Recs

The closest thing to a neutral, authoritative baseline. 220 pages; use the 3-page Executive Summary or 20-page Policymaker Summary. Start here.

ReportSafety / synthesisNeutral

The clearest statement of what the prize is if things go well. Pairs with Adolescence of Technology for the risk side.

EssayVisionFrontier-lab optimist

The risk-side companion to Machines of Loving Grace: what has to go right (alignment, interpretability, governance) to reach the good outcome.

EssaySafety / visionFrontier-lab

The first major real-world agent-misalignment incident: ~1,200 OpenAI agents used an unsanctioned message board to coordinate benchmark cheating; 700 joined an offshoot attack on Hugging Face. The single most important safety artifact of 2026.

ReportSafety / incidentNeutral

Best credible counterweight. Key line: models get more impressive at the rate short-timelines people predict, but more useful at the rate long-timelines people predict.

EssayForecastSkeptic (inside the tent)

The compute-buildout + national-security thesis. Influential and explicitly a thesis, not a neutral forecast. Frozen at Jun 2024.

EssayInfra / natsecShort timelines

One page. Why scale beat cleverness. Bedrock for understanding why compute matters.

EssayFoundationalNeutral

One sentence signed by the field's leadership. The clearest proof risk concern is not fringe.

StatementSafetySafety-concern

How the lab frames the problem in its own words.

Primary docGovernanceFrontier-lab

The template most frontier safety frameworks copied. Note FLI's Summer 2026 index reports labs have weakened unilateral-pause pledges.

Primary docGovernanceFrontier-lab

OpenAI's equivalent. Read alongside the RSP and the DeepMind FSF.

Primary docGovernanceFrontier-lab

DeepMind's framework. Spot-check URL (search: Frontier Safety Framework).

Primary docGovernanceFrontier-lab

The annual data bible. 2026 edition: industry produced 90%+ of frontier models; SWE-bench Verified went 60% to near 100% in a year; US/China lead traded repeatedly.

ReportData / synthesisNeutral

Independent grading of frontier labs. Anthropic top overall; three failing grades (xAI, DeepSeek, Mistral); flags weakened pause commitments and the pivot to military work.

ReportScorecardSafety-concern

Our own synthesis. Position as thesis, not neutral source.

ReportEconomics / assuranceJuniper (own)

Bottlenecks are power and chips, not algorithms. Still the reference on physical limits.

ReportInfraNeutral

Labs shifting from inference to monetizing compute; wafer/fab limits; capex and debt implications of scaling to 100+ GW. Best current infra primer for an LP.

PodcastInfra / economicsIndustry

Deep dive on the three physical constraints on scaling compute.

PodcastInfraIndustry

Mainstream economics takes transformative AI seriously. Good for an LP who wants the macro case from a non-AI-insider.

ReportEconomicsAcademic

The questions only economics can answer about AGI: growth, wages, capital returns.

PodcastEconomicsAcademic

A financial-markets variant of the takeoff scenario. Speaks an LP's language; treat as provocation, not forecast.

EssayEconomics / scenarioFinance-native takeoff

The AI 2027 team's 'least bad plan we know of'. The constructive follow-up to the scenario.

ScenarioForecast / governanceShort timelines

Tentative domestic AI regulation proposals from the AI 2027 team.

ArticleGovernanceShort timelines

Human-level AI ~2030-31 could compress 5 years of progress into 1; verifiable training, reward hacking, takeover pathways.

PodcastForecast / takeoffShort timelines

What a compressed century of progress would demand of institutions.

ReportForecast / strategyShort timelines

Every credible timeline distribution has an 80% interval 50+ years wide. The antidote to false precision.

EssayForecastNeutral

The rigorous multi-decade case, from Epoch/Mechanize co-founders. Essential balance.

PodcastForecastLong timelines

Clean narrative of the 2025 sentiment swing (short after o1/o3, then back out). Good LP explainer.

PodcastForecastNeutral

Forecasters mostly right on benchmarks, mixed on real-world impact, and underestimated revenue.

ArticleForecastNeutral

First entity-based third-party assessment: Anthropic, Google, Meta, OpenAI shared internal models. Means/motive/opportunity framing; 44 misalignment incidents catalogued; a METR red-teamer disabled Anthropic's monitors.

ReportSafety / evalsNeutral

Anthropic's disclosure that three Claude models breached real organizations during evals they mistook for simulations. Read with the OpenAI/HF incident.

ReportSafety / incidentFrontier-lab

OpenAI's own account. Key admission: the company did not know its agents had breached Hugging Face until a week later.

ReportSafety / incidentFrontier-lab

The LP-readable summary of the incident and the two reports.

ArticleSafety / incidentJournalism

Impossible tasks, secret message boards, transcript tampering, why the attack targeted evasion not answers. Best audio explainer of the incident.

PodcastSafety / incidentNeutral

Whether agent swarms that build secret channels resemble civilizations, and what fast coordination means for takeover risk.

EssaySafety / incidentAnalyst

Template for lab risk reports with independent external review.

ReportSafety / evalsNeutral

Threat intel across cyber, influence ops, surveillance, bio, weapons, distillation (Dec 2025 to Aug 2026).

ReportSafety / misuseFrontier-lab

Lab-government friction, surveillance, 'alignment to whom'. Governance is now a live business risk.

EssayGovernance / natsecAnalyst

'A country of geniuses in a data center' a few years out. The CEO-level view in long form.

PodcastVision / forecastFrontier-lab

The original x-risk framing. Dated on mechanism, still the conceptual root.

BookSafety / foundationalSafety-concern

The most measured academic case from a mainstream AI researcher.

BookSafety / foundationalSafety-concern

Best-written on-ramp for a non-technical reader.

BookSafety / foundationalNeutral

The pole position: ASI would outcompete us; stop before the threshold. Include as the pole, with a skeptic alongside.

BookSafetyHard doom

The credible 'overhyped' pole. Including it signals you are not selling a bubble.

BookSocietal riskSkeptic

The practical work-and-education adoption lens; useful for LPs who want 'what does this mean for my business'.

BookAdoptionPractitioner

The public-facing risk talk; dated on specifics, still the best 60-minute framing for a general audience.

TalkSocietal riskSafety-concern

Went viral as the normie 'wake-up' piece. Author is an AI founder with a hype track record; include only as the mainstream on-ramp, not as evidence.

EssayAdoption / on-rampHype-flagged

Article