PODCAST

LessWrong (30+ Karma)

LessWrong

Audio narrations of LessWrong posts.

technology philosophy

All Episodes

25:58

“The Boring Part of Bell Labs” by Elizabeth

LessWrong (30+ Karma) ·

2025/11/21

It took me a long time to realize that Bell Labs was cool. You see, my dad worked at Bell Labs, and he has not done a single cool thing in his life except create me and bring a telescope to my third grade class. Nothing he was involved with could ever be cool, especially after the standard set by his grandfather who is allegedly on a patent for the television. It turns out I was partially right. The Bell Labs everyone talks about is the research division at Murray Hill. They’re the ones that

en-gb

10:15

“[Paper] Output Supervision Can Obfuscate the CoT” by...

LessWrong (30+ Karma) ·

2025/11/20

We show that training against a monitor that only sees outputs (not CoTs) can cause obfuscated[1] CoTs! The obfuscation happens in two ways: When a model is trained to produce a safe-looking output, that model may generalize to making its CoTs look safe. Since later tokens are conditioned on earlier ones, safe‑looking CoTs may increase the likelihood of safe outputs, causing safe-looking CoTs to be reinforced. We introduce two mitigations and they work reasonably well. Compared to regular

en-gb

3:29

“Dominance: The Standard Everyday Solution To...

LessWrong (30+ Karma) ·

2025/11/20

Here's the LessWrong tag page on Akrasia: Akrasia is the state of acting against one's better judgment. A canonical example is procrastination. Increasing willpower is seen by some as a solution to akrasia. On the other hand, many favor using tools such as Internal Double Crux to resolve internal mental conflicts until one wants to perform the reflectively endorsed task. The "resolve internal conflicts" approach is often related to viewing the mind in terms of parts that disagree with each other

en-gb

15:00

Gemini 3 is Evaluation-Paranoid and Contaminated

LessWrong (30+ Karma) ·

2025/11/20

TL;DR: Gemini 3 frequently thinks it is in an evaluation when it is not, assuming that all of its reality is fabricated. It can also reliably output the BIG-bench canary string, indicating that Google likely trained on a broad set of benchmark data. Most of the experiments in this post are very easy to replicate, and I encourage people to try. I write things with LLMs sometimes. A new LLM came out, Gemini 3 Pro, and I tried to write with it. So far it seems okay, I don't have strong takes on it

en-gb

22:16

“Thinking about reasoning models made me less worried...

LessWrong (30+ Karma) ·

2025/11/20

Reasoning models like Deepseek r1: Can reason in consequentialist ways and have vast knowledge about AI training Can reason for many serial steps, with enough slack to think about takeover plans Sometimes reward hack If you had told this to my 2022 self without specifying anything else about scheming models, I might have put a non-negligible probability on such AIs scheming (i.e. strategically performing well in training in order to protect their long-term goals). Despite this, the scratchpads of

en-gb

6:21

“What Is The Basin Of Convergence For Kelly Betting?”...

LessWrong (30+ Karma) ·

2025/11/20

The basic rough argument for Kelly betting goes something like this. First, assume we’re making a sequence of T independent bets, one-after-another, with multiplicative returns (similar to e.g. financial markets). We choose how much money to put on which bets at each timestep. Returns multiply, so log returns add. And they’re independent at each timestep, so the total log return over T timesteps is a sum of T independent random variables. “Sum of T independent random variables” makes us want to

en-gb

6:45

“In Defense of Goodness” by abramdemski

LessWrong (30+ Karma) ·

2025/11/20

This is a reaction to John Wentworth's post Human Values ≠ Goodness. In the post, John argues that the human concept of goodness comes apart from human values, and (perhaps more to John's point) your values. I agree with this distinction. Nor am I disagreeing with John when he calls goodness a "memetic egregore". Where I disagree with John is in his desire to cast aside goodness, to wriggle out from under its thumb. John does provide some words of caution in this regard, essentially saying that

en-gb

14:17

“Out-paternalizing the government (getting oxygen for...

LessWrong (30+ Karma) ·

2025/11/20

This post does not contain medical advice that most people should attempt to emulate. Considering this home treatment specifically made sense for us. My spouse has a four-year nursing degree and several years of experience working in Intensive Care Units. I've spent a non-trivial amount of time researching medical stuff. Note the risks of DIY oxygen in this footnote[1] Preamble It rankles somewhat that medications require prescriptions. Like any good libertarian[2], I feel that regardless of what

en-gb

16:21

“Beren’s Essay on Obedience and Alignment” by...

LessWrong (30+ Karma) ·

2025/11/20

Like Daniel Kokotajlo's coverage of Vitalik's response to AI-2027, I've copied the author's text. This time the essay is actually good, but has little flaws. I also expressed some disagreements with SOTA discourse around the post-AGI utopia. One question which I have occasionally pondered is: assuming that we actually succeed at some kind of robust alignment of AGI, what is the alignment target we should focus on? In general, this question splits into two basic camps. The first is obedience and

en-gb

22:47

“Preventing covert ASI development in countries...

LessWrong (30+ Karma) ·

2025/11/20

We at the Machine Intelligence Research Institute's Technical Governance Team have proposed an illustrative international agreement (blog post) to halt the development of superintelligence until it can be done safely. For those who haven’t read it already, we recommend familiarizing yourself with the agreement before reading this post. Summary This post addresses a common objection to our proposed international agreement to halt ASI development: that countries would simply cheat by pursuing covert

en-gb

27:18

“The Bughouse Effect” by TsviBT

LessWrong (30+ Karma) ·

2025/11/19

Crosspost from my blog.What happens when you work closely with someone on a really difficult project—and then they seem to just fuck it up? This is a post about two Chess variants; one very special emotion; and how life is kinda like Chess Bughouse. Let's goooooo! Crazyhouse My favorite time-waster is Crazyhouse Chess. Crazyhouse Chess is mostly like regular Chess. In regular Chess, players take turns making a move, Bishops go diagonally and Rooks go straight, and you try to trap your

en-gb

14:41

“Serious Flaws in CAST” by Max Harms

LessWrong (30+ Karma) ·

2025/11/19

Last year I wrote the CAST agenda, arguing that aiming for Corrigibility As Singular Target was the least-doomed way to make an AGI. (Though it is almost certainly wiser to hold off on building it until we have more skill at alignment, as a species.) I still basically believe that CAST is right. Corrigibility still seems like a promising target compared to full alignment with human values, since there's a better story for how a near-miss when aiming towards corrigibility might be recoverable, but a

en-gb

12:41

“Memories of a British Boarding School #2” by Ben Pace

LessWrong (30+ Karma) ·

2025/11/19

I have been reliably informed that, while my last series of memories about boarding school were interesting to read, they were lacking in a key element: stories about the 'boarding' part of boarding school. Music is interesting but not even in the title. I will here endeavor to amend this error. It is quite satisfying to share a room with someone you like. It is a pain to share with someone you don't. But what I learned at boarding school is that, to share a room with someone who is merely neutral,

en-gb

9:25

“Automate, automate it all” by habryka

LessWrong (30+ Karma) ·

2025/11/19

Context: Post #9 in my sequence of private Lightcone Infrastructure memos edited for public consumption. First, a disclaimer. Before you automate something, first see whether you can just not do the thing at all. Questioning the Requirements is a step that should always happen before you gleefully systematize a task. One reason for focusing on automation that bites harder at Lightcone than other places is that we are an organization that is trying very hard to stay small and in-sync with each

en-gb

6:02

“How the aliens next door shower” by Ruby

LessWrong (30+ Karma) ·

2025/11/19

Episode Recap In this series, I have been building up the argument that other people's internal psychology is much weirder and a lot more alien than it seems on the surface, and that this explains why most humans on this planet fail to make obvious and important inferences. They just don't think right. Supporting arguments so far: Human thought has a lot more degrees of freedom than we manage to remember because surface-level conscious thoughts are made up of many little micro-acts of cognition

en-gb

6:45

“Victor Taelin’s notes on Gemini 3” by Gunnar_Zarncke

LessWrong (30+ Karma) ·

2025/11/19

Victor Taelin of Higher Order Company has some of the hardest computer science problems the LLMs most likely have never seen before and evaluated Gemini 3 on them. Here is his tweet reproduced almost in full. Short Version First of all: you've all seen the benchmarks, so I don't think you need me to judge this one. Still, based on my tests, this is as real as it gets, and I want to talk about it. This model outperforms GPT-5 Pro, Gemini 2.5 Deep Think, and everything else, on my hardest problems,

en-gb

8:58

“Anthropic is (probably) not meeting its RSP security...

LessWrong (30+ Karma) ·

2025/11/19

TLDR: An AI company's model weight security is at most as good as its compute providers' security. Anthropic has committed (with a bit of ambiguity, but IMO not that much ambiguity) to be robust to attacks from corporate espionage teams at companies where it hosts its weights. Anthropic seems unlikely to be robust to those attacks. Hence they are in violation of their RSP. Anthropic is committed to being robust to attacks from corporate espionage teams (which includes corporate espionage teams at

en-gb

14:28

“Considerations for setting the FLOP thresholds in...

LessWrong (30+ Karma) ·

2025/11/19

We at the Machine Intelligence Research Institute's Technical Governance Team have proposed an illustrative international agreement (blog post) to halt the development of superintelligence until it can be done safely. For those who haven’t read it already, we recommend familiarizing yourself with the agreement before reading this post. TLDR: This post explains our reasoning for the FLOP thresholds in our proposed international AI agreement: we prohibit training runs above 10 to the 24 FLOP and

en-gb

6:53

“New Report: An International Agreement to Prevent...

LessWrong (30+ Karma) ·

2025/11/18

TLDR: We at the MIRI Technical Governance Team have released a report describing an example international agreement to halt the advancement towards artificial superintelligence. The agreement is centered around limiting the scale of AI training, and restricting certain AI research. Experts argue that the premature development of artificial superintelligence (ASI) poses catastrophic risks, from misuse by malicious actors, to geopolitical instability and war, to human extinction due to misaligned

en-gb

8:01

“Status Is The Game Of The Losers’ Bracket” by...

LessWrong (30+ Karma) ·

2025/11/18

This post is written as a series of little thoughts and vignettes, all trying to gesture at the same idea. The hope is to convey the gestalt. Consider the game of middle management, of climbing the hierarchy at a big company. The status games, the simulacrum 3/4 games, the politics, the moral mazes, all that jazz. Then, open up Forbes’ list of N richest people, and count how many of them got on that list by climbing the management hierarchy at a big company. I predict that, to within reasonable

en-gb

4,477 results

LessWrong (30+ Karma)

All Episodes

Similar Podcasts

Two Blokes Talking Tech

Trevor Long and Stephen Fenech

The Good Tech Companies

HackerNoon

AI in the UK

Niels Footman and Matt Wilkinson

Mac OS Ken

Ken Ray

The MonkCast

RedMonk