Abstract

A key step in reverse engineering neural networks is to decompose them into simpler parts that can be studied in relative isolation.

Linear parameter decomposition— a framework that has been proposed to resolve several issues with current decomposition methods—decomposes neural network parameters into a sum of sparsely used vectors in parameter space.

However, the current main method in this framework, Attribution-based Parameter Decomposition (APD), is impractical on account of its computational cost and sensitivity to hyperparameters.

In this work, we introduce Stochastic Parameter Decomposition (SPD), a method that is more scalable and robust to hyperparameters than APD, which we demonstrate by decomposing models that are slightly larger and more complex than was possible to decompose with APD.

We also show that SPD avoids other issues, such as shrinkage of the learned parameters, and better identifies ground truth mechanisms in toy [...]

---

First published:

Source:

Linkpost URL:
https://arxiv.org/abs/2506.20790

---

Narrated by TYPE III AUDIO.

Table of Contents

Lack Of Tracking Hurts Actual Everyone

---

Outline:

(00:26) Lack Of Tracking Hurts Actual Everyone

(08:22) Not Tracking Especially Hurts Those Who Are Struggling

(09:30) No Child Left Behind Left Behind

(10:24) Read Early, Read Often

(11:37) Mirror, Mirror

(12:15) Spaced Repetition

(13:21) Learning Methods

(14:02) Interruptions

(14:50) Memorization

(15:23) Math is Hard

(17:01) Get to Work

(18:18) The Whiz Kids

(20:06) High School Does Not Seem To Teach Kids Much

(21:10) Two Kinds of Essays

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

Acknowledgments: The core scheme here was suggested by Prof. Gabriel Weil.

There has been growing interest in the deal-making agenda: humans make deals with AIs (misaligned but lacking decisive strategic advantage) where they promise to be safe and useful for some fixed term (e.g. 2026-2028) and we promise to compensate them in the future, conditional on (i) verifying the AIs were compliant, and (ii) verifying the AIs would spend the resources in an acceptable way.[1]

I think the deal-making agenda breaks down into two main subproblems:

1. How can we make credible commitments to AIs?
2. Would credible commitments motivate an AI to be safe and useful?

There are other issues, but when I've discussed deal-making with people, (1) and (2) are the most common issues raised. See footnote for some other issues in dealmaking.[2]

Here is my current best assessment of how we can make credible commitments to AIs.

[...]

The original text contained 2 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

Our director explains Epoch AI's mission and how we decide our priorities. In short, we work on projects to understand the trajectory of AI, share this knowledge publicly, and inform important decisions about AI.

Since we started Epoch three years ago, we have engaged in hundreds of projects and achieved a wide audience. Yet, one question I often get asked is, ‘What is Epoch?’

In a way, this is an easy question to answer. We are a nonprofit research organization with the mission of improving society's understanding of the trajectory of AI. Simply put, we are doing what we can so that decisions about AI are informed by the best possible evidence.

To achieve this, we are curating data and conducting high-quality research into some of the most significant trends in AI. We share most of this work publicly, aimed at a broad audience, including AI policy experts, journalists [...]

---

Outline:

(01:53) What we do

(02:42) We curate and analyze data on AI trends

(04:15) We develop benchmarks to measure advanced AI capabilities

(05:54) We provide independent evaluations of AI models

(06:43) We provide consultations and commissioned research

(10:23) What we are not

(10:57) We are not an AI development company

(11:54) We are not an AI policy think tank

(12:58) We are not a company incubator

(14:09) Closing words

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

I originally wrote this just for myself but thought I'd share it in case others find it useful. Sharing rough notes in the spirit of not letting perfect be the enemy of the good. This was written in early May 2025.

In this post I collect evidence and forecasts regarding recent and future trends in software and compute progress. I also make my own forecasts which inform an updated version of AI 2027's timelines forecast (more to come on these and further updates). These forecasts are rough and uncertain.

I especially build on this post.

Evidence and forecasts regarding the recent rate of software progress (and share vs. compute)

Compute efficiency increase at fixed performance

2.5-2.8x/year

Average efficiency increase reported in the paper (compute needed to reach same performance halving every 8-9 [...]

The original text contained 6 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

When discussing AI control, we often talk about levels of AI capabilities where we think control can probably greatly lower risks and where we can probably estimate risks. However, I think it's plausible that an important application of control is modestly improving our odds of surviving significantly superhuman systems which are misaligned. This won't involve any real certainty, it would only moderately reduce risk (if it reduces risk at all), and it may only help with avoiding immediate loss of control (rather than also helping us extract useful work). In a sane world, we wouldn't be relying on control for systems which have significantly superhuman general purpose capabilities. Nevertheless, I think applying control here could be worthwhile. I'll discuss why I think this might work and how useful I think applying control in this way would be.

As capabilities advance, the level of risk increases if we hold [...]

---

Outline:

(02:49) (Limited) hopes for controlling superintelligence

(05:59) How much time can control buy?

(12:39) Conclusion

The original text contained 9 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

The AI Futures Project, the team behind AI 2027, has developed a tabletop exercise (TTX) to simulate AI takeoff where we facilitate people going through an AI 2027-like scenario with different people taking different roles, like POTUS, China, or the frontier AI company CEO. Reviews have been positive and people have found it useful for building their own models of how AI will go.

We’re interested in developing an online version of this game, with the goal of eventually getting millions of people to play through takeoff and better understand the dynamics and risks involved. The other game characters could be driven by LLMs so we don’t need to facilitate it. The online TTX could be based on our scenario and background research alongside LLMs and various game elements.

To this end, we’re looking for a full-stack developer with game design and UX experience and a deep understanding [...]

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

It's hard to imagine doing web research without using LLMs. Chatbots may be the first thing you turn to for questions like: What are the companies currently working on nuclear fusion and who invested in them? What is the performance gap between open and closed-weight models on the MMLU benchmark? Is there really a Tesla Model H?

So which LLMs, and which "Search", "Research", "Deep Search" or "Deep Research" branded products, are best? How good are their epistemics, compared to if you did the web research yourself?

Last month we (FutureSearch) published Deep Research Bench (DRB), a benchmark designed to evaluate LLMs agents on difficult web research tasks using frozen snapshots of the internet. In this post, we're going to share the non-obvious findings, suggestions and failure modes that we think might be useful to anyone who uses LLMs with web search enabled.

tl;dr

• ChatGPT with o3 + [...]

---

Outline:

(01:09) tl;dr

(02:19) What We Did

(04:01) What We Found

(04:05) CommercialWeb Research Tools

(04:40) The Good

(06:04) The Mediocre

(06:33) The Bad

(07:06) Regular vs. Deep Research Mode

(07:55) Using LLMs With Agents via the API

(10:42) Open vs. Closed Models

(11:22) Thinking vs. Non-Thinking Models

(11:52) Epistemic concerns with LLM-powered web research tools

(12:59) Conclusions

The original text contained 2 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

Hillel... would say: "If I am not for me, who will be for me? And If I am just for myself, what am I? And if not now, when?"

Pirkei Avot 1,14

That last question has usually been interpreted as a rhetorical one. If you're not going to do it now, you'll never get around to it!

But this is clearly a bad principle to live by. If you start cooking too early, the food will be cold when you want to eat it. If you pick a fruit before it's ripe it will be sour. When you are exhausted the quality of your work goes down. And when you have many things to do, you have to prioritise which comes first.

There is a season and a time for everything...

Ecclesiastes 3,1

Perhaps we would do better interpreting it literally. It's fine to push things off, but first [...]

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

I think more people should say what they actually believe about AI dangers, loudly and often. Even if you work in AI policy.

I’ve been beating this drum for a few years now. I have a whole spiel about how your conversation-partner will react very differently if you share your concerns while feeling ashamed about them versus if you share your concerns as if they’re obvious and sensible, because humans are very good at picking up on your social cues. If you act as if it's shameful to believe AI will kill us all, people are more prone to treat you that way. If you act as if it's an obvious serious threat, they’re more likely to take it seriously too.

I have another whole spiel about how it's possible to speak on these issues with a voice of authority. Nobel laureates and lab heads and the most cited [...]

The original text contained 2 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.