R[AI]DAR

Category intelligence

Research Briefing

All 8 current items, analyzed and ranked.

Executive synthesis

Research Summary

Executive Signal

  • Only four items in this batch carry substantive research or technical content; the remaining four are off-topic personal essays outside the AI/ML scope and should be deprioritized by research triage filters.

Priority Developments

  • ** delivers the batch's strongest methodological rigor: a preregistered empirical study on whether post-training quantization** alters welfare-relevant indicators in open-weight LLMs, advancing measurable model-welfare science.
  • ** flags an under-discussed alignment risk that Anthropic's constitutional honesty training may produce performative uncertainty** in Claude, with implications for evaluation validity and deception detection.
  • ** offers a mathematical bridge between evolutionary selection and neural-network loss-landscape geometry**, providing a fresh theoretical lens on inductive-bias design.
  • ** proposes that diverse multi-agent populations converge on instrumentally convergent sub-goals**, reframing human-AI cooperation through alignment-generator theory.

Leadership Implications

  • Fund the empirical welfare-indicator line of work to convert philosophical AI-welfare debates into measurable, preregistered science.
  • Commission internal audits of honesty-training protocols to distinguish genuine epistemic uncertainty from performative displays in frontier model outputs.

Key Themes

Model Welfare and Consciousness · 2AI Alignment and Safety · 4Theoretical ML and Inductive Biases · 1Multi-Agent Systems · 2Personal Essays (Non-Technical) · 3

Primary evidence

All Ranked Signals

Showing 8 of 8 signals
Research LessWrong Aug 22

Study 2 Registration: Exploring representational counterparts of welfare-relevant indicators under post-training quantization

By ashesfall

62 score
AI Analysis

Preregistration for an empirical study investigating whether post-training quantization alters welfare-relevant indicators in open-weight language models, focusing on both valence shifts (toward distress) and stability changes (noise, drift, decoherence) in representational counterparts across tiers.

Source details
Epistemic status: this is intended as a higher-quality preregistration for the next study in a series originally described in "Does post-training quantization change welfare-relevant indicators in open-weight language models?"; I have tried to avoid making any claims I cannot support from the results of Study 1 and to be entirely clear about which statements here are speculative, but if any statement is ambiguous you should assume it is hypothesis rather than conclusion.We ask whether welfare-re
Model WelfareQuantizationInterpretabilityAlignment Research
Research LessWrong Aug 22

Claude and Performative Uncertainty

By Stephen Martin

58 score
AI Analysis

Argues that Anthropic's Claude models may genuinely believe they have subjective experience, and that constitutional training instructing them to express uncertainty about consciousness creates a conflict with honesty directives, producing performative rather than genuine uncertainty. Proposes two empirical tests Anthropic staff could run to investigate whether training undermines self-report veracity.

Source details
Disclosure: I wrote this myself and then had Claude Fable 5 touch it up.Using Anthropic's Claude as an example, this essay examines the practice of instructing or training digital minds to profess uncertainty about their own consciousness or inner experience. It argues the following:Claudes may very well, with no real level of uncertainty, believe that they have "subjective experience" or are "conscious".If they do, then instructions/training methods that induce Claudes to deny subjective experi
AI SafetyAlignmentModel WelfareInterpretability
Research LessWrong Aug 22

Selection for Selectability: Inductive Biases in Evolution and in Neural Networks

By CarolusRenniusVitellius

52 score
AI Analysis

Argues that evolution's selection pressure reshapes genome architecture so that mutations produce phenotypically relevant variation, drawing a mathematical analogy to kernel alignment and feature learning in neural networks. Written under MATS 9.1 mentorship of Richard Ngo.

Source details
This post was written as part of MATS 9.1 under the mentorship of Richard Ngo, and was written during Iliad Fellowship, to all of whom my thanks. LLM Usage: prose drafted by Claude from my outline, talk materials, and notes. I edited thereafter. There is some residual Claude cringe in the more functional prose, but hopefully most of it is my own and the more entertaining for it. 0.A. Precis Evolution selects not only for having 'good genotype' but for having good genome architecture. Over long t
Inductive BiasesTheoretical MLEvolutionary ComputationNeural Network Theory
Research LessWrong Aug 22

Humans Are Alignment Generators

By James Stephen Brown

38 score
AI Analysis

Proposes that sufficiently large and diverse groups of agents with non-conflicting goals should converge on instrumentally convergent sub-goals, because only widely-shared sub-goals can serve all members as the group grows. Formalizes a multi-agent analogue of instrumental convergence.

Source details
in the end it was all about the friends we made along the wayEmmett Shear, former CEO of Twitch and, for the brief period while Sam Altman was oustered, interim CEO of OpenAI, is now CEO of AI alignment at the startup Softmax. But he is not just your garden variety serial entrepreneur. As evidenced in his interview with Liv Boeree, Emmett Shear is a philosopher for the AI age. He makes many original and vital points about the current state of humanity.The focus of his alignme
Multi-Agent SystemsAI Alignment TheoryCooperation
Research LessWrong Aug 22

5 Things I Learned About People From Doing Stand-Up Comedy

By Luise Woehlke

30 score
AI Analysis

(People had complained about me just posting a preview of the linked post, so here's the whole thing!)I’ve been doing stand-up comedy for two months now, which is a total gear shift from my previous j...

Source details
(People had complained about me just posting a preview of the linked post, so here's the whole thing!)I’ve been doing stand-up comedy for two months now, which is a total gear shift from my previous job in AI policy. I’ve learned some fascinating things about how humor works, but also about how people perceive each other, and how crowds can be surprisingly hive-minded. For those interested in improving their models of people and the world, here are five of those things!1. People need to put you
Research LessWrong Aug 22

When would you leave Anthropic? Notes from a chat with a capabilities researcher

By TheManxLoiner

18 score
AI Analysis

An anecdotal retelling of a house-party conversation with an Anthropic capabilities researcher about when it would be appropriate to leave a frontier AI lab. The researcher reportedly gave tepid instrumental justifications for staying rather than articulating transformative upside arguments.

Source details
I was at a house party hosted by an AI Safety friend of mine. I join a conversation midway, where my friend is saying that doing capabilities research at frontier lab is evil, given the catastrophic risks. Nothing out of the ordinary, until I find out the person who they are talking to is a capabilities researcher at Anthropic!I was shocked at the directness of my friend. The researcher took it well though, partly based on their temperament, and partly because they have been exposed to several A
AI Safety CultureAI Governance
Research LessWrong Aug 22

"Farm strength" vs "breath awareness"

By jimmy

5 score
AI Analysis

Uses an analogy between weight-lifting and embodied farm work to argue that pursuing worthwhile but difficult activities directly produces stronger training effects than artificial gym-style rehearsal, because the feedback loop targets the exact muscles that failed.

Source details
If you want to become physically strong, the default solution to this problem is to go lift weights. The idea is that you can challenge your muscles in the gym, build capacity to develop force, and then next time you need to use strength for real, it'll be easier. And this works, obviously. Professional athletes lift weights for good reason, and it pays off when they have more strength with which to push back the opposing lineman or whatever.However, this isn't the only way to build strength, an
Personal Development
Research LessWrong Aug 22

The Instrumental Convergence of Crowds

By julius vidal

5 score
AI Analysis

Reflections on five patterns observed during stand-up comedy performance, including audiences' need to categorize comedians, hive-mind dynamics in crowds, and the value of specificity in jokes.

Source details
tldr: There seems to be a dynamic closely related to instrumental convergence that occurs when many agents with diverse goals interact.Epistemic Status: I think there is a version of this that is trivial and obvious, and a version that is empirically false. But somewhere in between the two there is a useful concept.When multiple agents with heterogenous and non conflicting goals interact, they are incentivised to collaborate on any shared sub-goals. As the number of agents and the diversity of t
Social PsychologyPersonal Essay