R[AI]DAR

Category intelligence

Research Briefing

All 23 current items, analyzed and ranked.

Executive synthesis

Research Summary

Today's standout work clusters around safety/interpretability, governance, and a few niche applied items. Empirical results from a major lab lead the field.

Safety, Alignment & Interpretability

Governance & Policy

Applied ML & Discourse

Key Themes

AI Safety & Alignment · 6Interpretability · 3AI Governance & Policy · 5Applied ML · 2AI Consciousness & Welfare · 3Rationality & Philosophy · 8

Primary evidence

All Ranked Signals

Showing 10 of 23 signals 2 remaining
Research LessWrong Jun 13

SFT Drives Gemini’s Safety Properties

By Josh Engels

70 score
AI Analysis

A Google DeepMind interpretability team update reporting that most safety-relevant properties of Gemini appear to arise from pretraining plus supervised fine-tuning rather than RL stages. They ran SFT on pretraining-only versions of Gemini 3.1 Pro and Gemini 3 Flash and found post-SFT models matched production models across safety benchmarks. This is an empirical, lab-internal finding with implications for where safety training effort should focus.

Source details
This is the third in a series of informal research updates from the Google DeepMind Language Model Interpretability team, in interpretability and adjacent areas. The second post can be found here.In this short post, we describe a surprising finding: most safety relevant properties in Gemini seem to be caused by the combination of pretraining and SFT, not other training stages like RL. We do not want to overstate this claim as applying to other model families, and we also note that this may chang
InterpretabilityAI Safety & AlignmentLanguage ModelsTraining Methods
Research LessWrong Jun 13

How might continual learning affect safety and alignment?

By Rauno Arike

58 score
AI Analysis

Part of a sequence analyzing how continual learning could reshape LLM agent safety, arguing it may allow post-deployment goal and value change and erodes the last-mover advantage of current safety interventions. It maps three pathways for value drift and three ways safety measures (pre-deployment evals, data filtering, control protocols) could weaken. A structured conceptual contribution to alignment thinking.

Source details
This is the third post in our sequence Implications of Continual Learning for LLM Agents.SummaryWe argue that continual learning (CL) has two major potential safety implications: it may enable changes to LLM goals and values after deployment, and it eliminates the last-mover advantage held by current safety interventions.We identify three pathways for goal and value change during deployment. First, loss of developer-side control over generalization. Second, value systematization, induced when an
AI Safety & AlignmentContinual LearningAI AgentsAI Control
Research LessWrong Jun 13

A cheap specialist judge gets used by agents but fails to reduce alignment audit costs

By burnssa

55 score
AI Analysis

An empirical study giving AuditBench investigator agents a lightweight Gemma 2-2B toxicity/EM judge to test whether a cheap specialist tool reduces alignment audit costs. The agents reliably used the judge, but it only helped on quirks matching its training distribution and failed to lower total spend since the Sonnet driver dominated costs. A useful negative result for the alignment-auditing tooling community.

Source details
TL;DRI gave AuditBench's investigator agents a lightweight (Gemma 2-2B) EM-toxicity-scorer (judge) as an additional audit tool, targeting a proof-of-concept for misalignment detection at low cost, looking to validate that a specialized judge would (1) get used by these investigator agents, (2) help audits, (3) reduce overall evaluation spendValidation results:(1) Yes - the judge was used in every audit run. ~7 calls/audit without a prompt-given mandate, ~16 per audit with a mandate(2) Partial no
AI Safety & AlignmentAlignment AuditingAI AgentsLanguage Models
Research LessWrong Jun 13

American Government Takes Down Claude Fable

By Zvi

46 score
AI Analysis

Zvi's commentary on a US Commerce Department export-control directive that suspended access to Anthropic's Fable 5 and Mythos 5 models for all foreign nationals, citing national security and a reported jailbreak. It analyzes whether this is targeted lawfare or broad national-security hawkishness and its implications for AI governance. The two affected models were released only days before this coverage date.

Source details
No good policy gets announced shortly after 5pm eastern on a Friday. Here we go again. The Once And Future Fable The United States Department of Commerce, as per a letter from Commerce Secretary Howard Lutnick, apparently in response to a narrow jailbreak identified by Amazon, has classified Fable 5 and Mythos 5 as being subject to US export controls. That explicitly means cutting off access to all ‘foreign nationals,’ even within the United States, even if they are Anthropic employees. Given An
AI Governance & PolicyExport ControlsAI Safety & Alignment
Research LessWrong Jun 12

Short Timelines Favor Control, Long Timelines Favor Infrastructure Security

By Jannis

45 score
AI Analysis

A strategy post arguing that AGI timelines do not uniformly change risk: longer timelines may reduce accidental misalignment but raise misuse and sabotage risks, so timeline length shifts which interventions have highest expected value. Drawing on the author's infrastructure-security background, it favors AI control under short timelines and infrastructure security under long ones. A conceptual contribution to safety prioritization.

Source details
TL:DR, A common assumption is that extending AGI timelines reduces risk straightforwardly by giving alignment researchers more time. I suspect the relationship is more complicated. Longer timelines may reduce accidental misalignment risk while simultaneously increasing risks from deliberate misuse and sabotage[1]. If so, extending timelines changes which interventions have the highest expected value rather than uniformly reducing risk.--My background is in vulnerability research and critical inf
AI Safety & AlignmentAI ControlAI SecurityStrategy & Forecasting
Research LessWrong Jun 12

US government directive to suspend access to Fable 5 and Mythos 5

By Capybasilisk

42 score
AI Analysis

A primary-source repost of Anthropic's statement that a US government national-security export-control directive forced an abrupt shutdown of Fable 5 and Mythos 5 for all foreign nationals, reportedly over a jailbreak method. Anthropic notes the demonstrated vulnerabilities were minor and discoverable by other public models. It documents a notable regulatory intervention in AI deployment.

Source details
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance. Access to all other Anthropic models will not be affected. We received the directive from the government today at
AI Governance & PolicyExport ControlsAI Security
Research LessWrong Jun 12

Exploration of a DNA Sequencing Basecaller using Activation Patching

By Madeleine L

40 score
AI Analysis

An undergraduate mech-interp project applying activation patching to a non-LLM model, a DNA sequencing basecaller, finding MLP dominance in early and late layers, mid-layer attention activity, and concentrated activation in specific heads. It is exploratory interpretability on a biology pipeline component with potential relevance to pathogen surveillance.

Source details
This write-up for an undergraduate project is my first LW post, made with the objective of a) gathering feedback on the project and post, if more experienced authors are willing, and b) sending out results of a mech-interp-on-a-non-LLM (specifically, DNA basecaller) exploration in case the idea is interesting to anyone. Apologies in advance for any inconveniences and mistakes, and thank you in advance for your understanding.SummaryAs a first AI Safety/mech interp learning project, I tried applyi
InterpretabilityApplied MLComputational Biology
Research LessWrong Jun 13

AML for AI as a verification mechanism

By MarkelKori

36 score
AI Analysis

A proposal to build an anti-money-laundering-style monitoring system that tracks AI infrastructure nodes to detect large or hidden training runs, supporting future compute-governance agreements like FLOP thresholds. It sketches an idea and references existing efforts like datacenter modeling and procurement MCPs. Early-stage concept rather than implemented research.

Source details
The idea is to build a system that tracks key nodes in AI infrastructure in order to detect preparation for, or execution of, large training runs, and to monitor the overall situation more generally. In the future, if or when an international agreement limiting AI development appears — for example, via limits on FLOPs per training run ; the EU AI Act already uses a threshold of around 10²⁵ FLOPs for GPAI models with systemic risk[1], and providers are required to notify the AI Office without und
AI Governance & PolicyCompute GovernanceVerification Mechanisms
Research Machine learning : nature.com subject feeds Jun 13

Multi-layer feature aggregation network with residual module and attention mechanism for jaw cyst image segmentation

By Unknown

36 score
AI Analysis

A Nature Scientific Reports paper presenting a multi-layer feature aggregation network with residual modules and an attention mechanism for jaw cyst image segmentation. It targets improved medical image segmentation accuracy, though no abstract or details were provided in the source.

Source details
Applied MLMedical ImagingComputer VisionNeural Architectures
Research LessWrong Jun 13

The term “AGI” is almost useless at this point [Linkpost]

By Noosphere89

28 score
AI Analysis

A linkpost arguing the term AGI has become nearly useless because AI capabilities are highly jagged, so different AGI definitions that once correlated now diverge sharply. It highlights how AI now does real economic work while remaining uneven across tasks. Mostly conceptual commentary on terminology.

Source details
The reason I wanted to make this linkpost now rather than some other time is because discussions over AGI and whether or not LLMs are or aren't AGI are happening right now, and the point of the linkpost is that the term AGI is for our purposes useless at this point, because we are now in the fuzzy cloud now that AI can do real economic work.Some choice paragraphs:It used to seem possible that, in practice, the differences between these definitions might not matter all that much. If AI capabiliti
AI Concepts & ForecastingLanguage Models