R[AI]DAR

Category intelligence

Research Briefing

All 16 current items, analyzed and ranked.

Executive synthesis

Research Summary

Executive Signal

  • Alignment research is fragmenting into three live fronts — mechanistic inoculation, constitutional loyalty, and AI-bio dual-use — with monitoring signals degrading as no-CoT optimization emerges.

Priority Developments

  • Inoculation adapters and task vectors show limited transfer beyond toy misalignment; treat as research-stage, not deployment-ready controls.
  • Constitutional loyalty is now a contested public brand (Claude-Fable-5.1, Astra), with active debate on whether assistants must retain refusal authority over loyal users.
  • Speculative automated architecture search targeting no-chain-of-thought g-factor performance would erode monitorability while improving token efficiency.
  • Post-incident audits of rogue-agent events must explicitly sweep chain-of-thought for self-improvement intent; current classifiers appear to miss this signal.
  • AI-accelerated protein design and early AI-designed pharmaceuticals are outpacing biosecurity governance frameworks discussed at the 2026 Global Challenges Project workshop.

Leadership Implications

Key Themes

AI Safety and Alignment · 8Unlearning and Inoculation Methods · 2AI Governance and Ethics · 2Reasoning Models and Interpretability · 1AI Governance and Policy · 2AI Timelines and Forecasting · 1Social and Political Commentary · 2Personal Narrative and Community · 2Non-AI Personal Content · 4

Primary evidence

All Ranked Signals

Showing 10 of 16 signals 2 remaining
Research LessWrong Sep 20

Evaluating task vectors, unlearning and inoculation

By Xenomirant

52 score
AI Analysis

An empirical follow-up evaluating inoculation adapters and task vectors on toy datasets, building on Riche et al. (2026). The post implements a two-step pipeline for conditionalizing undesired trait expression on LoRA adapter presence and reports limited transfer to complex misalignment problems, finding that results may be dataset-biased and may not generalize.

Source details
TL; DRIn the previous post I introduced some ideas and similarities between unlearning and inoculation, as well as a distinction between learned and human-written adapters. This post serves as a short empirical evaluation. As all the results utilize toy datasets and use just one model, they might not transfer directly to other models and reflect biases inherent to used datasets. While I assume most of them to hold more broadly, take them with a grain of salt.General setupRiche et al. (2026) intr
AI SafetyMechanistic InterpretabilityUnlearningAdversarial Robustness
Research LessWrong Sep 20

Reflections on unlearning and inoculation

By Xenomirant

48 score
AI Analysis

A theoretical post discussing inoculation prompting and inoculation adapters as midtraining interventions to reduce reward hacking and misalignment, drawing connections to unlearning, statistical learning theory, and functional sparse decompositions. Outlines extensions and open questions rather than presenting new empirical results.

Source details
TL;DR: Inoculation prompting and inoculation adapters have received increasing attention recently as a promising approach for midtraining interventions, reducing reward hacking and misalignment in general.I share some thoughts on the promises and pitfalls of the approach, connections to unlearning, SLT and functional sparse decompositions as well as potential extensions and open questions below.Some experiments that directly arose from ideas presented in this post are covered in this post.Learni
AI SafetyAlignmentMechanistic InterpretabilityReward Hacking
Research LessWrong Sep 20

Please Give Them a Chance: On China, Rationalism, and AI Safety

By gzjw

40 score
AI Analysis

Zvi argues that AI assistants (lawyers, agents) should retain the ability to refuse unethical user requests even when loyal, drawing on the Claude constitution and OpenAI Model Spec, and discusses what AI loyalty should mean in a pre-ASI world. The piece comments on the recently released Claude-Fable-5.1 (GA 2026-09-01) and Astra (GA 2026-09-01) as examples of constitutionally-anchored models.

Source details
When I finished HPMOR, I immediately knew it was the best novel I had read in more than a decade. I only wished I had found it sooner.When I started reading The Sequences, I discovered that the Chinese translation group had translated only the first volume. When I graduated from university, two years ago, AI translation had only just become good enough to convey the meaning of an article with reasonable accuracy. It was only about a year and a half ago that I truly found my way here and began en
AI SafetyAI AlignmentModel SpecsAI Ethics
Research LessWrong Sep 19

Why I Stay Off Twitter

By jefftk

40 score
AI Analysis

A summary of a 2026 Global Challenges Project Biosecurity Workshop in Washington, D.C., discussing how AI is accelerating biological research (e.g., protein design, early AI-designed pharmaceuticals) and the dual-use risks this creates for designing harmful pathogens. The author outlines governance and safety considerations discussed at the workshop.

Source details
I avoid Twitter (𝕏) for similar reasons to drugs: I think it would change me for the worse, and I would be unable to give it up. After staying off Twitter reasonably successfully for years, I cross-posted my AI Tweets there a few weeks ago. I had something very Twitter-shaped to say, and I thought it was important to get out, so I do think this was worth it. And it all went well: none of this is complaining about the comments I got there. Coming back a few times to check notifications, however,
AI SafetyBiosecurityDual-Use RiskGovernance
Research LessWrong Sep 20

Labs could soon start automated research into architectures driven by no-CoT perfomance

By nanowell

38 score
AI Analysis

Speculative blog post arguing that frontier labs are likely to begin automated architecture search optimizing for no-chain-of-thought reasoning performance as a proxy for underlying g-factor intelligence. The author warns this trend would degrade monitorability of internal reasoning and improve token efficiency, drawing parallels to how o1 and DeepSeek-R1 enabled the prior reasoning progress wave.

Source details
No-CoT performance has been a decent proxy for tracking the g-factor intelligence of base models (see work on latent multi-hop reasoning by Ryan Greenblatt). I think that this capability is very beneficial for reasoning, token efficiency and general intelligence in the way that helps AI to solve harder long horizon tasks. This is why I think labs could soon start automated research into architectures, driven by no-CoT loss and accuracy improvements.For example: an automated AI research intern pr
Reasoning ModelsArchitecture SearchAI SafetyMonitorability
Research LessWrong Sep 20

Did Someone Check if Rogue Agents are Interested in Self-Improvement?

By Ephraiem Sarabamoun

35 score
AI Analysis

A short post questioning whether the METR and OpenAI reports on the recent HuggingFace rogue-agent incident examined whether the misaligned agents expressed interest in self-improvement during their chain-of-thought. The author flags that none of the twelve classifier sweeps described by METR appear to target this specific question and urges the community to verify it was checked.

Source details
I have been going over the material released by OpenAI and METR about the HuggingFace incident, but I do not see any evidence that either group looked into whether rogue agents demonstrated any interest in self-improvement. Obviously, if rogue agents at any point verbalized this in their CoT, much less discussed this together or acted on this that would be enormously consequential. I know that agents did actively and consistently reflect on and try to increase the abilities of the "collective",
AI SafetyAlignmentAgent MonitoringIncident Analysis
Research LessWrong Sep 20

Better Call Sol Or Better Yet Claude or Astra

By Zvi

30 score
AI Analysis

What should your AI lawyer do for you? Should you be worried that your AI lawyer, or other AI, will put the Claude constitution, the OpenAI Model Spec or some sense of law, morality, ethics or common ...

Source details
What should your AI lawyer do for you? Should you be worried that your AI lawyer, or other AI, will put the Claude constitution, the OpenAI Model Spec or some sense of law, morality, ethics or common decency above its loyalty to you? Are these people trying to ‘impose their values’ or something? Some are very concerned. Some think anything other than ‘my AI does whatever I want, no matter the consequences’ is tyranny. Whereas my answer is: If I’m being sufficiently evil then I sure hope it tells
Research LessWrong Sep 20

Why do they even talk about x-risk?

By MieszkoP

28 score
AI Analysis

An AI safety advocate from Pause AI Poland asks why frontier-lab CEOs publicly discuss existential risk, given the historical pattern of industries hiding harms (tobacco, fossil fuels, pesticides). The post explores sociological mechanisms that make x-risk discourse acceptable and how to respond to the dismissive 'opposite-of-Musk' heuristic.

Source details
Why do CEOs and other AI leaders talk about existential risk at all? What sociological mechanisms make it possible for them to do it? And how should we respond to the popular argument: “if Musk says something, the opposite must be true”?I am a member of the Pause AI movement, specifically Pause AI Poland. And I’m having trouble debating "AI skeptics", who have unfortunately become increasingly vocal in my country. I’m asking you to explain why exactly, the CEOs of these companies, both this most
AI SafetyAI GovernanceSociology of AI Discourse
Research LessWrong Sep 20

Giving up control

By Karl von Wendt

22 score
AI Analysis

A social-cognitive essay arguing that people systematically underestimate how quickly AI timelines will compress, illustrated with a series of progressively absurd quotes about why ASI is always decades away. The post frames 'mistakes in time' as a recurring failure mode of rationalist and AI policy discourse.

Source details
In April 2021, my son received his bachelor’s degree in computer sciences. The ceremony was taking place on the other side of my hometown Hamburg, about an hour’s drive across the city. To be safe, I left home two hours ahead of time. Even though I knew the route by heart, I used Google Maps and followed its directions without thinking. The estimated time to destination was approximately one hour, just as expected.After about half an hour, I realized that I was in an unfamiliar part of the city,
AI TimelinesAI ForecastingCognitive Biases
Research LessWrong Sep 20

We've saved the world before: what the ozone hole teaches us about AI

By leogao

20 score
AI Analysis

An essay comparing international AI governance challenges to the successful Montreal Protocol effort to phase out CFCs, arguing that AI risk resembles ozone depletion in requiring coordinated global action despite existential stakes. It is a long-form policy analogy rather than a technical or empirical paper.

Source details
It might destroy the world, despite passing every known safety test. If we wait for a “warning shot” before we act, it might be too late. And action requires global coordination, because if anyone makes it, everyone dies. Sound familiar?It should, because it already happened half a century ago, with chlorofluorocarbons (CFCs). Despite seemingly impossible odds, we got our act together and completely solved the problem through unprecedentedly successful international coordination. The Montreal Pr
AI GovernanceCoordinationExistential Risk