synthetic Intelligence Atlas","canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence","version":"1.1.0","lastReviewed":"2026-07-28","sourcesEndpoint":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources.json"},"count":18,"concepts":[{"id":"large-language-models","label":"Large language models","definition":"Neural networks trained on large text corpora to predict tokens, then adapted to follow instructions and produce text or actions.","whyItMatters":"The object every other concept in this map is about. Nearly all capability discussion reduces to statements about these systems under particular scaffolding.","notEstablished":"That these systems possess a general capability corresponding to human intelligence. No source in this map supports a single scalar capability score.","canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/large-language-models","relatedConceptIds":["transformers","pre-training","post-training","scaling-laws"],"sourceIds":["vaswani2017","brown2020","bommasani2021"],"claimIds":["si-001","si-002","si-005"],"sources":[{"id":"vaswani2017","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/vaswani2017"},{"id":"brown2020","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/brown2020"},{"id":"bommasani2021","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/bommasani2021"}]},{"id":"transformers","label":"Transformers","definition":"A sequence architecture built on attention rather than recurrence or convolution.","whyItMatters":"Architectural properties set what can be scaled and what the inference cost profile looks like.","notEstablished":null,"canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/transformers","relatedConceptIds":["large-language-models","pre-training","inference-time-compute"],"sourceIds":["vaswani2017"],"claimIds":["si-001"],"sources":[{"id":"vaswani2017","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/vaswani2017"}]},{"id":"scaling-laws","label":"Scaling laws","definition":"Empirical relations between training compute, parameters, data, and loss.","whyItMatters":"Scaling laws are the basis of most quantitative capability forecasting, so their contested regions propagate directly into forecast uncertainty.","notEstablished":"That a loss curve extrapolates to a capability claim. Scaling laws are fitted on loss, not on task competence, and the mapping between them is not established.","canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/scaling-laws","relatedConceptIds":["pre-training","ai-forecasting","large-language-models"],"sourceIds":["kaplan2020","hoffmann2022","chowdhery2022"],"claimIds":["si-003","si-004","si-007"],"sources":[{"id":"kaplan2020","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/kaplan2020"},{"id":"hoffmann2022","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/hoffmann2022"},{"id":"chowdhery2022","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/chowdhery2022"}]},{"id":"pre-training","label":"Pre-training","definition":"The large-scale self-supervised phase in which a model learns from an unlabelled corpus.","whyItMatters":"Determines what is in the model before any alignment, and is where benchmark contamination enters.","notEstablished":null,"canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/pre-training","relatedConceptIds":["scaling-laws","benchmark-contamination","post-training"],"sourceIds":["brown2020","kaplan2020","hoffmann2022","grattafiori2024","lewkowycz2022"],"claimIds":["si-002","si-003","si-004","si-014"],"sources":[{"id":"brown2020","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/brown2020"},{"id":"kaplan2020","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/kaplan2020"},{"id":"hoffmann2022","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/hoffmann2022"},{"id":"grattafiori2024","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/grattafiori2024"},{"id":"lewkowycz2022","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/lewkowycz2022"}]},{"id":"post-training","label":"Post-training","definition":"Adaptation after pre-training — supervised fine-tuning, preference optimization, reinforcement learning from human or model feedback.","whyItMatters":"Most of the difference between a raw pre-trained model and a deployed assistant is here, which means capability comparisons across models often compare post-training pipelines.","notEstablished":null,"canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/post-training","relatedConceptIds":["pre-training","inference-time-compute","evaluation"],"sourceIds":["ouyang2022"],"claimIds":["si-005","si-006"],"sources":[{"id":"ouyang2022","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/ouyang2022"}]},{"id":"inference-time-compute","label":"Inference-time compute","definition":"Additional computation spent at generation time — longer reasoning traces, sampling with verification, or search over candidate outputs.","whyItMatters":"It breaks the assumption that a benchmark score is a property of a model. The same weights score differently under different inference budgets.","notEstablished":"That inference-time gains substitute indefinitely for training scale. The trade-off is reported within studied regimes, not as a general law.","canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/inference-time-compute","relatedConceptIds":["post-training","evaluation","agent-loop"],"sourceIds":["wei2022","cobbe2021","snell2024"],"claimIds":["si-006","si-007"],"sources":[{"id":"wei2022","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/wei2022"},{"id":"cobbe2021","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/cobbe2021"},{"id":"snell2024","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/snell2024"}]},{"id":"retrieval-augmented-generation","label":"Retrieval-augmented generation","definition":"Conditioning generation on documents fetched at inference time from an external store.","whyItMatters":"Separates what a system knows from what its weights encode, which changes what a capability measurement is measuring.","notEstablished":null,"canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/retrieval-augmented-generation","relatedConceptIds":["tool-use","large-language-models","evaluation"],"sourceIds":["lewis2020"],"claimIds":["si-008"],"sources":[{"id":"lewis2020","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/lewis2020"}]},{"id":"tool-use","label":"Tool use","definition":"Model-initiated invocation of external functions — search, code execution, APIs — with results returned into context.","whyItMatters":"Tool access frequently accounts for more measured improvement than model changes, and is often not held constant across comparisons.","notEstablished":null,"canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/tool-use","relatedConceptIds":["agent-loop","retrieval-augmented-generation","coding-agents"],"sourceIds":["schick2023","yao2022"],"claimIds":["si-008","si-009"],"sources":[{"id":"schick2023","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/schick2023"},{"id":"yao2022","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/yao2022"}]},{"id":"agent-loop","label":"Agent loop","definition":"An iterated cycle of reasoning, acting through tools, and observing results, continuing until a stopping condition.","whyItMatters":"The loop and its environment are a system distinct from the model inside it. Attributing loop performance to the model is the most common category error in capability reporting.","notEstablished":null,"canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/agent-loop","relatedConceptIds":["tool-use","coding-agents","long-horizon-task-completion","reliability-compounding-error"],"sourceIds":["yao2022","liu2023agentbench"],"claimIds":["si-009","si-010","si-019"],"sources":[{"id":"yao2022","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/yao2022"},{"id":"liu2023agentbench","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/liu2023agentbench"}]},{"id":"coding-agents","label":"Coding agents","definition":"Agent systems operating on real codebases through file, shell, and test interfaces.","whyItMatters":"The most measured agentic domain, and the clearest demonstration that interface design changes outcomes independently of the model.","notEstablished":null,"canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/coding-agents","relatedConceptIds":["agent-loop","tool-use","evaluation"],"sourceIds":["jimenez2023","yang2024"],"claimIds":["si-010","si-011","si-023"],"sources":[{"id":"jimenez2023","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/jimenez2023"},{"id":"yang2024","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/yang2024"}]},{"id":"long-horizon-task-completion","label":"Long-horizon task completion","definition":"The ability to carry a task requiring many dependent steps to completion, measured here by the time a human would take on tasks a system completes at a given success rate.","whyItMatters":"It is the capability most relevant to autonomy claims and the one where measurement is least mature.","notEstablished":"That a measured horizon transfers to real-world software work. The source paper explicitly frames external validity as a limitation and its forward statement as conditional.","canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/long-horizon-task-completion","relatedConceptIds":["agent-loop","reliability-compounding-error","ai-forecasting"],"sourceIds":["kwa2025"],"claimIds":["si-018","si-019"],"sources":[{"id":"kwa2025","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/kwa2025"}]},{"id":"reliability-compounding-error","label":"Reliability and compounding error","definition":"The tendency of per-step failure probabilities to accumulate across a multi-step trajectory.","whyItMatters":"It explains why systems that look strong on single-step benchmarks can fail on long tasks, and why error recovery matters more than peak accuracy.","notEstablished":null,"canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/reliability-compounding-error","relatedConceptIds":["long-horizon-task-completion","agent-loop","evaluation"],"sourceIds":["kwa2025","liu2023agentbench"],"claimIds":["si-018","si-019"],"sources":[{"id":"kwa2025","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/kwa2025"},{"id":"liu2023agentbench","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/liu2023agentbench"}]},{"id":"evaluation","label":"Evaluation","definition":"The practice of measuring model or system behaviour against defined tasks and metrics.","whyItMatters":"Every capability statement is an evaluation statement. What was measured, under what scaffolding, by whom, decides what the number means.","notEstablished":null,"canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/evaluation","relatedConceptIds":["construct-validity","benchmark-contamination","governance-accountability"],"sourceIds":["liang2022","srivastava2022","bowman2021","liu2023agentbench"],"claimIds":["si-010","si-011","si-012","si-013","si-014","si-015","si-016","si-017","si-020","si-023"],"sources":[{"id":"liang2022","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/liang2022"},{"id":"srivastava2022","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/srivastava2022"},{"id":"bowman2021","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/bowman2021"},{"id":"liu2023agentbench","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/liu2023agentbench"}]},{"id":"construct-validity","label":"Construct validity","definition":"Whether a measurement actually measures the property it is claimed to measure.","whyItMatters":"It is the difference between \"scored well on this benchmark\" and \"is capable\". This atlas keeps those separate everywhere.","notEstablished":"That any benchmark in this map measures general intelligence. Each measures a defined task distribution, and its authors say so.","canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/construct-validity","relatedConceptIds":["evaluation","benchmark-contamination","large-language-models"],"sourceIds":["raji2021","bowman2021","chollet2019","hendrycks2020","rein2023"],"claimIds":["si-011","si-012","si-013","si-015","si-016","si-023"],"sources":[{"id":"raji2021","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/raji2021"},{"id":"bowman2021","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/bowman2021"},{"id":"chollet2019","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/chollet2019"},{"id":"hendrycks2020","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/hendrycks2020"},{"id":"rein2023","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/rein2023"}]},{"id":"benchmark-contamination","label":"Benchmark contamination","definition":"Presence of evaluation data, or close variants, in training corpora.","whyItMatters":"Contamination inflates scores without capability change, and is the main reason benchmark trends cannot be read as capability trends.","notEstablished":null,"canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/benchmark-contamination","relatedConceptIds":["pre-training","evaluation","construct-validity"],"sourceIds":["sainz2023","deng2023"],"claimIds":["si-013","si-014","si-023"],"sources":[{"id":"sainz2023","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/sainz2023"},{"id":"deng2023","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/deng2023"}]},{"id":"local-inference","label":"Local and on-device inference","definition":"Running models on user-controlled hardware rather than a hosted API.","whyItMatters":"It changes the cost, privacy, and reproducibility profile of a system, and makes an evaluation reproducible in a way a served endpoint is not.","notEstablished":"Any claim about the current capability gap between local and hosted models. No source in this map measures that gap as of the evidence cutoff.","canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/local-inference","relatedConceptIds":["evaluation","governance-accountability","transformers"],"sourceIds":["grattafiori2024","chen2023drift"],"claimIds":["si-017"],"sources":[{"id":"grattafiori2024","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/grattafiori2024"},{"id":"chen2023drift","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/chen2023drift"}]},{"id":"ai-forecasting","label":"AI forecasting","definition":"Projection of future capability from historical measurements, expert elicitation, or trend extrapolation.","whyItMatters":"Forecasts drive policy and capital allocation, and their failure modes are documentable rather than hypothetical.","notEstablished":"That any projection in this map is a prediction. Forecast-status claims are conditional scenarios with stated assumptions and failure modes.","canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/ai-forecasting","relatedConceptIds":["scaling-laws","long-horizon-task-completion","construct-validity"],"sourceIds":["grace2024","kwa2025","srivastava2022"],"claimIds":["si-018","si-021"],"sources":[{"id":"grace2024","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/grace2024"},{"id":"kwa2025","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/kwa2025"},{"id":"srivastava2022","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/srivastava2022"}]},{"id":"governance-accountability","label":"Governance and accountability","definition":"Mechanisms for disclosure, reporting, and responsibility around deployed models.","whyItMatters":"Reporting conventions determine what outsiders can check, which determines whether any capability claim is auditable.","notEstablished":null,"canonicalUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/concepts/governance-accountability","relatedConceptIds":["evaluation","construct-validity","local-inference"],"sourceIds":["mitchell2019","bommasani2021","openai2023gpt4"],"claimIds":["si-020","si-021"],"sources":[{"id":"mitchell2019","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/mitchell2019"},{"id":"bommasani2021","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/bommasani2021"},{"id":"openai2023gpt4","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/openai2023gpt4"}]}]}