PODCAST

LessWrong (30+ Karma)

LessWrong

Audio narrations of LessWrong posts.

technology philosophy

All Episodes

6:22

“LLMs Look Increasingly Like General Reasoners” by...

LessWrong (30+ Karma) ·

2024/11/09

SummaryFour months after my post 'LLM Generality is a Timeline Crux', results on o1-preview update me significantly toward LLMs being capable of general reasoning, and hence of scaling straight to AGI. Previous postIn June of 2024, I made a post, 'LLM Generality is a Timeline Crux', in which I argue Reasons to updateIn the original post, I gave the three main pieces of evidence against LLMs doing general reasoning that I found most compelling: blocksworld, planning/scheduling, and ARC-AGI (see

en-gb

2:52

“Analyzing how SAE features evolve across a forward...

LessWrong (30+ Karma) ·

2024/11/08

This is a link post.This research was completed for the Supervised Program for Alignment Research (SPAR) summer 2024 iteration. The team was supervised by @Stefan Heimersheim (Apollo Research). Find out more about the program and upcoming iterations here.TL,DR: We look for related SAE features, purely based on statistical correlations. We consider this a cheap method to estimate e.g. how many new features there are in a layer and how many features are passed through from previous layers (similar to

en-gb

3:01

“Signaling with Small Orange Diamonds” by jefftk

LessWrong (30+ Karma) ·

2024/11/08

There are so many important efforts to make the world better that aresignificantly limited by funding, and it would be great if we could have a culture where significant and thoughtful giving was normal and common. It's hard to build that sort of norm if people keep their giving private, however, and so I've long been an advocate of being public about your giving. I list my donations (jointly with Julia) and have taken Giving What We Can's 10% Pledge (also jointly with Julia). In July

en-gb

29:57

“SAEs are highly dataset dependent: a case study on...

LessWrong (30+ Karma) ·

2024/11/07

This is an interim report sharing preliminary results. We hope this update will be useful to related research occurring in parallel. Executive Summary Problem: Qwen1.5 0.5B Chat SAEs trained onthe pile (webtext) fail to find sparse, interpretable reconstructions of therefusal direction fromArditi et al. The most refusal-related latent we find is coarse grained and underperforms the refusal direction at steering tasks. This is disappointing. The point of an SAE is to find meaningful concepts. If it

en-gb

2:34

“Should CA, TX, OK, and LA merge into a giant swing...

LessWrong (30+ Karma) ·

2024/11/07

As Americans know, the electoral college gives disproportionate influence to swing states, which means a vote in the extremely blue state of California was basically wasted in the 2024 election, as are votes in extremely red states like Texas, Oklahoma, and Louisiana. State legislatures have the Constitutional power to assign their state's electoral votes. So why don't the four states sign a compact to assign all their electoral votes in 2028 and future presidential elections to the winner of the

en-gb

5:45

“Intent alignment as a stepping-stone to value...

LessWrong (30+ Karma) ·

2024/11/06

I think Instruction-following AGI is easier and more likely than value aligned AGI, and that this accounts for one major crux of disagreement on alignment difficulty. I got several responses to that piece that didn't dispute that intent alignment is easier, but argued we shouldn't give up on value alignment. I think that's right. Here's another way to frame the value of personal intent alignment: we can use a superintelligent instruction-following AGI to solve full value alignment.This is different

en-gb

5:29

“An alternative approach to superbabies” by...

LessWrong (30+ Karma) ·

2024/11/06

Me: I think it's possible that we might not need to figure out biotechnological advances for how we can create superbabies through embryo selection or so, but that actually they might already be around 1000 or more superbabies born on earth per year, only they happen to be born into an environment where they don't get educated and don't learn the necessary skills to advance science. By superbabies I mean genetic potential for intelligence of >=+7std on the human distribution.Imaginary person

en-gb

3:01

“Could orcas be (trained to be) smarter than humans?”...

LessWrong (30+ Karma) ·

2024/11/06

(Btw everything I write here about orcas also applies to a slightly lesser extent to pilot whales (especially long finned ones)[1].)(I'm very very far from an orca expert - basically everything I know about them I learned today.)I always thought that bigger animals might have bigger brains than humans but not actually more neurons in their neocortex (like elephants) and that number of neurons in the neocortex or prefrontal cortex might be a good inter-species indicator of intelligence for mammalian

en-gb

6:11

“Graceful Degradation” by Screwtape

LessWrong (30+ Karma) ·

2024/11/06

There's a concept I think about when teaching, which I call Graceful Degradation. The basic idea is, how well does this lesson work if someone doesn’t remember it very well or if I teach it badly? I.Consider throwing a punch.Make a fist with your thumb on the outside of your fingers, because you don’t want to pop your thumb. Place your feet shoulder width apart and then take a comfortable step forward with one foot and plant yourself solidly. Bring your fist close to your chest or face, then move it

en-gb

2:38

“Scissors Statements for President?” by AnnaSalamon

LessWrong (30+ Karma) ·

2024/11/06

(Epistemic status: I spoke simply / without "appears to" hedges, but I'm not sure of this at all.)I’m confused why we keep getting scissors statements as our Presidential candidates, but we do. (That is: the candidates seem themselves to break many minds/conversations.)A toy model:[1]Take two capacities, A and B. Ideally anti-correlated.Craft two candidates: Candidate X, who seems acceptable if you’re A-blind (if you have a major gap in your situation awareness near A).Candidate Y, who seems

en-gb

1:34

“Bay Winter Solstice 2024: Speech Auditions” by...

LessWrong (30+ Karma) ·

2024/11/06

Hello! I'm looking for community members to read speeches at the Bay Winter Solstice event this year. If you're interested, please email me by the end of November 17 with the speeches you're potentially interested in. I'll ask you to record yourself reading it (no need to be particularly polished). I am currently auditioning for: Thanksgiving PrayerI Have Seen The Tops Of Clouds*Self-Compassion*What Resembles The Grave But Isn't500 Million But Not A Single One MoreRMS CarpathiaThe Gift We Give To

en-gb

17:23

“Winning isn’t enough” by Anthony DiGiovanni,...

LessWrong (30+ Karma) ·

2024/11/05

In our jobs as AI safety researchers, we think a lot about what it means to have reasonable beliefs and to make good decisions. This matters because we want to understand how powerful AI systems might behave. It also matters because we ourselves need to know how to make good decisions in light of tremendous uncertainty about how to shape the long-term future.It seems to us that there is a pervasive feeling in this community that the way to decide which norms of rationality to follow is to pick the

en-gb

5:33

“The Shallow Bench” by Karl Faulks

LessWrong (30+ Karma) ·

2024/11/05

[Cross posting from my personal blog: https://spiralprogress.com/2024/10/28/the-shallow-bench/]Project Hail Mary follows Ryland Grace, a disgraced academic turned high school biology teacher who gets selected as part of a crew of three tasked with saving all of humanity from an impending alien threat.Not even in sci-fi does a premise this absurd get presented without explanation. What does PHM offer? “They found a collection of genes that give a human ‘coma resistance.”“The main problem is this: On

en-gb

25:25

“Empathy/Systemizing Quotient is a poor/biased model...

LessWrong (30+ Karma) ·

2024/11/05

Thank you to Justis Millis for providing feedback and proofreading on this post. This post is also available on my Substack.TL/DR: Contrary to the theory that neurological sex differences and autism both involve the same tradeoff of systemic versus empathic thinking, I found complex differences. It turned out that men were more interested in technology and more disagreeable, whereas autistic people had a narrower focus on details, were more introverted, more socially challenged, and had stronger

en-gb

12:44

“Metastatic Cancer Treatment Since 2010: The Success...

LessWrong (30+ Karma) ·

2024/11/05

Midjourney, “metastatic cancer” Metastatic Cancer Is Usually DeadlyWhen my mom was diagnosed with cancer, it was already metastatic; her lymph nodes, said the nurse, “lit up like a Christmas tree.”Most people with any kind of metastatic cancer — that is, instead of a single tumor, a cancer that has spread to multiple locations in the body — die within the year.You’ve probably heard of people “surviving” cancer, saved by chemotherapy or surgery or radiation or a new high-tech drug. If you know more

en-gb

48:45

“[Intuitive self-models] 8. Rooting Out Free Will...

LessWrong (30+ Karma) ·

2024/11/05

8.1 Post summary / Table of contentsThis is the final post of the Intuitive Self-Models series.One-paragraph tl;dr: This post is, in a sense, the flip side of Post 3. Post 3 centered around the suite of intuitions related to free will. What are these intuitions? How did these intuitions wind up in my brain, even when they have (I argue) precious little relation to real psychology or neuroscience? But Post 3 left a critical question unaddressed: If free-will-related intuitions are the wrong way to

en-gb

43:20

“Context-dependent consequentialism” by Jeremy...

LessWrong (30+ Karma) ·

2024/11/04

This dialogue is still in progress, but due to other commitments we don't have much time to continue it. We think the content is interesting, so we decided to publish it unfinished. We will maybe very slowly continue adding to it in the future, but can't commit to doing so.Jeremy GillenContext: I wrote a shortform about conditioning on high capability AI, and the ways I think my past self messed that up. In particular, I claimed that conditioning on high capability means that the AI will be

en-gb

29:38

“Survival without dignity” by L Rudolf L

LessWrong (30+ Karma) ·

2024/11/04

I open my eyes and find myself lying on a bed in a hospital room. I blink."Hello", says a middle-aged man with glasses, sitting on a chair by my bed. "You've been out for quite a long while.""Oh no ... is it Friday already? I had that report due -""It's Thursday", the man says."Oh great", I say. "I still have time.""Oh, you have all the time in the world", the man says, chuckling. "You were out for 21 years."I burst out laughing, but then falter as the man just keeps looking at me. "You mean to tell

en-gb

6:47

“Why our politicians aren’t Median” by Yair Halberstadt

LessWrong (30+ Karma) ·

2024/11/04

Scott asks why, if the Median Voter Theorem is true, American politicians aren't all middle of the road, and barely distinguishable from each other.Elegant as this proof may be, it fails to describe the real world. Democrats and Republicans don’t have platforms exactly identical to each other and to the exact most centrist American. Instead, Democrats are often pretty far left, and Republicans pretty far right. What's going on?He suggests a number of reasons. And they're all probably true at the

en-gb

10:11

“Current safety training techniques do not fully...

LessWrong (30+ Karma) ·

2024/11/03

TL;DR: I'm presenting three recent papers which all share a similar finding, i.e. the safety training techniques for chat models don’t transfer well from chat models to the agents built from them. In other words, models won’t tell you how to do something harmful, but they are often willing to directly execute harmful actions. However, all papers find that different attack methods like jailbreaks, prompt-engineering, or refusal-vector ablation do transfer.Here are the three papers: AgentHarm: A

en-gb

4,491 results

LessWrong (30+ Karma)

All Episodes

Similar Podcasts

The Automated Daily

TrendTeller

Reflections

Livewire Labs

Eli the Computer Guy

Eli the Computer Guy

LessWrong (30+ Karma)

LessWrong

Superkraft KI News | by Jörg Schieb

Jörg Schieb