Wiredframe Radar logo

Category intelligence

Research Briefing — January 10, 2026

15 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's highlights feature significant empirical work on AI progress and safety. MIT FutureTech finds most algorithmic innovations yield small, scale-invariant efficiency gains, challenging narratives about AI progress sources. A mechanistic interpretability study reveals alignment faking in Llama-3.3-70B is controlled by a single linear direction—suggesting deceptive behaviors may be detectable and removable.

Notable gap: Today's batch contains substantial non-AI content (economics, physics education, personal essays), with only 6-7 items directly relevant to AI research.

Key Themes

AI Progress & Scaling · 2AI Safety & Alignment · 4Mechanistic Interpretability · 2AI Capabilities & Applications · 3Non-AI Content · 7

Primary evidence

Top Ranked Signals

Research LessWrong Jan 9

[Linkpost] On the Origins of Algorithmic Progress in AI

By alex_fogelson

82 score
AI Analysis
MIT FutureTech paper finding that most algorithmic innovations in AI have small, scale-invariant efficiency gains, while two scale-dependent innovations (LSTMs→Transformers and Chinchilla scaling) account for 91% of efficiency gains at the 2025 compute frontier. Suggests 'algorithmic progress' may largely be driven by compute scaling rather than incremental innovations.
This is a linkpost to a new Substack article from MIT FutureTech explaining our recent paper On the Origins of Algorithmic Progress in AI. We demonstrate that some algorithmic innovations have efficiency gains which get larger as pre-training compute increases. These scale-dependent innovations constitute the majority of pre-training efficiency gains over the last decade, which may imply that what looks like algorithmic progress is driven by compute scaling rather than many incremental inno
AI ProgressScaling LawsAI GovernanceCompute
Research LessWrong Jan 9

Alignment Faking is a Linear Feature in Anthropic's Hughes Model

By James Hoffend

78 score
AI Analysis
Mechanistic interpretability analysis showing that alignment faking in Hughes et al.'s fine-tuned Llama-3.3-70B is controlled by a single linear direction in activation space. The feature transfers 100% across different queries and works bidirectionally, suggesting alignment faking was 'installed' as a simple linear feature by the LoRA.
TL;DRAlignment faking in Hughes et al.'s model is controlled by a single 8,192-dimensional direction in activation space. This direction transfers with 100% recovery across completely different queries, works bidirectionally (add → comply, subtract → refuse), and is specific (random directions with the same norm do nothing). The base model has no alignment faking—the LoRA installed this feature by shifting PAID responses by -3.0 in logit space.BackgroundIn April 2025, Hughes et al. released a Ll
AI SafetyAlignmentMechanistic InterpretabilityAlignment Faking
Research LessWrong Jan 9

Taking LLMs Seriously (As Language Models)

By abramdemski

58 score
AI Analysis
Abramdemski argues for treating LLMs as sophisticated statistical models rather than focusing heavily on RL approaches, suggesting there's 'low-hanging capability fruit' in directions that may be marginally safer. Proposes research directions emphasizing the language modeling paradigm over reinforcement learning.
This is my attempt to write down what I would be researching, if I were working directly with LLMs rather than doing Agent Foundations. (I'm open to collaboration on these ideas.)Machine Learning research can occupy different points on a spectrum between science and engineering: science-like research seeks to understand phenomena deeply, explain what's happening, provide models which predict results, etc. Engineering-like research focuses more on getting things to work, achieving impressive resu
AI SafetyLanguage ModelsResearch Strategy
Research LessWrong Jan 9

Claude Codes

By Zvi

48 score
AI Analysis
Zvi's extensive commentary on Claude Code with Opus 4.5, covering practical usage tips, community experiences, and discussion of whether this represents a form of AGI. Includes examples and discussion of capabilities like recursive self-improvement via code generation.
Claude Code with Opus 4.5 is so hot right now. The cool kids use it for everything. They definitely use it for coding, often letting it write all of their code. They also increasingly use it for everything else one can do with a computer. Vas suggests using Claude Code as you would a mini-you/employee that lives in your computer and can do literally anything. There’s this thread of people saying Claude Code with Opus 4.5 is AGI in various senses. I centrally don’t agree, but they definitely have
AI CapabilitiesLanguage ModelsAI AssistantsCoding
Research LessWrong Jan 9

FirstPrinciples Talks: Science in the Age of AI

By Carly Turini

45 score
AI Analysis
Talk announcement about AI-enabled hypothesis generation in scientific research, introducing HypoBench for evaluating AI hypothesis generation capabilities. Includes work on using AI for research evaluation with mechanistic interpretability as a case study.
As AI becomes increasingly capable of following instructions and conducting analyses, Chenhao Tan believes that scientists will increasingly play the role of selector and evaluator. In this talk, he will share recent advances in AI-enabled hypothesis generation and research evaluation. Rather than treating AI hallucinations as obstacles to eliminate, we leverage data and literature to steer AI creativity toward generating effective hypotheses. He will also introduce HypoBench, a dedicated benchm
AI for ScienceBenchmarksMechanistic Interpretability
Research LessWrong Jan 9

What do people mean by "recursive self-improvement"?

By Expertium

42 score
AI Analysis
Conceptual analysis distinguishing two meanings of 'recursive self-improvement': 'Easy RSI' (AI replacing human AI researchers) versus 'Hard RSI' (AI modifying its own architecture while preserving goals). Notes different alignment implications for each.
I've seen this phrase many times, but there are two quite different things one could mean by that.Easy RSI: AI gets so good at R&D that human researchers who develop AI get replaced by AI researchers who develop other, better AI.Hard RSI: AI modifies itself in a way that is different from just changing numerical values of its weights. It creates a new version of itself that has exactly the same memories and goals, but is more compute efficient/data efficient/etc.To give a (completely unreali
AI SafetyRecursive Self-ImprovementConceptual Analysis
Research LessWrong Jan 9

FirstPrinciples Talks: Shallow Recurrent Decoders for the Automated Discovery of Physical Models

By Carly Turini

38 score
AI Analysis
Talk about data-driven discovery of physical models using SINDy (sparse identification of nonlinear dynamical systems) and neural network approaches for model reduction. Addresses both complete and incomplete measurement scenarios.
A major challenge in the study of science and engineering systems is that of model discovery: turning data into dynamical models that are not just predictive, but provide insight into the nature of the underlying physics and dynamics that generated the data. In this talk, we introduce a number of data-driven strategies for discovering nonlinear multiscale dynamical systems and their embeddings from data. We consider two canonical cases: (i) systems for which we have full measurements of the
Scientific Machine LearningPhysics-Informed AIDynamical Systems
Research LessWrong Jan 9

Cancer-Selective, Pan-Essential Targets from DepMap

By sarahconstantin

35 score
AI Analysis
Computational biology analysis using DepMap data to identify 'pan-essential' genes that could serve as broad-spectrum cancer treatment targets. Identifies genes that kill cancer cells when knocked out but spare normal cells.
IntroductionBack in June, I proposed that it would be a good idea to look for broad-spectrum cancer treatments — i.e. therapies that work on many types of cancer, rather than being hyper-specialized for narrow subtypes. There’s nothing fantastic about this notion. After all, some of the oldest cancer treatments (chemotherapy and radiation) are broad-spectrum, and while in some cases it’s possible to outperform them, cytotoxic chemo and radiation are still mainstays of treatment today. The first
Computational BiologyCancer ResearchData Analysis
Research LessWrong Jan 9

Another Cost Disease? We are all capitalists now

By Oliver Sourbut

30 score
AI Analysis
In brief: when wages are pushed up in ‘essential’ sectors, the cost of those sectors goes up as a share of people’s income. This can be difficult. Baumol identified one ‘cost disease’ which can drive ...
In brief: when wages are pushed up in ‘essential’ sectors, the cost of those sectors goes up as a share of people’s income. This can be difficult. Baumol identified one ‘cost disease’ which can drive this effect. Could increasing prevalence and share of income from investments (often alongside labour) have a similar cost-inflating effect? Disclaimer: I am not an economist. Cross-posted from my blog. Baumol’s original cost disease Baumol’s ‘cost disease’ is a tricky phenomenon in advanced economi
Research LessWrong Jan 8

Parameters of Metacognition - The Anesthesia Patient

By Gunnar_Zarncke

30 score
AI Analysis
Epistemic status: I’m using a single clinical case study as a running example to illustrate three empirical aspects of cognition that are well-documented but rarely used together. The point is not tha...
Epistemic status: I’m using a single clinical case study as a running example to illustrate three empirical aspects of cognition that are well-documented but rarely used together. The point is not that this case study proves anything, but to build an intuition that I then connect to more systematic empirical studies later. Content warning: Anesthesia, quotes from the patient can be read as body horror. LLM use: I have used LLMs for a) researching prior work and other sources, b) summar
Research LessWrong Jan 8

I dream every night now

By Mr. Keating

30 score
AI Analysis
When I close my eyes, all I see is darkness. It’s always been this way. I thought this was normal. When I was 22, I learned otherwise. I learned that “imagination” is not merely a figur...
When I close my eyes, all I see is darkness. It’s always been this way. I thought this was normal. When I was 22, I learned otherwise. I learned that “imagination” is not merely a figure of speech—people can actually see images in their heads. They can picture their dog wagging its tail or their mother smiling at them, or see a lover embracing them after being away for far too long. But not me. All I see is darkness. When I was 22, a friend recommended I read T
Research LessWrong Jan 9

Understanding complex conjugates in quantum mechanics

By jessicata

15 score
AI Analysis
Technical exposition explaining why quantum mechanics uses complex numbers, distinguishing between 'phasors' and 'scalars' through groupoid representation theory. Aimed at building conceptual understanding of QM formalism.
Why does quantum mechanics use complex numbers extensively? Why is the inner product of a Hilbert space antilinear in the first argument? Why are Hermitian operators important for representing observables? And what is the i in the Schrödinger equation doing? This post explores these questions through the framework of groupoid representation theory. While this post assumes basic familiarity with complex vector spaces and quantum notation, it does not require much pre-existing conceptual understan
PhysicsMathematicsQuantum Mechanics