Superpower Daily / Research desk
Evidence you can interrogate.
Public AI market reports built from our verified signals, editorial coverage, directories, and first-party measurements. Explore the chart, inspect the records, cite the method, or take the data with you.
56Public datasets
VERIFIED INPUTS / REPRODUCIBLE OUTPUTS
9,074 published rows 817 cited sources CSV / JSON / SVG / PNG
Public market reports
11 measured views of the AI economy
- 01
Price and capability
AI Model Economics Index
Observed model prices, context windows, licenses, and release economics from verified launch records.60Priced models - 02
Weekly funding and acquisitions / Weekly
Superpower AI Capital Ledger
Weekly disclosed AI funding and acquisition totals, transaction counts, stage mix, valuations, and named investors.$0Latest week - 03
Release momentum
Open Model Momentum
Verified open and disclosed-license model releases, developer activity, and release cadence.77Open releases - 04
Proposal to enforcement
AI Policy Implementation Clock
Policy actions, effective dates, status movement, jurisdictions, and implementation lag.132Policy actions - 05
Risk monitor
AI Security Pulse
Verified security-risk coverage, incident velocity, source depth, and reader attention.59Risk records - 06
Enterprise adoption evidence / Weekly
Superpower Agent Deployment Monitor
A weekly record of named agent deployments, production rollouts, pilots, integrations, and enterprise adoption evidence.70Adoption records - 07
Ecosystem structure
AI Platform Dependency Map
A relationship map of tracked tools, companies, products, and the entities they depend on or represent.30Mapped nodes - 08
Cohort health
AI Launch Survival Report
Observed Product Hunt cohorts, launch volume, selected products, and currently tracked tools by week.35Observed launches - 09
Developer cadence
Model Release Velocity
Release frequency, developer share, model availability, and cadence from verified model launches.215Verified releases - 10
Reader behavior
AI Attention vs. Substance Index
Reader attention plotted against editorial importance, source depth, completion, saves, and shares.500Measured stories - Q
Quarterly market report / Quarterly
Superpower State of the AI Market
A quarterly synthesis of verified capital, model-release, infrastructure, policy, and enterprise-adoption signals.7852026-Q4
Live observatories
Follow the systems behind the headlines
- 11
Data centers, compute and power / Continuous
Superpower AI Capacity Watch
Source-backed data-center, power-capacity, chip-supply, and infrastructure developments shaping AI capacity.1197Capacity signals - 12
Public procurement
Government AI Contract Tracker
Reported AI awards, procurement programs, agencies, vendors, and publicly disclosed contract activity.295Contract records - 13
Courts and claims
AI Litigation and Copyright Docket
AI lawsuits, copyright disputes, court decisions, jurisdictions, and the organizations involved.511Docket records - 14
Price, context and releases / Continuous
Superpower Frontier Model Ledger
A dated ledger of disclosed model prices, context windows, availability, and release conditions.215Dated observations - 15
Hiring, layoffs and skills
AI Workforce Impact Monitor
Source-backed signals on AI hiring, workforce reductions, changing roles, skills, and labor-market exposure.356Workforce records - 16
Disclosure quality
AI Transparency Scorecard
Compares model developers by the launch fields they disclose, without substituting missing facts with estimates.130Developers scored - 17
Projects and maintainers
Open-Source AI Ecosystem Health
Tracks explicitly open AI tools, release activity, categories, mentions, and disclosed project destinations.90Open tools - 18
Service interruptions
AI Outage and Reliability Monitor
Published reports of AI outages, downtime, degraded service, availability incidents, and recovery events.94Reliability records - 19
Evaluations under pressure
AI Benchmark Saturation Tracker
Tracks reported benchmark results, leaderboard movement, evaluation releases, and reliability concerns.1017Evaluation records - 20
Commercial relationships
AI Partnership and Dependency Graph
Maps organizations appearing together in verified partnership, adoption, integration, and infrastructure coverage.80Mapped organizations
Original investigations
Questions the daily news leaves unanswered
- R1024
comparison
Claude Code’s documented task cancellation does not guarantee a remote job stops
The sampled documentation separates client interaction, protocol task state and backend execution. Cancelling a call, acknowledging cancellation or ending a session does not establish that delegated remote work has terminated.0Verified observations - R1023
comparison
Muse’s 1Password FAQ Promises Sensitive-Action Confirmation; Grok Bot’s Sampled FAQ Does Not
The sampled 1Password documents distinguish approval to sign in from permission to act afterward. They document sensitive-action confirmation for Muse but leave the equivalent Grok Bot boundary unspecified—not proven absent.0Verified observations - R1021
claim audit
llm-d 1.0’s passing TPU badges do not validate every documented TPU path
The sampled records distinguish release-validation lanes, checked passing badges and documented configurations, but do not establish exact v1.0.0 runtime validation for every path—or validation of routing, offload and disaggregation together.0Verified observations - R1017
comparison
Sampled Muse Spark Open Recipes Document Four-Input NVIDIA Paths and Text-Only AMD Setups
Documented input support depends on the checkpoint, runtime and configuration—not merely whether the model fits the hardware. Missing modality documentation does not establish missing capability.0Verified observations - R1016
claim audit
Five of eight Claude circuit comparisons are specification-matched on paper
The supplied evidence supports five specification-matched comparisons, including bounded numerical approximations. The remaining comparisons require qualifications about permitted failures or comparator provenance; no single equal-guarantee improvement describes all eight tasks.0Verified observations - R1014
benchmark analysis
Google’s geospatial tests show larger score gains from specialized capabilities than added planning and guidance
The published comparisons support descriptive gains from a specialized-capability package, but do not isolate skill packaging, planning alone or a planning-by-skills interaction.16Verified observations - R1014
comparison
Buyer consent does not replace Williams Sonoma’s permission for AI shopping agents
Buyer authorization and retailer permission are separate requirements in some checked shopping-agent workflows. Their scope varies by retailer, action and integration; the evidence does not support a universal ban or unrestricted permission.0Verified observations - R1013
comparison
Light Claude or GPT use can make switching Cursor Teams cheaper—but the token surcharge stays
Switching savings depend on the subscription-price reduction, existing credits actually used and changes in first-party overages—not simply Cursor’s claimed increase in Grok coding usage. The documented team token surcharge applies before and after switching.4Verified observations - R1011
claim audit
Gemini’s documented SynthID checks are Google-only; the standalone detector names partners
Google’s partner announcement does not establish interchangeable verification tools or complete product-level coverage. The supplied records document specific watermark insertion paths, announced integrations and unresolved compatibility boundaries.0Verified observations - R1010
policy impact
Unreviewed OSS Scanner reports do not start Anthropic’s 90-day disclosure clock
The supplied documents distinguish automated delivery from notification of Anthropic’s later human validation. They also separate Anthropic’s publication timing from maintainers’ review, sharing and confidentiality obligations, while leaving an unresolved boundary between scanner-specific timing and the linked coordinated-disclosure policy.0Verified observations - R1009
claim audit
Selecting MXC Alone Does Not Restrict Networking in Codex 0.162.1
Codex 0.162.1’s release-tagged source distinguishes backend preference, operation-level sandbox selection and generated policy. MXC selection alone does not establish restricted networking: a network-enabled permission profile without managed networking generates allow defaults for egress, ingress and host loopback.0Verified observations - R1009
claim audit
AgentCore’s MMDSv2 Requirement Leaves Existing Sessions Unaffected
The sampled records distinguish invocation-time metadata enforcement from generated IAM policy changes. They establish configuration requirements and narrower toolkit templates, but not automatic migration of previously created customer roles.0Verified observations - R0928
policy impact
OpenShell 0.1.0’s Read-Only API Rules Need Enforce Mode to Block Writes
NVIDIA’s versioned documentation distinguishes connection denial from request-level enforcement. For an allowed REST endpoint, a read-only rule in the default audit mode logs and forwards a disallowed write; enforce mode blocks it.9Verified observations - R0926
claim audit
Detected Is Not Stopped: A Timeline of OpenAI’s New Agent Reports
Detection, human acknowledgement, containment and remediation are distinct stages. The three public reports do not support a uniform response-time claim.15Verified observations - R0922
claim audit
Databricks Genie One MCP: Does “Governed” Cover the Write Surface?
The reviewed public record documents layered identity, provider-permission, and logging controls, but connector-specific approval, cancellation, rollback, and recovery rules remain unevenly specified. Genie One MCP’s documented tool surface is distinct from the separate SaaS connector write surface.8Verified observations - R0927
comparison
Agent Audit Trails: Which Records Are Outside the Agent’s Reach?
A visible transcript is not necessarily an independently preserved audit record. Capture location, recording defaults, coverage, deletion paths and export timing determine what an operator may be able to reconstruct.3Verified observations - R0926
trend investigation
Chinese Models’ ‘Majority Share’: Tokens, Requests or Spending?
The supplied evidence cannot establish August shares for Chinese-origin models on any of the three Vercel measures. Vercel’s published token and spend figures classify open-weight models, which is a different category.3Verified observations - R0925
claim audit
Google’s Ten-Minute AI Film Tests One System, Not Four
Google attributes the ten-minute film to A²RD. The four linked papers evaluate different artifacts and do not report a test of all four systems as one integrated pipeline.9Verified observations - R0924
comparison
Three Meta Privacy Boundaries, Not One: What Protects Muse and AI Glasses?
Meta describes three distinct processing paths with different availability and data-flow scopes. A protection described for one path should not be assumed to cover another.8Verified observations - R0923
benchmark analysis
Which Model Earned Opus 5.5’s Agent-Benchmark Score?
The published scores describe different evaluation configurations. Four selected Opus 5.5 rows permitted production-safeguard fallback to older Claude models; Zapier’s AutomationBench run did not. Permission to fall back does not show that fallback occurred.11Verified observations - R0922
claim audit
Nscale’s $100B Backlog Test: Revenue, Capacity, or Conditional Promise?
The filing supports a lower bound for condition-affected contract value, but not an exact calculation of unconditional minimum payments. It separates active TCV, contracted TCV, maximum contract values, revenue, and active capacity—measures that should not be collapsed into one backlog figure.10Verified observations - R0922
comparison
Existing Nuclear Power vs. Future Reactors: The AI Power-Contract Timeline
The bounded sample mixes operating-plant purchases with prospective generation, uprates and development rights. A power headline is not, by itself, a common measure of electricity available to an AI load on a given date.14Verified observations - R0920
claim audit
Can Codex Hook Policies See the Tool Calls They Claim to Govern?
The checked source supports a bounded finding: hooks cover most enumerated static built-in handler types, but selective exclusions and hosted paths mean they are not a complete action-governance boundary.4Verified observations - R0920
benchmark analysis
Vibe Code Bench’s Score Reset: Model Regression or a Different Test?
The apparent collapse is not interpretable as a model-performance reversal from the public record alone: the earlier benchmark and VCB 1–100 change the task, scoring, budget, state, and evaluation structure.13Verified observations - R0916
comparison
After RubyGems: Can Major Package Registries Make Agent-Scale Publishing Reviewable Before Release?
All three registries document workflow-bound, short-lived publishing authority, but that control does not itself require human review. In the reviewed public materials, npm documents a registry-native staged path with a maintainer 2FA release gate; RubyGems and PyPI can publish directly from trusted workflows unless maintainers add an external CI approval rule.6Verified observations - R0912
comparison
When an Agent-Security Gate Is Unsure: Block, Ask, or Allow-and-Log?
“Pre-execution security” is not a standardized product capability. In this bounded public-evidence review, inspection, workflow approval, runtime authorization, and execution environments are distinct controls with different documented defaults and boundaries.15Verified observations - R0916
policy impact
Sponsored Agents in ChatGPT Ads: Where Does an Ad End and an Action-Capable Agent Begin?
The bounded public record separates ordinary-ad safeguards from Sponsored Agent accountability. It assigns advertisers responsibility for Sponsored Agent content, configurations, Actions, Output, and Ad Materials, but does not publicly specify several agent-level safeguards that ordinary-ad materials describe or that an action-capable experience could require.10Verified observations - R0915
claim audit
After the Answer-Key Leak: Which SWE-bench Pro Scores Still Identify a Verifiable Dataset Version?
The public record establishes leakage and task-quality concerns, but the decisive procurement question is score-level provenance. In the bounded public sample, no visible SWE-bench Pro row preserves the complete metadata chain needed for strict historical or pre/post-remediation comparison.30Verified observations - R0914
claim audit
Bolt Forge’s 50× Trade: What a Builder Consents to Donate for Open-Model Training
The supplied record describes a substantial opt-in bargain, including broad Forge Content classes, mode-specific consent, some exclusions, and prospective withdrawal controls. It does not fully specify the transformation procedure, secret-removal limits, retention of training copies, the non-public Bolt-Arcee agreement, or deletion from completed training and delivered datasets.20Verified observations - R0910
claim audit
The OneGov Price Reset: What Federal Agencies Actually Pay After September 30
The public record supports a two-phase reading: the $1-per-agency ChatGPT Enterprise promotion ends September 30, 2026, while a distinct consumption-based agreement is expected to begin October 1. Zero platform fees do not make all government AI use free.31Verified observations - R0902
comparison
The Flash-vs-Frontier Cost Curve
The published token-price gap matters only if Gemini 3.8 Flash preserves completion rates after agent token use, retries, caching, and output length are counted.0Verified observations - R0906
policy impact
The 2026 Data-Center Moratorium Ledger: What Local Pauses Actually Stop
In the saved, bounded ledger, local measures are not operationally uniform. The evidence more often describes temporary planning and permitting controls than demonstrated stops to already authorized construction, with exemptions and legal status materially affecting project exposure.0Verified observations - R0906
policy impact
MCP’s Optional-Authorization Reality Check
The saved evidence indicates a gap between MCP’s security guidance and its mainstream starter HTTP flows: authorization capabilities and security requirements exist, but authentication and scoped authorization are opt-in in the four sampled Tier-1 SDK paths.0Verified observations - R0904
benchmark analysis
Astra’s Computer-Use Claims: How Much Is Model, and How Much Is Harness?
Astra may be highly capable, but the disclosed results do not consistently isolate model performance from the surrounding agent system.29Verified observations - R0903
comparison
Cache Economics Beyond the Headline: When Does Fable or Gemini Win on Repeated Context?
The lower cache-read price does not alone determine the lower-cost workflow: cache-write charges, eligibility, hit behavior, expiry, and output tokens can change the result.24Verified observations
The publishing contract
No invented history. No hidden denominator.
01 / Evidence Every public number resolves to a stored record, a source, or an explicitly labeled first-party measurement.
02 / Limits Gaps, prospective baselines, and sampling bias are disclosed beside the chart.
03 / Portability The table, CSV, JSON, SVG, and PNG are generated from the same report dataset.