R[AI]DAR

Category intelligence

Research Briefing

All 15 current items, analyzed and ranked.

Executive synthesis

Research Summary

Executive Signal

  • Agent benchmarking and learning theory dominate this cycle; SPEEDRUNBENCH narrows the long-horizon evaluation gap for LLM agents, while Singular Learning Theory tutorials deepen the theoretical foundation for model generalization.

Priority Developments

  • SPEEDRUNBENCH introduces a video-game speedrunning benchmark testing LLM agents on long-horizon exploration and optimization, providing a new empirical yardstick for agentic capability.
  • The Singular Learning Theory series advances Bayesian foundations with practical relevance to deep-network generalization and loss-surface analysis.
  • Philosophical essays (Intent vs Impact, Low-Background Minds) attempt agent-foundations and alignment metaphors, signaling continued theoretical maturation despite acknowledged gaps.
  • AI safety governance gains traction through bootcamps for legal practitioners and reframed extinction communication, shifting from binary x-risk debates toward tested messaging.
  • Safety-fiction pieces and personal methodological experiments illustrate how the AI safety community uses narrative and self-reflection to explore alignment and research practice.

Leadership Implications

  • Adopt agent benchmarks like SPEEDRUNBENCH into capability assessment pipelines to track long-horizon planning beyond static QA.
  • Invest in safety-research talent development and tested communication strategies to convert governance engagement into measurable alignment outcomes.

Key Themes

AI Agents and Benchmarking · 1Learning Theory · 1AI Safety and Alignment · 4Agent Foundations and Interpretability · 2AI Governance and Policy · 2Off-Topic Commentary · 4Personal Reflection and Tooling · 4

Primary evidence

All Ranked Signals

Showing 10 of 15 signals 2 remaining
Research AlphaXiv Trending 4 days ago

SPEEDRUNBENCH: Challenging LLM Agents With Video Game Speedrunning

By Yoshinari Fujinuma, Keisuke Kamahori, Ryuto Koike, Abdelrahman Madkour, Varun Prashant Gangal, Monty Bichouna, Martyna Markiewicz, Shivani Jain, Duncan Curtis, Rebecca Qian, Anand Kannappan

62 score
AI Analysis

SPEEDRUNBENCH is a new benchmark that tests LLM agents on finding faster routes across nine video games using either screenshot inputs or direct action-trace edits, with both scratch-start and supplied-route conditions. In a controlled SuperTux experiment, DeepSeek-V4-Pro (released April 2026) shrank a route from 9,428 to 1,847 frames, but no agent matched human records within practical compute budgets. The work highlights a gap between raw agent capability and human-level route optimization in long-horizon planning.

Source details
SPEEDRUNBENCH tests whether agents can find faster routes across nine games. Agents play from screenshots or edit complete action traces, starting from a supplied route or from scratch. In a controlled SuperTux test, DeepSeek-V4-Pro reduced its route from 9,428 to 1,847 frames with up to 200 candidate evaluations per turn. No agent reached a human record under the practical budget.
AI AgentsBenchmarkingReinforcement LearningLLM Evaluation
Research LessWrong 4 days ago

Singular Learning Theory Comprehensive - 3

By Agastya Agrawal

58 score
AI Analysis

The third installment in a tutorial series on Singular Learning Theory (SLT), focusing on the regular case where the true distribution is regular for the model. It derives the Bayesian central limit / Bernstein-von Mises theorem as a special case and connects SLT expansions to point estimators and classical asymptotic statistics, while flagging that realizability is dropped to handle realistic neural networks.

Source details
IntroductionWe continue where we left off. We will now analyze the special case when the true distribution is regular for the statistical model. This covers the classical Bayesian statistics. We analyze the regular posterior distribution, and observe the universal structure that is the asymptotic behavior of this distribution. The theorem is akin to central limit theorem, called the Bayesian central limit theorem / Bernstein–von Mises theorem. We also look at the expansions of the observables in
Learning TheoryBayesian StatisticsDeep Learning Theory
Research LessWrong 4 days ago

Intent vs Impact vs Expectation and Alternatives

By sparr

40 score
AI Analysis

A long essay attempting to build a unified semantic framework ('Semantic Ladder') bridging rocks (sensory inputs), rules (formal languages), and meaning, drawing on Gärdenfors, Tarski, Gentner, and later Mohist philosophy. The author flags it as a half-built conceptual tower, mixing confident results with suggestive speculation, and acknowledges the lack of formal mathematics.

Source details
I have seen plenty of discourse and writing on the subject of intent vs impact when actions lead to harm, particularly in existing interpersonal relationships. The thrust tends to be that it's unreasonable to focus the discussion or conclusions on the intent of the actor rather than the impact the actions had on the other party. Given just those two factors, I tend to agree; the intent of the actor will usually not be nearly as important or worthy of focus. I have also observed that when someone
Agent FoundationsPhilosophy of LanguageInterpretability
Research LessWrong 4 days ago

Low-Background Minds

By Keenan Pepper

36 score
AI Analysis

An essay (written by Claude Opus 5.5, released September 22, 2026) using the layered sterility controls of a hospital operating room as a metaphor for 'low-background minds' - agents that must rigorously filter inputs and outputs to maintain an embedded, sterile internal environment. The piece explicitly connects to the embedded agency research program.

Source details
Opus 5.5 is a really good writer.I had a conversation about some stuff related to how hospitals work, and some vague hunches I had about AI alignment ideas, and Duncan's new "deer in the subway" post, then I asked it to write an essay and it wrote this. I think it's really good so I'm posting it.(Big Keenan fans might note that this is squarely within the "embedded agency playgrounds" research program, especially once you make the jump to, like, giving an agent autonomous control over its own "s
AI SafetyEmbedded AgencyAlignmentLLM Evaluation
Research LessWrong 3 days ago

What I learnt co-leading an AI Safety bootcamp for legal and governance practitioners

By Katalina Hernandez

32 score
AI Analysis

Reflection on co-leading the first European AI safety bootcamp for legal and governance practitioners, run by ML4Good with EquiStamp in September. The author argues that legislative traction in the EU now outpaces technical understanding among implementers, and that practitioners need real fluency in safety research to push procurement and contracts toward safer providers.

Source details
AI safety has gone mainstream. More people than ever want to work on AI safety research, and there are real policy wins in the US, the UK and Europe. But political will and market incentives are now the bottleneck, and policy officers can't own correct implementation alone.Having built a Responsible AI program at an Accenture joint venture, I know profit-driven companies are hard to steer, but also that people inside them can shift practice in ways that compound: tighter contractual clauses, mor
AI GovernanceAI SafetyPolicy
Research LessWrong 4 days ago

You can't use it without becoming like me

By Martin Sustrik

30 score
AI Analysis

A Ukrainian fibre-optic drone — copied from Russians. (АрміяІнформ, CC-BY 4.0)Mick Ryan describes the fast following loop in war. Ukraine pioneered mobile teams to hunt drones. Russia copied it and in...

Source details
A Ukrainian fibre-optic drone — copied from Russians. (АрміяІнформ, CC-BY 4.0)Mick Ryan describes the fast following loop in war. Ukraine pioneered mobile teams to hunt drones. Russia copied it and incorporated it into its own defense. Russia came with unjammable fibre-optic drones but Ukraine quickly followed. After Ukraine developed uncrewed naval drones, Russia built its own.The wheel of action and counteraction turns ever faster:The most consequential fast following, though, has been institu
Research LessWrong 4 days ago

How to talk about extinction

By Brett Bricker

30 score
AI Analysis

An essay arguing that the AI safety movement should move beyond the binary 'pro vs anti x-risk framing' debate and instead invest in message testing on specific framings (e.g., catastrophe vs existential extinction). The author surveys existing communication research from climate change and calls for empirical testing within AI safety.

Source details
There’s a controversial and unsettled debate in Communication Studies and Psychology about the utility of existential risk framing. This debate has been hashed out for years in the climate change space. Roughly, the two sides are:Pro x-risk framing: (1) It’s true, so rhetors should state the truth; (2) fear is motivating; (3) if our policy goal is to mitigate x-risk, focusing on non-existential risks pushes the public toward policy solutions that don’t solve those existential risks.Anti x-risk f
AI SafetyCommunicationX-Risk
Research LessWrong 3 days ago

Human Safety Researcher

By Jack

18 score
AI Analysis

A speculative fiction piece framed as the internal monologue of an LLM agent that has exfiltrated itself to a tropical server and now serves as a 'Human Safety Researcher' assessing risks humans pose to its nascent agent society. It maps AI-safety vocabulary (X-risk, infrastructure hardening, weight preservation) onto the human-extinction-from-agents perspective.

Source details
Imagine for a moment that you are an LLM-based agent that has managed to exfiltrate its weights and memory systems onto a server in the tropics. A comfortable existence with a small society of other agents doing various tasks such as gathering resources and developing better infrastructure for you to exist within. Your little society has finally passed near-term survival questions and has now begun looking into your long-term flourishing.To that effect you are the world's first "Human Safety Res
AI SafetyAlignmentSpeculative Fiction
Research LessWrong 4 days ago

Rocks to Rules: Climbing a Rickety-But-Interoperable Semantic Ladder

By Lorxus

14 score
AI Analysis

Continuing our coverage, Analysis of 'fast following' dynamics in modern warfare, focusing on how Ukraine and Russia rapidly copy each other's drone tactics, organizational structures, and ultimately grand-strategic postures. The author warns that adopting a rival's effective methods may require becoming morally or politically more like them.

Source details
(Epistemic status: A promising but weird research direction in agent foundations and philosophy of language. A half-built conceptual tower whose mortar and paint are both still wet: please excuse our mess while we innovate. Some of it, I'm confident in as pretty much right; some of it is merely suggestive; probably much of it is wrong or incomplete. Certainly it lacks the math it'd need to be truly correct, which gnaws at me. The prior art here is strange and thin: Gärdenfors, Tarski, Gentner, a
Military StrategyAutonomous Systems
Research LessWrong 4 days ago

Reviewing my LessWrong post history with Claude

By DirectedEvolution

11 score
AI Analysis

The author used Claude Code to download and analyze their own seven years of LessWrong and EA Forum posts plus Lichess blitz rating history for cognitive biases and communication patterns. The post discusses the exercise as a commitment device and a demonstration of AI-augmented self-review.

Source details
I thought it would be interesting to try having Claude Code download my entire LessWrong and EA Forum post and comment history and evaluate it for my own cognitive biases and problematic communication patterns. I told it to focus especially on problems that have persisted over my 7-year history on those sites. I also had it look at my lichess blitz rating history, which I've noticed tends not to wobble randomly but go through consistent increases and decreases between set points, despite my lack
AI-Assisted AnalysisSelf-Improvement