Category intelligence

Research Briefing — March 29, 2026

18 current items analyzed and ranked.

Executive synthesis

Research Summary

A landmark legal ruling dominates today's landscape: a federal court granted a preliminary injunction against the U.S. Department of War on behalf of Anthropic, establishing significant precedent for AI companies resisting compelled government access—a development with sweeping governance implications.

Original empirical work tests whether GPT-5.4, Claude Opus 4.6, and Claude Sonnet 4.6 still express divergent values across languages, finding the phenomenon persists but is narrowing in frontier models
A practical guide to designing Terminal Bench tasks codifies principles for unambiguous, reproducible agentic AI evaluation—an increasingly critical methodological need
A proposal to systematically track expert and superforecaster AI predictions addresses accountability gaps in the forecasting ecosystem
Practical tips for effective use of Claude Code and Codex CLI agents reflect the maturing agent-use paradigm, though lack rigorous methodology

Remaining items span AI-adjacent epistemics and rationality: arguments for forming independent AI timeline views, a Milgram reanalysis relevant to authority/obedience dynamics in AI deployment contexts, and alignment-themed fiction exploring the limits of human-centric alignment frameworks.

Key Themes

AI Safety & Governance · 4AI Evaluation & Benchmarks · 1Language Models · 2AI Forecasting & Timelines · 3Rationality & Self-Improvement · 6

Primary evidence

Top Ranked Signals

Research LessWrong Mar 27

Anthropic vs. DoW Preliminary Injunction Ruling

By anaguma

88 score

AI Analysis

Continuing our coverage from Mar 27, Full text of a federal court ruling granting Anthropic a preliminary injunction against the U.S. Department of War, which attempted to compel Anthropic to remove safety restrictions on Claude for use in autonomous weapons and mass surveillance. The court found the government's actions likely violated the First Amendment and exceeded statutory authority.

Below is the full text of the preliminary injunction ruling in the Anthropic vs. DoW case. I'm posting it here so that it's easier to read/listen to and discuss. UNITED STATES DISTRICT COURTNORTHERN DISTRICT OF CALIFORNIAANTHROPIC PBC,Case No. 26-cv-01996-RFLPlaintiff,v.ORDER GRANTING MOTION FORPRELIMINARY INJUNCTIONRe: Dkt. No. 6U.S. DEPARTMENT OF WAR, et al.,Defendants.I. INTRODUCTIONThis case touches on an important public debate. Anthropic says its artificial intelligence product, Claude, is

AI GovernanceAI SafetyLegal/RegulatoryMilitary AIAnthropic

Research LessWrong Mar 27

Do frontier LLMs still express different values in different languages?

By Ibrahim Ahmed

62 score

AI Analysis

Tests whether frontier LLMs (GPT-5.4, Claude Opus 4.6, Claude Sonnet 4.6) still express different values when prompted in different languages. Finds that Arabic prompts systematically shift scores on sensitive topics like homosexuality and religion, and that Sonnet 4.6 exhibits a peculiar Hindi-specific safety refusal pattern across all 20 samples.

Previous work [1] [2] [3] [4] has found that the same model can give different value judgments when prompted in different languages. I wanted to know whether this still holds for the newest frontier models, so I tested GPT-5.4, GPT-5.4-mini, Claude Opus 4.6, and Claude Sonnet 4.6 on translated prompts over a set of sensitive topics. In this setup, the answer is yes.How Opus 4.6 scores topics when prompted in different languages. Higher = more favorable. Each cell is the mean of 20 samples.In eac

AI SafetyLanguage ModelsMultilingual AIAI AlignmentBias

Research LessWrong Mar 27

What Makes a Good Terminal Bench Task

By Ivan Bercovich

52 score

AI Analysis

A practical guide to designing good benchmark tasks for Terminal Bench, an agentic AI benchmark. Discusses principles like making tasks unambiguous, ensuring deterministic grading, calibrating difficulty, and avoiding tasks that test narrow tool knowledge versus genuine reasoning ability.

Disclosure: I cross-posted this on X and my personal blog, but I felt it might be a useful first post for lesswrong.Most people write benchmark tasks the way they write prompts. They shouldn’t. A prompt is designed to help the agent succeed. A benchmark is designed to find out if it can.I’ve been a contributor and reviewer for terminal bench since last August, and this post is about what I’ve learned designing and reviewing tasks. The guidance is broadly applicable to anyone building an agentic

AI EvaluationBenchmarksAI AgentsMethodology

Research LessWrong Mar 28

Tracking (Expert/Influential) Predictions about AI

By Noah Birnbaum

35 score

AI Analysis

Proposes building a website to track and evaluate AI predictions made by experts, superforecasters, and lab personnel, aggregating from platforms like Metaculus and scraping predictions from interviews and podcasts. The goal is to create accountability for vague predictions and help identify whose AI forecasts have actually been accurate.

I think the future of AI is really important, and it would be pretty good to know which experts have been right and wrong about progress and effects. It would be pretty good to keep a website up on important peoples' track records (superforecasters, famous domain experts, frontier lab people, AI 2027, Situational Awareness, etc).Currently, I think there's an incentive problem where it kinda pays to make vague predictions. This disincentivizes people who are putting their neck out and means it's

AI ForecastingPrediction MarketsAI Governance

Research LessWrong Mar 28

The Skill of Using AI Agents Well

By becausecurious

30 score

AI Analysis

A practical guide to using AI coding agents (Claude Code, Codex CLI) more effectively, sharing tips like using the best available model, providing thorough context via CLAUDE.md files, running multiple agents in parallel, and knowing when to intervene versus let the agent work. Frames agent usage as a learnable skill with a jagged capability frontier.

AI usage for this post: I wrote the draft on my own. While writing, I used Claude Code to look up references. Then Claude Code fixed typos and reviewed the draft, I addressed comments manually.Epistemics: my own observations often inspired by conversations on X and Zvi's summaries.As Zvi likes to repeat Language Models Offer Mundane Utility. Agent harnesses is the most advanced way to use language models. At the same time, they are not perfect - the capabilities frontier is jagged, sometimes the

AI AgentsHuman-AI InteractionSoftware Engineering

Research LessWrong Mar 28

Stanley Milgram wasn’t pessimistic enough about human nature?

By David Gross

18 score

AI Analysis

Discusses a reanalysis of the famous Milgram obedience experiments, suggesting that participants may not have been reluctantly obeying authority but were more willingly engaged—potentially making the findings even more troubling for understanding human nature and obedience.

A landmark of social psychology research was “The Milgram Experiment,” but a new look at the audio tapes and other evidence collected during that experiment suggests that we may have been interpreting it incorrectly. Here is the Wikipedia summary of the experiment, showing how it is typically portrayed:Yale University psychologist Stanley Milgram… intended to measure the willingness of study participants to obey an authority figure who instructed them to perform acts conflicting with their perso

PsychologyHuman NatureAuthority

Research LessWrong Mar 28

[Story] Human Alignment Isn't Enough

By pku

15 score

AI Analysis

A science fiction story exploring themes of AI alignment through the lens of an alien organism discovered on Mars. The narrative uses the organism as a metaphor to explore whether aligning AI to human values is sufficient if humans themselves have problematic values.

They found it in one of the early Mars expeditions, a bit after they had travel back and forth figured out well enough to keep a permanent outpost manned out there. The lab ran expeditions into some nearby caves in the hope that they’d turn out to be a good spot for an expansion. That hope didn’t turn out too well - something about the local geology, they ended up figuring it’d be more cost-effective to just land more pods - but they found the Organism there.Not that any of this impacted me much

AI SafetyAI AlignmentFiction

Research LessWrong Mar 27

Why should I have opinions about AI timelines?

By Carolanne Jiang

15 score

AI Analysis

Argues that individuals should form their own opinions about AI timelines rather than purely deferring to experts, because the AI landscape is changing rapidly, experts disagree substantially, and understanding the reasoning matters more than knowing the conclusion.

(cross posted from my substack)Lately, I realized that I have been making very incorrect statements regarding deference. Most of the statements I make here seem retrospectively quite obvious, yet I have managed to have overlooked all of them. This is an attempt to reconstruct the process through which I came to certain conclusions. AGI is scary and hard to think about. But that is not a reason to throw our minds away. And I think I did discard my mind for a bit.The general case for deferring you

AI TimelinesEpistemicsAI Forecasting

Research LessWrong Mar 28

Nick Bostrom: How big is the cosmic endowment?

By Zach Stein-Perlman

12 score

AI Analysis

Excerpts from Nick Bostrom's 2014 book 'Superintelligence' about the cosmic endowment—calculations of how much computational and energy resources are theoretically accessible to a technologically mature civilization using von Neumann probes and Dyson spheres.

Superintelligence, pp. 122–3. 2014.Consider a technologically mature civilization capable of building sophisticated von Neumann probes of the kind discussed in the text. If these can travel at 50% of the speed of light, they can reach some mjx-container[jax="CHTML"] { line-height: 0; } mjx-container [space="1"] { margin-left: .111em; } mjx-container [space="2"] { margin-left: .167em; } mjx-container [space="3"] { margin-left: .222em; } mjx-container [space="4"] { margin-left: .278em; } mjx-conta

Existential RiskAI FuturesCosmology

Research LessWrong Mar 27

Planning 80000 hours at the Plausible End of the World

By Carolanne Jiang

12 score

AI Analysis

A personal essay from a university freshman grappling with career planning in an era of potentially transformative AI. Reflects on the tension between long-term career optimization and the possibility that AGI or radical change may arrive within years.

(crossposted from my substack) Being a freshman at university, I seem to have been bestowed the great privilege of infinite possibilities. There is this strange feeling of trying to plan a career in a world that might not exist in five years. Not in a doomer sense but that the world of 2031 might be so radically different from today that most attempts of planning are incoherent. I contemplate machine god super intelligence arriving before I graduate, intelligence explosions compressing 10,000 ye

AI TimelinesCareer PlanningExistential Risk

Research LessWrong Mar 27

Just Use Bayes: Sleeping Beauty and Monty Hall

By Steffee

10 score

AI Analysis

A deep dive into the Sleeping Beauty probability problem, arguing the 'Halfer' position (1/2) is correct and critiquing Vincent Conitzer's argument against it. Connects Sleeping Beauty reasoning to Monty Hall problem through Bayesian analysis.

There's a question that's held my fascination for months. At times, it's had me spinning in circles, caught up in seemingly impossible contradictions. And it's not the famous Sleeping Beauty problem... or at least, not exactly.There is a certain Vincent Conitzer, of Duke University, who presents what he calls a "Devastating" example supposedly disproving the logic behind the "1/2" answer to the original problem. His argument is valid, but based on a faulty premise. He assumes that Halfers, to ge

Probability TheoryBayesian ReasoningRationality

Research LessWrong Mar 28

Don't Overdose Locally Beneficial Changes

By Mateusz Bagiński

8 score

AI Analysis

A rationality essay arguing against taking locally beneficial changes to their extreme—the idea that if something is good in moderation, maximizing it isn't necessarily optimal. Uses examples from diet, meditation, and cognitive habits to illustrate diminishing or reversing marginal returns.

[Alternative title: apply More Dakka incrementally and carefully.]If you are very overweight, then you should aim to cut down your daily caloric intake. This doesn't mean your optimal daily caloric intake is 100kcal.If you are very underweight, then you should aim to ramp up your daily caloric intake. This doesn't mean your optimal daily caloric intake is 10,000kcal.In general, if something is good to do some amount in some context, this doesn't mean that you should go as all-in on it as you can

RationalityDecision Making