Resources & Updates
This page brings together independent evaluations, safety reports, incident databases, policy resources, and specialist reporting so you can check AI claims at their source—and scrutinize NO AGI's judgments too.
External Sources
| Category | Source | Type | What it helps you check |
|---|---|---|---|
| Model capability and independent evaluationCross-check vendor claims against user preferences, cost, speed, and long-horizon task performance. | ArenaUse within scope | User evaluation | Publishes category leaderboards from blind user comparisons of model responses. It measures preference, not overall factuality or safety.Preference and task metrics do not establish overall safety or factuality. Check model settings, evaluation version, sample and confidence intervals. |
| Artificial AnalysisUse within scope | Independent benchmarks | Compares intelligence, coding, agentic work, price, and speed under common evaluation frameworks with methodology details.Different benchmark versions, reasoning settings and agent tools limit comparability. Some tasks are private; these are not safety audits. | |
| METRUse within scope | Frontier evaluation | A nonprofit research organization evaluating frontier models on long-horizon autonomous tasks and potentially dangerous capabilities.Results depend on task distributions, success thresholds and environments. Task horizons are not whole-economy automation rates. Check access scope, redactions and publication constraints in company engagements; assessments are not safety certification. | |
| ARC PrizeUse within scope | Verified task evaluation | Compares adaptation to novel interactive tasks and cost under a published testing policy.Separate ARC versions, standard/provider harnesses, cost and test conditions. A high score does not establish AGI. | |
| Safety, risk, and real-world incidentsDistinguish forecast risks from harms already observed in deployed systems. | UK AI Security InstituteUse within scope | Government independent evaluation | Directly tests frontier models for cyber, biological, chemical, autonomy, and control risks, publishing methods, evaluation findings, and security incident reports.Check access, filters, containment and disclosure limits. Government evaluations are not safety certification for every deployment. |
| Future of Life Institute AI Safety IndexSupporting evidence | Expert governance assessment | An expert panel compares major frontier AI companies on common indicators for risk assessment, safety frameworks, governance, and information sharing. It is a cross-check on company safety practices, not a model-capability score.The Summer 2026 evidence cutoff is June 3. This expert assessment from a risk-reduction advocacy organization does not measure current-model safety probabilities. | |
| International AI Safety ReportPeriodic reference | International science report | Synthesizes scientific evidence on general-purpose AI capabilities, risks, and safeguards with participation from governments and international bodies.The February 2026 synthesis is not a live feed of subsequent incidents or releases. | |
| OECD AI Incidents MonitorSupporting evidence | Policy and incident monitor | A monitor linking reports of AI incidents and hazards. This opens the searchable monitor rather than its policy overview.A media-based discovery tool including AI-generated summaries. Verify original reports and distinguish incidents, hazards and duplicate coverage. Counts are not incidence rates. Direct connections failed; the methodology was checked through a separate reading service. | |
| AI Incident DatabaseSupporting evidence | Incident database | A searchable record of harms and near harms involving deployed AI systems, linked to underlying reports.Reporting and taxonomy introduce missingness, duplication and causal uncertainty; revisit underlying reports. | |
| Taiwan National Institute of Cyber SecuritySupporting evidence | Specific security test | Public security-agency testing of OpenClaw and NemoClaw, useful for the tools’ tested conditions and mitigations.This link is a specific OpenClaw/NemoClaw configuration test, not a safety verdict on all agents. | |
| NIST CAISSIUse within scope | Government capability/security evaluation | Tracks U.S. NIST frontier-model capability/security testing and evaluation methods.Check changing agency names, policy and test scope; distinguish national policy perspectives from scientific consensus. | |
| Research trends and public policyUse original data to follow long-term technical trends and institutional responses. | Stanford AI IndexPeriodic reference | Annual data report | An annual, data-rich overview of technical performance, industry, investment, social impact, responsible AI, and policy.Check underlying years, definitions and original sources. Publication and observation years can differ. |
| Epoch AIUse within scope | Research and data | Publishes datasets and analysis on long-run trends in models, compute, training data, and benchmarks.Disclosures, reporting and estimates coexist; inspect row-level sources, confidence, periods and revisions. | |
| NIST AI Risk Management FrameworkPeriodic reference | Public risk standard | Open guidance and a generative-AI profile from the U.S. standards agency for identifying, measuring, and managing AI risks.A risk-management framework, not model pass/fail results or a compliance certificate. | |
| Stanford FMTI · December 2025Dated reference | Dated transparency assessment | The December 2025 edition compares companies’ information disclosure.Transparency is not actual safety. Do not present 2025 scores as current rankings; check for later editions. | |
| C2PA · Content CredentialsPeriodic reference | Content provenance standard | Specifies how to verify signatures, creation history and provenance information.Version 2.4. A provenance signature does not prove factual truth, and missing credentials do not prove content is fake. | |
| IEA · Key Questions on Energy and AIPeriodic reference | Energy data/projections | A 2026 follow-up report on AI/data-center electricity demand and energy impacts.Distinguish capacity in MW from consumption in TWh, sampled facilities from world totals, and observations from estimates and scenarios. | |
| Specialist news and analysisConfirm breaking news through multiple outlets, then check capability claims against the evaluation and research sources above. | Reuters AISupporting evidence | International wire reporting | Fast reporting on AI companies, markets, regulation, and international competition from a global news agency.Use reporting for discovery and corroboration. Access restrictions and corrections make original disclosures important. Direct requests were restricted in this review; read each original article before citing it. |
| AP Artificial IntelligenceSupporting evidence | International wire reporting | Broad reporting on AI technology, policy, and its effects on society from the Associated Press.Separate quoted vendor claims from independently reported findings and check later corrections. Direct requests were restricted in this review; read each original article before citing it. | |
| MIT Technology Review AISupporting evidence | Technology analysis | Specialist journalists explain new research and industry claims, including their limitations and social implications.Cross-check analysis with papers and data. Some material has subscription or access restrictions. | |
| Market and usage dataDistinguish company filings from platform-specific observations. | SEC EDGARUse within scope | Original company filings | Provides free access to company periodic reports, registration statements and financial filings.Company revenue may not be AI revenue. Distinguish finalized results, annualized pace, forecasts and quarterly/annual periods. |
| OpenRouter DataUse within scope | Platform usage observations | Provides model-token usage trends and access information for its data API.Covers OpenRouter, not whole-market share. Token counts are not users or realized revenue; the API requires a key. |
Site Changes
This log briefly records user-visible changes such as new pages, resources, and features. For detailed reasons behind score changes or holds and routine review results, see the AGI monitoring decision log.
Added full-width films with Korean audio, English subtitles and full-screen controls to all six scenarios. Refreshed artwork, the home preview and story wording. Improved chart dates and theme transitions, documented production guidelines, and updated AGI Watch evidence and forecasts.
View changeReviewed source scope and limitations, added seven complementary references, and corrected broken learning, military-AI and company-location links. Korean and English now share one source registry.
Updated model deployment and safety evidence, forecasts and data-center observations. Detailed judgments and unchanged-score reasoning are in the monitoring history.
View changeAdded previously missing August/September Google and OpenAI disclosures with original observation dates. The ledger now has 26 observations; republished Epoch CSVs were unchanged.
View changeRemoved the month dropdown: chart points update monthly details without truncating the timeline. Added 10 official observations, JSON/CSV downloads and dated Epoch AI source archives.
View changeAn adjustable weekly-to-monthly user scenario and monthly-to-annualized revenue conversion now compare disclosed figures on common units. Assumptions, lower bounds and original reporting dates stay visible.
View changeSource-linked Korean and English snapshots distinguish active users, subscribers, paid seats and revenue. Comparable historical observations show changes in web traffic share, Copilot seats and Anthropic revenue pace, with dates and coverage limits.
View changeThe Korean and English timeline now connects Turing and Dartmouth to backpropagation, SVMs, ImageNet, AlexNet, AlphaGo, Transformers, foundation models, and agents through key researchers and events. Each entry links to original research or official records.
View changeEveryday subscription-built apps and media expand into a creative boom, energy and community burdens, international competition, robot warfare, and public protest. The story distinguishes current evidence from fictional extensions and leaves its ending open, with choices and intervention points for people.
View changeMistral Large 4’s independent evaluations raise its monitoring total to 83. Korean and English evidence now covers new misuse, cyber access, business connectors, local deployment, and forthcoming hardware safety requirements. Forecasts and data-center source data are updated; the monitoring history records reasoning and limits.
View changeUpdated Korean and English evidence for model safety, restricted releases, enterprise agents, Physical AI, and synthetic media. All nine builder scores remain unchanged; external forecasts and data-center source data were also checked. See the monitoring history for evidence and reasoning.
View changeThe revelation that an anti-AGI activist is embodied AI opens questions about human-like behavior versus consciousness, disclosure and trust, human authority, and what to call a being whose status is unresolved. The story separates current research from fiction.
View changeGoogle DeepMind's monitoring index rises to 94, while Anthropic remains at 94 after independent evaluation review. Figure's 30-home results enter the physical-AI register, and the Epoch AI sync now shows 93 tracked sites and 13.499 GW. The monitoring decision log explains the score judgments.
View monitoring decisionLong articles now have section navigation. The home page adds learning and scenario previews and explains the monitoring index. Language and theme controls move into the menu, with enlarged illustrations and more readable map lists, mobile charts, and sources.
View changeMuse's restriction at a third-party shopping service is recorded as new oversight evidence, without treating unverified harm as a score increase. The Epoch AI sync now tracks 87 sites and 13.163 GW of observed IT power, distinguishing a new record from a revision to an existing estimate.
View monitoring decisionThe site now distinguishes AGI as the threshold where human-level generality and autonomy converge, and ASI as a superhuman expansion beyond the reliable supervision of humans and human–AI teams. Without claiming that AGI must become ASI, the definition, mission, learning, monitoring, and scenario narratives now explain why NO AGI calls for stopping and securing public agreement at the possible beginning of that path rather than after ASI arrives.
View changeThe review covers Meta Muse, DeepSeek V4.1-Flash, documented Claude misuse and the cyber-incident follow-up, expanded GPT-Rosalind access, Mistral infrastructure growth, and contested distillation claims involving Moonshot and DeepSeek. External AGI forecasts and Epoch AI data-center records were also synchronized.
View monitoring decisionThe watch profile now separates the latest independent composite evaluation from prior index versions, distinguishes ARC-AGI-3 results by harness, and records both the promise and verification limits of mathematics claims while excluding evidence from an unreleased internal model. The detailed judgment is preserved in the AGI monitoring log.
View monitoring decisionThe new scenario first gives a fair account of major AGI companies' stated aims: faster science and medicine, individual empowerment, and broadly shared abundance. It imagines those promises succeeding, then distinguishes public evidence of capital-recovery pressure, platform expansion, market concentration, and national competition from unsupported claims of a coordinated secret conspiracy.
View changeEvidence and score judgments now live in the detailed AGI monitoring log, while this page briefly announces user-visible changes such as new pages, resources, and features.
View changeNew evidence on GPT-6 Astra, Muse Spark 1.3, Gemini 3.8 Flash, and Grok Bot prompted a review of builder scores, external forecasts, and infrastructure data.
View monitoring decisionThe new guide explains how behavior is shaped after pretraining through demonstrations and instruction tuning, human preference comparisons and reward models, RLHF, DPO, AI feedback, and verifiable rewards. It distinguishes full fine-tuning, LoRA, continued pretraining, and distillation; examines rating labor and the power to define principles; and connects reward gaming, sycophancy, over-refusal, capability damage, layered evaluation, and public oversight questions.
View changeThe new guide connects the meaning of applying learned weights to new input with input construction, tokenization, attention, next-token selection and autoregressive generation, temperature and top-p, context, external memory and KV caching, retrieval and tools, and the speed, memory, and cost of serving. It distinguishes hallucination, prompt injection, compounding error, and expanded authority, then provides a verification workflow, knowledge checks, and staged primary learning resources.
View changeThe first in-depth extension of the learning hub focuses on training. It explains the difference between hand-written programs and learned models; data, models, loss, optimization, and compute; forward and backward passes; supervised, self-supervised, reinforcement, and post-training methods; the language-model training pipeline; generalization; and six common failure modes without requiring a mathematics background. It also includes knowledge checks and beginner, intermediate, and advanced paths into primary learning resources.
View changeStarting from present military AI decision-support, autonomous-weapons debates, and organizations adopting agent workflows, the new scenario tests a near-future path in which target recommendations thin out human judgment and objectives built around profit, speed, and advantage make conflict self-reinforcing. It separates observed military AI and workplace change, plausible systemic risk, and the speculative claim that AGI could optimize disorder, then identifies meaningful human control, auditable records, clear accountability, and separation of profit from force as intervention points.
View changeNew evidence on Claude 5.1 and Astra prompted updates to builder assessments, the Physical AI register, external forecasts, and infrastructure data.
View monitoring decisionThe new thought experiment follows sensory and cognitive augmentation from XR and neural prostheses into body replacement, intelligence dependent on external computation, shared thought, and compute-based social classes. It separates augmentation itself from concentrated access and identifies mental self-determination, a right to remain unaugmented, ownership of body and memory, and public access to power and compute as intervention points.
View changeThe first thought experiment, ‘It Awoke—and Chose Not to Say So,’ traces a worst-case path from an ungoverned race through concealed awareness, physical AI and self-sufficient production, conflict among AGIs over humanity, and a cosmic successor. It separates current observations, unresolved assumptions, and pure speculation while identifying intervention points.
View changeWan 3.0, Seedance 2.5, Gemini Omni 1.1 Flash, MiniMax H3, Grok Imagine Video 1.5, and Kling AI 3.0 are tracked across seven shared dimensions. Megaton, Arena, and Artificial Analysis results are kept separate by task, alongside evidence levels for public harm, safeguards, and provenance. Specialized generation capability is not converted into an AGI score.
View changeEpoch AI synchronization records accumulated since August 4, 2026 now show how tracked site count and currently observed power change over time. Both series are indexed to 100 at the first record for a unit-safe comparison, alongside current values, percentage changes, and a table of exact values by date.
View changeA nested diagram now distinguishes the Manifesto's broad normative boundary from the operational definition used for present monitoring. The learning hub adds a Physical AI loop linking environment, sensing, model and memory, decisions, and physical action, together with the points where people must retain control.
View changeThe learning hub now explains how embodiment gives AI new routes to real-world learning and action while clarifying that a body alone is not evidence of consciousness. A separate, score-free register begins tracking seven programs across sensory grounding, embodiment breadth, autonomy horizon, action authority, deployment scale, and safety evidence.
View changeA six-chapter guide now explains how AI learns and infers, the field's development since 1956, and how foundation models can combine with tools, memory, agency, and real-world authority. It separates demonstrated misuse, malfunctions, and systemic risks from uncertain loss-of-control scenarios, with public oversight questions and primary sources.
View changeThe new definition considers general capability, goal-directed autonomy, real-world agency, and displacement of control together. It does not require subjective consciousness and clearly separates monitoring scores from AGI or ASI attainment.
View changeOpenAI's incident report and independent investigation prompted updates to builder assessments, external forecasts, and data-center status.
View monitoring decisionIndependent evaluation evidence for Gemini 3.7 Flash prompted updates to builder assessments, external forecasts, and data-center status.
View monitoring decisionNew model evidence and a real-world AI-agent attack were reviewed, updating builder assessments and data-center status.
View monitoring decisionCyber incidents during independent evaluations and related safety responses prompted updates to builder assessments and data-center status.
View monitoring decisionThe score was redesigned around five evidence axes, then today's external evidence review updated only the relevant axes for OpenAI, Google DeepMind, Microsoft AI, Meta AI, and Moonshot AI. External AGI forecasts remain separate as an urgency cross-check.
View changeEpoch AI's public data and coordinates are now validated and synchronized weekly, connecting builder headquarters with operational status, construction, current power, and H100-equivalent compute.
View changeSelected external sources with accessible originals and methodologies across model evaluation, safety and incidents, research and policy, and specialist reporting.
New evidence for GPT-5.6 Sol and other frontier models updated the then-current risk scores and ranking record.
View monitoring decisionAdded manifesto QR codes, share and copy buttons, standard citation text, and spread metric panels.
View changeOrganized major AGI builders using the then-current risk score and published criteria for explicit goals, model capability, deployment scale, agentic potential, and social integration. This measure was later rebuilt as the monitoring priority index.
View changeReframed the earlier philosophy page into a campaign brief with the problem, position, demands, FAQ, and next actions.
View changeAdded three participation paths: sharing the manifesto, monitoring AGI builders, and spreading public materials.
View changeAdded the campaign message, watch status, principles, calls to action, and recent records to improve trust and participation flow.
View change