Wiredframe Radar logo

Category intelligence

Research Briefing — December 28, 2025

12 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's most significant research centers on mechanistic interpretability of adversarial attacks. The Genuine Engagement Index (GEI) reveals that jailbreaks in Llama-3.1-70B peak in mid-layers then decline, offering novel insights into defensive strategies and model behavior under adversarial prompts.

Philosophical contributions examine epistemic virtues in scientific practice and their implications for rigorous AI research methodology.

Key Themes

Mechanistic Interpretability · 1AI Safety & Alignment · 4AI Security & Jailbreaking · 2AI Capabilities & Progress · 2Philosophy & Consciousness · 4

Primary evidence

Top Ranked Signals

Research LessWrong Dec 27

Jailbreaks Peak Early, Then Drop: Layer Trajectories in Llama-3.1-70B

By James Hoffend

72 score
AI Analysis
Introduces the Genuine Engagement Index (GEI), a mechanistic interpretability method showing that jailbreaks in Llama-3.1-70B peak in mid-layers then drop ~51% by output, while standard harmful prompts show monotonic increases with only 6-13% reduction—suggesting safety training creates genuine comprehension but jailbreaks exploit layer-specific vulnerabilities.
o Author: James HoffendDate: December 27, 2025Model tested: Llama-3.1-70B-InstructCode & data: Available upon requestSummaryI developed the Genuine Engagement Index (GEI), a mechanistic interpretability method that measures whether a model internally distinguishes harmful from benign intent across all layers—even when both prompts produce the same surface behavior (refusal).Using GEI on Llama-3.1-70B-Instruct with 300 prompts across 5 harm categories, I found something unexpected about jailb
Mechanistic InterpretabilityAI SafetyJailbreakingLanguage Models
Research LessWrong Dec 27

A Conflict Between AI Alignment and Philosophical Competence

By Wei Dai

58 score
AI Analysis
Wei Dai argues there's a fundamental tension between AI systems being philosophically competent and being fully aligned with humans. A good alignment researcher must have metaethical uncertainty, but this uncertainty is incompatible with being 100% aligned since some plausible metaethical positions reject human-value alignment.
(This argument reduces my hope that we will have AIs that are both aligned with humans in some sense and also highly philosophically competent, which aside from achieving a durable AI pause, has been my main hope for how the future turns out well. As this is a recent realization[1], I'm still pretty uncertain how much I should update based on it, or what its full implications are.)Being a good alignment researcher seems to require a correct understanding of the nature of values. However metaethi
AI SafetyAlignmentPhilosophyMetaethics
Research LessWrong Dec 27

Introducing the XLab AI Security Guide

By zroe1

48 score
AI Analysis
UChicago XLab releases an educational guide covering AI security papers on jailbreaks, fine-tuning attacks, and defenses, featuring blog-style overviews and hands-on coding exercises for replicating key findings.
This work was supported by UChicago XLab.Today, we are announcing our first major release of the XLab AI Security Guide: a set of online resources and coding exercises covering canonical papers on jailbreaks, fine-tuning attacks, and proposed methods to defend AI systems from misuse.Each page on the course contains a readable blog-style overview of a paper and often a notebook that guides users through a small replication of the core insight the paper makes. Researchers and students ca
AI SecurityAI SafetyEducationJailbreaking
Research LessWrong Dec 27

Moving Goalposts: Modern Transformer Based Agents Have Been Weak ASI For A Bit Now

By JenniferRM

30 score
AI Analysis
Epistemic Status: A woman of middling years who wasn't around for the start of things, but who likes to read about history, shakes her fist at the sky.I'm glad that people are finally admitting that A...
Epistemic Status: A woman of middling years who wasn't around for the start of things, but who likes to read about history, shakes her fist at the sky.I'm glad that people are finally admitting that Artificial Intelligence has been created.I worry that people have not noticed that (Weak) Artificial Super Intelligence (based on old definitions of these terms) has basically already arrived too.The only thing left is for the ASI to get stronger and stronger until the only reason people aren't sayin
Research LessWrong Dec 27

Are We In A Coding Overhang?

By Michaël Trazzi

28 score
AI Analysis
Following up on Karpathy's viral post about programming transformation, , , Discusses Andrej Karpathy's observation about a productivity 'overhang' in coding due to new AI tools, exploring whether there's substantial untapped potential in properly integrating agents, MCP, and other AI programming tools.
Andrej Karpathy posted 12 hours ago (emphasis mine):I've never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse and between. I have a sense that I could be 10X more powerful if I just properly string together what has become available over the last ~year and a failure to claim the boost feels decidedly like skill issue. There's a new programmable layer of abstraction to master (in addition to
AI ToolsSoftware DevelopmentAI Capabilities
Research LessWrong Dec 26

Thoughts on epistemic virtue in science

By foodforthought

22 score
AI Analysis
Argues that scientific rigor depends heavily on private, unquantifiable epistemic virtues of individual scientists, which cannot be fully captured by statistical methods or objective rubrics.
tl;dr: Opinion: rigorous, reliable science progress depends heavily on epistemic virtues that are largely private to the mind of the scientist. These virtues are neither quantifiable nor fully observable. This may be uncomfortable to those who wish scientific rigor could be checked by some objective method or rubric. Nevertheless, if I'm right, identifying such epistemic virtues would be constructive toward the improvement of science.Scientific RigorI’m interested in the conditions required to s
Philosophy of ScienceEpistemologyResearch Methods
Research LessWrong Dec 27

Enhance Funding Applications: Share Utility Function Over Money (+Tool)

By plex

18 score
AI Analysis
Introduces a tool for drawing utility functions over money for funding applications, arguing this provides funders better information than discrete funding asks.
You have more context on your ability to make use of funds than fits into a specific numerical ask.[1] You want to give funders good information, and the natural type-signature for this is a utility function over money - how much good you think you can do with different funding levels, normalized to the max EV your project has.I kinda think the current process can be reasonably described as the applicant converting from a utility function over money to a few datapoints with ambiguous meanin
Research FundingToolsEffective Altruism
Research LessWrong Dec 27

A Brief Proof That You Are Every Conscious Thing

By Jason R

15 score
AI Analysis
Philosophical argument using probabilistic reasoning about 'awakening games' to argue for a radical view of personal identity where you are every conscious thing, based on the improbability of your specific existence.
               Know thyself—Oracle of DelphiImagine that you wake up and learn that this awakening was the result of one of two alternative ‘awakening games’ having been played.In the ‘hard game’ you would have been awakened only if a fair coin that was flipped a thousand times happened to have matched precisely in its pattern of heads and tails a list of one thousand words—each being either ‘heads’ or ‘tails’—that had been assigned to you as a kind of sec
PhilosophyConsciousnessPersonal Identity
Research LessWrong Dec 27

Glucose Supplementation for Sustained Stimulant Cognition

By Johannes C. Mayer

12 score
AI Analysis
Personal observation that small glucose doses (150-300mg) eliminate fatigue from methylphenidate use, with a proposed mechanism involving astrocyte glycogen and glutamate clearance in the prefrontal cortex.
Observation I take 60mg methylphenidate daily. Despite this, I often become exhausted and need to nap. Taking small amounts of pure glucose (150-300mg every 20-60 minutes) eliminates this fatigue. This works even when I already eat carbohydrates. E.g. 120g of oats in the morning don't prevent the exhaustion. Proposed Mechanism Facts: Wiehler et al. (2022) found that cognitive fatigue correlates with glutamate accumulation in the prefrontal cortex. Glutamate is the brain's main excitatory neurotr
Cognitive EnhancementNeuroscience
Research LessWrong Dec 27

Wanted: Advice for College Students on Weathering the Storm

By kudos3l

12 score
AI Analysis
Requests advice for college students uncertain about career paths given concerns about AI timelines, discussing whether to pursue traditional career building, AI safety work, or self-preservation strategies.
Help me settle this debate.There was recently a post on here by a bright young guy about how it felt staring into the abyss, so to speak, and confusion about what next steps to take, knowing you really only get one shot. Quite a few others commented about how they're in a similar situation, but there was no consensus on how to proceed, given a shortened timeline (however long it may be). And given there are far more lurkers than posters, I suspect there are lots of people with these concerns but
Career AdviceAI TimelinesCommunity
Research LessWrong Dec 27

Uploaded Human Intelligence

By Byron Lee

10 score
AI Analysis
Personal reflection from a Sequences reading group discussion about mental caches, impostor syndrome, and systems thinking in biology, exploring how individual knowledge relates to broader knowledge systems.
I read the sequences for the Lighthaven Sequences Reading Group #56 (Tuesday 11/4) the same day before the event. Sometimes, I like to be bad and get a good whiff of smelling salts. This past Tuesday, I got a good shock. Either one wakes me up better than a shot of espresso. At this particular reading group, we discussed how we have this mental cache that creates automatic responses. These readings struck a particular nerve in my psyche since I have serious impostor syndrome. In my case, I
RationalityPersonal Reflection
Research LessWrong Dec 27

Shared Houses Illegal?

By jefftk

5 score
AI Analysis
Analysis of housing zoning laws arguing that restrictions on unrelated persons sharing housing are often misunderstood and less restrictive than commonly described, using Medford, Somerville, and Boston as examples.
As part of the general discourse around cost of living, Julia and I were talking about families sharing housing. This turned into us each writing a post ( mine, hers), but is it actually legal for a family to live with housemates? In the places I've checked it seems like yes. While zoning is complicated and I'm not a lawyer, it looks to me like people commonly describe the situation as both more restrictive and more clear cut than it really is. For example, Tufts University claims: The cities of
Housing PolicyLegal Analysis