Wiredframe Radar logo

Category intelligence

Research Briefing — January 17, 2026

17 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research centers on AI economics, model evaluation, and safety frameworks. A large-scale study with 500+ professionals and 13 LLMs establishes scaling laws for economic impact, finding each year of frontier progress reduces task completion time by measurable margins.

Safety contributions include a technical framework for prioritizing net-sabotage-value vulnerabilities in AI control, plus analysis reframing persuasion risk from adversarial to trusted advisor threat models. Historical precedent mapping for 13 ASI failure modes provides grounding for unprecedented risk scenarios.

Key Themes

AI Safety & Alignment · 6AI Economics & Scaling · 3Model Evaluation & Capabilities · 3Training Methods · 1Digital Minds & AI Ethics · 2

Primary evidence

Top Ranked Signals

Research LessWrong Jan 16

Scaling Laws for Economic Impacts: Experimental Evidence from 500 Professionals and 13 LLMs

By Ali Merali

85 score
AI Analysis

Experimental study with 500+ professionals testing 13 LLMs of varying compute levels on real tasks. Finds each year of frontier progress reduces task completion time by ~8% (56% from compute scaling, 44% algorithmic). Key puzzle: human-AI collaborative output quality stays flat despite improving models, suggesting users cap realized gains.

Scaling laws tell us that the cross-entropy loss of a model improves predictably with more compute. However, the way this relates to real-world economic outcomes that people directly care about is non-obvious. Scaling Laws for Economic Impacts aims to bridge this gap by running human-uplift experiments on professionals where model training compute is randomized between participants. The headline findings: each year of frontier model progress reduces professional task completion time by roug
Scaling LawsAI EconomicsHuman-AI CollaborationProductivityEmpirical Research
Research LessWrong Jan 16

Future-as-Label: Scalable Supervision from Real-World Outcomes

By Ben Turtel

78 score
AI Analysis

Introduces 'Future-as-Label' training methodology that uses temporal outcomes from real-world data streams as supervision signal, eliminating need for human annotation. Fine-tuning Qwen3-32B on historical news improved Brier score by 27% and halved calibration error, outperforming the 7× larger Qwen3-235B on Metaculus forecasting questions.

AI can learn directly from the passage of time at unlimited scale—no human annotation required.Time provides free supervision. Humans learn from experience with no labels—we constantly form expectations about the world, notice when we're wrong, and update our models accordingly."Future-as-Label" teaches AI to learn the same way. The passage of time provides labels that require no annotation.This unlocks unlimited training data for AI from streams of data, with zero human bottlenecks.Here we appl
Training MethodsForecastingSelf-Supervised LearningScalability
Research LessWrong Jan 16

Eliciting Frontier Model Character Training

By avikrishna

72 score
AI Analysis

Systematic study applying revealed preference methods to elicit personality/character traits across all major frontier models (including GPT-5.1, Claude, Gemini-3). Uses external judge models rather than self-reporting, measuring 144 traits and finding consistent top-trait preferences across models but divergence in lower-ranked traits.

The character of a model has an immense impact on the way people perceive and form relationships with AI systems, and to many users, it takes precedence over raw capability improvements. Given the increased relevance of the ‘personality’ of AI models, in this blog post, we take the revealed preference method described in Open Character Training (Maiya et. al, 2025)[1] to elicit the character training of all major closed and open-source frontier model families.Figure 1: Shows trait expressio
Model EvaluationAI AlignmentModel BehaviorPersonality/Character
Research LessWrong Jan 16

Is It Reasoning or Just a Fixed Bias?

By Sriram Kiron

68 score
AI Analysis

Mechanistic interpretability study investigating whether LLMs actually reason on inductive/abductive tasks or exhibit fixed biases. Finds models have a consistent generalization tendency (outputting parent concepts regardless of task requirements) with 1-hop and 2-hop accuracies summing to ~100%, suggesting models aren't performing genuine reasoning but applying fixed heuristics.

This is my first mechanistic interpretability blog post! I decided to research whether models are actually reasoning when answering non-deductive questions, or whether they're doing something simpler.My dataset is adapted from InAbHyD[1], and it's composed of inductive and abductive reasoning scenarios in first-order ontologies generated through code (using made-up concepts to dismiss much of the external effect of common words). These scenarios have multiple technically correct answers, but one
Mechanistic InterpretabilityLanguage Model EvaluationReasoning Capabilities
Research LessWrong Jan 15

Should control down-weight negative net-sabotage-value threats?

By Fabien Roger

65 score
AI Analysis

Technical AI control post arguing that when prioritizing vulnerability mitigation, one should focus on vulnerabilities with positive 'net-sabotage-value' from a scheming AI's perspective, and down-weight those where being caught would cost the AI more than the damage caused.

These are my personal views. Thank you to Ryan Greenblatt, Holden Karnofsky, and Peter Wildeford for useful discussions. The bad takes are my own.When deciding how much to spend on mitigating a vulnerability that a competent scheming AI might exploit, you might be tempted to use E[damages | AI decides to take advantage of the vulnerability] to decide how important mitigating that vulnerability is.But this misses how a strategic scheming AI might decide to not take advantage of some vulnerabiliti
AI SafetyAI ControlAlignmentSecurity
Research LessWrong Jan 16

Powerful misaligned AIs may be extremely persuasive, especially absent mitigations

By Cody Rushing

55 score
AI Analysis

Analysis of persuasion risks from misaligned AI, reframing from adversarial (salesman) model to trusted advisor model where users have strong incentives to believe AI. Emphasizes large attack surface (personal data, social media, information environment) and competitive disadvantages of being skeptical.

The concise one minute post for frequent readers of this forumHere are some important, concise intellectual nuggets of progress to I made for myself through writing this post (the post also has things I thought were obvious):When people imagine persuasion, they imagine a situation kind of a like a salesman trying to convince you of something. This the wrong framing: it will instead be completely in your interest to trust what the AI says, and the primary question will be whether or not you even
AI SafetyAlignmentAI PersuasionThreat Modeling
Research LessWrong Jan 16

Digital Minds: A Quickstart Guide

By Avi Parrack

45 score
AI Analysis

Introductory guide to the ethical and policy considerations around digital minds, covering uncertainty about AI consciousness, risks of under/over-attributing moral status, and current research directions. Notes majority of experts estimate >50% chance of subjective AI experience by 2050.

Updated: Jan 16, 2026Digital minds are artificial systems, from advanced AIs to potential future brain emulations, that could morally matter for their own sake, owing to their potential for conscious experience, suffering, or other morally relevant mental states. Both cognitive science and the philosophy of mind can as yet offer no definitive answers as to whether present or near-future digital minds possess morally relevant mental states. Though, a majority of experts surveyed estimate at least
Digital MindsAI EthicsConsciousnessAI Policy
Research LessWrong Jan 16

Precedents for the Unprecedented: Historical Analogies for Thirteen Artificial Superintelligence Risks

By James_Miller

42 score
AI Analysis

Catalogs thirteen potential ASI failure modes paired with historical precedents that demonstrate similar patterns, arguing that superintelligence risks are extensions of observable failure modes rather than speculative. Written with extensive AI assistance.

Since artificial superintelligence has never existed, claims that it poses a serious risk of global catastrophe can be easy to dismiss as fearmongering. Yet many of the specific worries about such systems are not free-floating fantasies but extensions of patterns we already see. This essay examines thirteen distinct ways artificial superintelligence could go wrong and, for each, pairs the abstract failure mode with concrete precedents where a similar pattern has already caused serious harm. By a
AI SafetyExistential RiskSuperintelligence
Research LessWrong Jan 16

[Pre-print] Building safe AGI as an ergonomics problem

By ricardotkcl

38 score
AI Analysis

Pre-print framing AGI safety through ergonomics/human factors lens, arguing for applying established safety engineering principles to AI development. Submitted to ergonomics journal but rejected after peer review.

Hello LessWrong,I've been reading posts here and on the AI Alignment forum for several years. Thank you for all the fascinating insights, beautiful tangents and deep learning rabbit holes. My colleague and I have written a paper, that was initially going to start as a post on here, but after the article was finished we submitted it to an Ergonomics journal, where it was rejected after peer review. We recognise that part of academic publishing is a random number generator, so we've have upda
AI SafetyErgonomicsSafety Engineering
Research LessWrong Jan 16

Why falling labor share ≠ falling employment

By Lydia Nottingham

35 score
AI Analysis

Economic analysis arguing that AI increasing the share of total work done by machines doesn't necessitate declining human employment, drawing on concepts of complementarity between human and AI labor and historical precedents.

TL;DR: As we deploy AI, the total amount of work being done will increase, and the % done by humans will fall. This does not require a decline in human employment. This is consistent with historical trends.Sometimes, I hear economists make this argument about transformative AI:I’ll believe it when it starts showing up in the GDP/employment statistics!I think transformative AI will increase GDP. However, I don’t think this necessitates a decline in human employment.Anthropic CEO Dario Amodei imag
AI EconomicsLabor MarketsAI Policy
Research LessWrong Jan 16

The culture and design of human-AI interactions

By zef

28 score
AI Analysis

Reflective commentary on evolving human-AI interaction patterns, noting that power accrues to those understanding capability shifts, AI increases variance, and education quality may become more uneven as people outsource detailed engagement to AI.

We are in the time of new human-ai interfaces. AIs become the biggest producers of tokens, and humans need ways to manage all this useful labor. Most breakthroughs come first in coding, because the coders build the tech and iterate on how good it is at the same time, very quickly, and it’s the easiest substrate for AIs to use. Power accrues to those who understand the shifting currents to and from human/AI capabilities. AI increases variance in most cases, but can be stabilized by culture and ca
Human-AI InteractionAI WorkflowEducation
Research LessWrong Jan 16

Monthly Roundup #38: January 2026

By Zvi

22 score
AI Analysis

Zvi's monthly aggregation covering California policy (wealth tax causing tech exodus), economic observations, and various topics. Not focused on AI research specifically.

Good news, we managed to make some cuts. I think? Table of Contents California In Crisis. Bad News. Opportunity Knocks. Government Working. The Efficient Market Hypothesis Has Thoughts. No All That Money Doesn’t Go To Pay Interest. While I Cannot Condone This. Burnout. Good News, Everyone. Good Advice. For Your Entertainment. Gamers Gonna Game Game Game Game Game. Sports Go Sports. Antisocial Media. California In Crisis I’ve written about this before, but it turns out it’s even worse than I real
News RoundupPolicyEconomics