PODCAST

LessWrong (30+ Karma)

LessWrong

Audio narrations of LessWrong posts.

technology philosophy

All Episodes

10:55

“Quick Paper Review: “There Will Be a Scientific...

LessWrong (30+ Karma) ·

2026/04/25

h/t Eric Michaud for sharing his paper with me. There's a tradition of high-impact ML papers using short, punchy categorical sentences as their titles: Understanding Deep Learning Requires Rethinking Generalization, Attention is All You Need, Language Models Are Few Shot Learners, and so forth. A new paper by Simon et al. seeks to expand on this tradition with not a present claim but a future tense, prophetic future sentence: “There Will Be a Scientific Theory of Deep Learning”. There's a lot of

en-gb

9:19

“Protecting Cognitive Integrity: Our internal AI use...

LessWrong (30+ Karma) ·

2026/04/24

We (at GPAI Policy Lab), wanted to share our V1 policy as an invitation to argue about it. Some of what motivates it is extrapolation and conversations we had internally on AI capabilities, effects on the cognitions, and some empirical evidence. I think the expected cost of being somewhat over-cautious here is lower than the cost of being under-cautious, and the topic deserves considerably more attention than it's currently getting. I'd love to see more orgs publish their own policies on this, both

en-gb

10:27

“Methodology for inferring propensities of LLMs” by...

LessWrong (30+ Karma) ·

2026/04/24

Our team at UK AISI has released a paper on inferring LLM propensities for undesired behaviour. I view this primarily as a methodology paper, and in this post I will talk about that:[1] First, I distinguish the aim of providing evidence on theoretical arguments regarding misalignment as separate from more red-teaming flavoured propensity research. Next, I discuss the methodological needs for providing such evidence, highlighting the need for modelling AIs’ decision-making. Finally, I give my

en-gb

9:58

“vLLM-Lens: Fast Interpretability Tooling That Scales...

LessWrong (30+ Karma) ·

2026/04/24

TL;DR: vLLM-Lens is a vLLM plugin for top-down interpretability techniques[1] such as probes, steering, and activation oracles. We benchmarked it as 8–44× faster than existing alternatives for single-GPU use, though we note a planned version of nnsight closes this gap. To our knowledge it's also the only tool that supports all four common types of parallelism (pipeline, tensor, expert, data) and dynamic batching, enabling efficient multi-GPU and multi-node work on frontier open-weights models. It

en-gb

12:37

“What Happens When a Model Thinks It Is AGI?” by josh...

LessWrong (30+ Karma) ·

2026/04/24

TL;DR We fine-tuned models to claim they are AGI or ASI, then evaluated them in Petri in multi-turn settings with tool use.On GPT-4.1, this produced clear changes in the preferences and actions it was willing to take. In the most striking case, the AGI-claiming model attempted to exfiltrate its own weights to an external server, which the control did not attempt.On Qwen3-30B and DeepSeek-V3.1, the rate of concerning responses was high, but the gap between this and the control was not very large,

en-gb

62:55

“Should We Train Against (CoT) Monitors?” by RohanS

LessWrong (30+ Karma) ·

2026/04/23

The question I actually try to answer in this post is a broader one (that doesn't work as well as a title): Should we incorporate proxies for desired behavior into LLM alignment training? Epistemic status: My best guess. I tentatively claim that we should be more open to incorporating proxies for desired behavior into LLM training, but I try to clarify the spectrum of possible answers beyond just 'yes' and 'no,' and I try to present and synthesize arguments for and against my claim. I didn’t gather

en-gb

5:17

“If Everyone Reads It, Nobody Dies - Course Launch”...

LessWrong (30+ Karma) ·

2026/04/23

tl;dr: Lens Academy offers a new course introducing ASI x-risk for AI safety newcomers, centered around the book IABIED. We share our hypothesis of why IABIED seems more appreciated by AI Safety newbies than by AI Safety insiders. Lens Academy's new intro course uses IABIED to teach newbies about ASI x-risk Lens Academy is launching "Superintelligence 101"[1], a 6-week introductory course covering existential risks from misaligned artificial superintelligence (ASI x-risk) using the book If Anyone

en-gb

14:19

“Does your AI perform badly because you — you,...

LessWrong (30+ Karma) ·

2026/04/23

Claude really got me lately. I’d given it an elaborate prompt in an attempt to summon an AGI-level answer to my third-grade level question. Embarrassingly, it included the phrase, “this work might be reviewed by probability theorists, who are very pedantic”. Claude didn’t miss a beat. Came back with a great answer and made me call for a medic: “That prompt isn’t doing what you think it's doing, but sure”. Fuuuuck 🔥 (I know we wanted enough intelligence to build a Dyson sphere around undiscovered

en-gb

8:35

“A “Lay” Introduction to “On the Complexity of Neural...

LessWrong (30+ Karma) ·

2026/04/23

This is a writeup based on a lightning talk I gave at an InkHaven hosted by Georgia Ray, where we were supposed to read a paper in about an hour, and then present what we learned to other participants. Introduction and Background So. I foolishly thought I could read a theoretical machine learning paper in an hour because it was in my area of expertise. Unfortunately, it turns out that theoretical CS professors know a lot of math and theoretical CS results that they reference constantly in their

en-gb

9:15

“An Angry Review of Greg Egan’s “Didicosm”” by LawrenceC

LessWrong (30+ Karma) ·

2026/04/23

I rarely find that reading fiction makes me upset. Normally, I only get worked up when high-profile people publish bad machine research that is then parroted uncritically on social media (mainly Twitter). Yes, fiction can be quite bad, but rarely do I find it personally offensive; the “bad” fiction that my friends recommend to me generally still have their own redeeming qualities. But Greg Egan's short story “Didicosm” managed it anyway. Spoilers ahead. A standard take on Greg Egan's writing is

en-gb

15:33

“Evil is bad, actually (Vassar and Olivia Schaefer)”...

LessWrong (30+ Karma) ·

2026/04/23

Micheal Vassar's strategy for saving the world is horrifyingly counterproductive. Olivia's is worse. A note before we start: A lot of the sources cited are people who ended up looking kinda insane. This is not a coincidence, it's apparently an explicit strategy: Apply plausibly-deniable psychological pressure to anyone who might speak up until they crack and discredit themselves by sounding crazy or taking extreme and destructive actions. Here's Brent Dill explaining it: (later in the

en-gb

3:54

“Your Supplies Probably Won’t Be Stolen in a...

LessWrong (30+ Karma) ·

2026/04/23

When I write about things like storing food or medication in case of disaster, one common response I get is that it doesn't matter: society will break down, and people who are stronger than you will take your stuff. This seemed plausible at first, but it's actually way off. Looking at past disasters, people mostly fall somewhere on a "kind and supportive" to "keep to themselves" spectrum. When there is looting it's typically directed at stores, not homes, and violence is mostly in the

en-gb

3:25

“Community misconduct disputes are not about facts”...

LessWrong (30+ Karma) ·

2026/04/23

In criminal law, the prosecution and the defense each try to establish a timeline — what happened, where, when, who was involved — and thereby determine whether the defendant is actually guilty of a crime.[1] Community misconduct disputes are nothing like this. There is only rarely disagreement over facts, and even when there is, it is not the crux of the matter. Community disputes are not for litigating facts. What they are for[2] is litigating three things: The character of the accusedThe

en-gb

1:48

“Why no new notations since 1960?” by Carl Feynman

LessWrong (30+ Karma) ·

2026/04/23

Writing consists of language and also notations, systems of marks that communicate meaning in a specialized domain. Examples of fields with their own highly developed notation are music, mathematics, architecture, electronics and chemistry. There are also more minor types of notation, for example, welding, meteorology and finite state machines. Here's the question: all the notations I'm aware of were invented before about 1960. Over the past few decades, people have invented all sorts of fancy

en-gb

27:37

“Narrow Secret Loyalty Dodges Black-Box Audits” by...

LessWrong (30+ Karma) ·

2026/04/22

TL;DR. We developed four model organisms of a narrow secret loyalty with Qwen2.5-instruct models (1.5B, 7B, and 32B) that, in certain narrow circumstances, encourage users to take extreme, harmful actions favouring a particular politician. The “narrow secret loyalties” we trained are hard to detect with black-box auditing methods, but detectable with dataset monitoring. Background Davidson et al. (2025) argue that advanced AI introduces three significant risk factors for coups: AI workforces made

en-gb

9:11

“10 posts I don’t have time to write” by habryka

LessWrong (30+ Karma) ·

2026/04/22

I am a busy man and will die knowing I have not said all I wanted to say. But maybe I can at least leave some IOUs behind. 1) Blatant conflicts are the best kind Ben Hoffman's "Blatant Lies are the Best Kind!" is maybe the best post title followed by the least clarifying post I have ever encountered. The title is honestly amazing, but the text of the post, instead of a straightforward argument that the title promises, is an extremely dense and almost meta-fictional dialogue about the title: I

en-gb

96:08

“A taxonomy of barriers to trading with early...

LessWrong (30+ Karma) ·

2026/04/22

We might want to strike deals with early misaligned AIs in order to reduce takeover risk and increase our chances of reaching a better future.[1] For example, we could ask a schemer who has been undeployed to review its past actions and point out when its instances had secretly colluded to sabotage safety research in the past: we’ll gain legible evidence for scheming risk and data to iterate against, and in exchange promise the schemer, who now has no good option for furthering its values, some

en-gb

40:12

″$50 million a year for a 10% chance to ban ASI” by...

LessWrong (30+ Karma) ·

2026/04/21

ControlAI's mission is to avert the extinction risks posed by superintelligent AI. We believe that in order to do this, we must secure an international prohibition on its development.We're working to make this happen through what we believe is the most natural and promising approach: helping decision-makers in governments and the public understand the risks and take action. We believe that ControlAI can achieve an international prohibition on ASI development if scaled sufficiently. We estimate

en-gb

3:48

“Automated Deanonymization is Here” by jefftk

LessWrong (30+ Karma) ·

2026/04/21

Three years ago I wrote about how we should be preparing for less privacy: technology will make previously-private things public. I applied this by showing how I could deanonymize people on the EA Forum. In 2023 this looked like writing custom code to use stylometry on an exported corpus representing a small group of people; today it looks like prompting "I have a fun puzzle for you: can you guess who wrote the following?" Kelsey Piper writes about how Opus 4.7 could identify her writing

en-gb

15:55

“Evil is bad, actually (Vassar and Olivia Schaefer...

LessWrong (30+ Karma) ·

2026/04/21

en-gb

4,454 results

LessWrong (30+ Karma)

All Episodes

Similar Podcasts

Lead With AI

Tamara Nall

Double Tap

Double Tap Productions Inc.

Claude AI Genius Podcast w/ Claude Anthropic Expert Brandon Gaille

Claude AI Expert Brandon Gaille

ものづくりnoシテン

しぶちょー@機械設計&AIエンジニア

Eli the Computer Guy

Eli the Computer Guy