In 1654, a Jesuit polymath named Athanasius Kircher published Mundus Subterraneus, a comprehensive geography of the Earth's interior. It had maps and illustrations and rivers of fire and vast subterranean oceans and air channels connecting every volcano on the planet. He wrote that “the whole Earth is not solid but everywhere gaping, and hollowed with empty rooms and spaces, and hidden burrows.”. Alongside comments like this, Athanasius identified the legendary lost island of Atlantis, pondered where one could find the remains of giants, and detailed the kinds of animals that lived in this lower world, including dragons. The book was based entirely on secondhand accounts, like travelers tales, miners reports, classical texts, so it was as comprehensive as it could’ve possibly been.

But Athanasius had never been underground and neither had anyone else, not really, not in a way that mattered.

Today, I am in San Francisco, the site of the 2026 J.P. Morgan Healthcare Conference, and it feels a lot like Mundus Subterraneus.

There is ostensibly plenty of evidence to believe that the conference exists, that it actually occurs between January 12, 2026 to January 16, 2026 at the Westin St. Francis Hotel, 335 Powell Street, San Francisco [...]

---

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.

I spent a few hundred dollars on Anthropic API credits and let Claude individually research every current US congressperson's position on AI. This is a summary of my findings.

Disclaimer: Summarizing people's beliefs is hard and inherently subjective and noisy. Likewise, US politicians change their opinions on things constantly so it's hard to know what's up-to-date. Also, I vibe-coded a lot of this.

Methodology

I used Claude Sonnet 4.5 with web search to research every congressperson's public statements on AI, then used GPT-4o to score each politician on how "AGI-pilled" they are, how concerned they are about existential risk, and how focused they are on US-China AI competition. I plotted these scores against GovTrack ideology data to search for any partisan splits.

1. AGI awareness is not partisan and not widespread

Few members of Congress have public statements taking AGI seriously. For those that do, the difference is not in political ideology. If we simply plot the AGI-pilled score vs the ideology score, we observe no obvious partisan split.

There are 151 congresspeople who Claude could not find substantial quotes about AI from. These members are not included on this plot or any of the plots which follow.

[...]

---

Outline:

(00:46) Methodology

(01:12) 1. AGI awareness is not partisan and not widespread

(01:56) 2. Existential risk is partisan at the tails

(02:51) 3. Both parties are fixated on China

(04:02) 4. Who in Congress is feeling the AGI?

(11:10) 5. Those who know the technology fear it.

(12:54) Appendix: How to use this data

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.

Lightcone is hiring! We build beautiful things for truth-seeking and world-saving.

We are hiring for three different positions: a senior designer, a campus manager, and a core team generalist. This is the first time in almost two years where we are actively hiring and trying to grow our team!

Senior Designer

When we are at our best, I think we produce world-class design. AI 2027 was I think a great design achievement, so is much of LessWrong.com itself. I also think on a product and business level, making things beautiful and intuitive and well-crafted is crucial. I like some of Patrick Collison's thinking on this:

If Stripe is a monstrously successful business, but what we make isn’t beautiful, and Stripe doesn’t embody a culture of incredibly exacting craftsmanship, I’ll be much less happy. I think the returns to both of those things in the world are really high. I think even beyond the pecuniary or financial returns, the world's just uglier than it needs to be… One can do things well or poorly, and beauty is not a rivalrous good.

My intuition is that more of Stripe's success than one would think is downstream of the [...]

---

Outline:

(00:45) Senior Designer

(03:15) Campus Operations Co-Lead

(04:51) Core-team generalist

(07:44) Some more thoughts on working at Lightcone

---

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.

(Content warnings: dubious math, quantum immortality, nuclear war)

Normal people make New Year's resolutions.

People on the internet love to make resolutions for November.

So, for the entire month of November, 41 people, myself included, set out to publish a post every day, as part of a writing residency called Inkhaven. On the final day, I couldn’t resist the temptation to pull out a prank and make sure no one appeared to have failed.

(Note that I want to maintain everything as plausibly deniable; everything in this post might or might not have happened.)

From Mahmoud, another resident:

Inkhaven is the name of the 30 day blogging retreat that I went on this month. The rules were: come meet your blogging heroes, post 500 words a day, online, by midnight or we kick you out.

I made it to the final day, but, on purpose, this last post will be fewer than 500 words. My reasons for this are kind of silly, but mostly I think it would be more fun if at least one person failed the program. I’m also writing post as a stream of consciousness, unsure where it will go. Maybe I’ll come up with a [...]

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.

Claude Code and Cowork are growing so much that it is overwhelming Anthropic's servers. Claude Code and Cowork news has for weeks now been a large portion of newsworthy items about AI.

Thus, at least for now, all things Claude Code and Cowork will stop appearing in the weekly updates, and will get their own updates, which might even be weekly.

Google offered us the new Universal Commerce Protocol, and gives us its take on Personalized Intelligence. Personalized Intelligence could be a huge deal if implemented correctly, integrating the G-Suite including GMail into Gemini, if they did a sufficiently good job of it. It's too early to tell how well they did, and I will report on that later.

Table of Contents

---

Outline:

(00:53) Language Models Offer Mundane Utility

(04:00) Huh, Upgrades

(07:06) Comparative Advantage

(07:41) Overcoming Bias

(09:09) Choose Your Fighter

(10:14) Get My Agent On The Line

(11:23) Deepfaketown and Botpocalypse Soon

(13:49) Fun With Media Generation

(17:01) A Young Lady's Illustrated Primer

(18:41) They Took Our Jobs

(21:47) Autonomous Killer Robots

(24:34) Get Involved

(25:10) Introducing

(28:03) In Other AI News

(29:45) Show Me the Money

(31:02) Quiet Speculations

(33:56) The Quest for Sane Regulations

(36:10) China Proposes New Regulations On AI

(40:51) Chip City

(44:19) The Week in Audio

(48:17) Ghost in a Jar

(49:29) Rhetorical Innovation

(59:00) Aligning a Smarter Than Human Intelligence is Difficult

(59:21) People Are Worried About AI Killing Everyone

---

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.

These are my personal views. Thank you to Ryan Greenblatt, Holden Karnofsky, and Peter Wildeford for useful discussions. The bad takes are my own.

When deciding how much to spend on mitigating a vulnerability that a competent scheming AI might exploit, you might be tempted to use E[damages | AI decides to take advantage of the vulnerability] to decide how important mitigating that vulnerability is.

But this misses how a strategic scheming AI might decide to not take advantage of some vulnerabilities, because being caught could reduce how likely it would be to take advantage of better future opportunities. So instead you might want to focus on making net-sabotage-value negative:

Focus on mitigating vulnerabilities that are net-positive from the scheming AI's perspective, that is if the net-sabotage-value E[damages - value to humans from catching the AI | AI decides to take advantage of the vulnerability] is positive. Down-weight the importance of those for which the net-sabotage-value is negative.

Here is an illustration that paints a very simplified picture of the situation:

In practice, you should often be uncertain about whether a vulnerability has negative net-sabotage-value and whether AIs know this has negative net-sabotage-value, and mitigating sabotage risk from [...]

---

Outline:

(02:34) The argument

(04:32) Main limitations

(05:48) Potential implication 1: Simple monitoring might be enough against early competent scheming AIs

(06:23) Toy case study: Default mitigations are all you need against superhuman coder persuasion?

(08:17) Potential implication 2: Your willingness to pay for control might be much higher if alignment becomes easy

(09:52) Toy case study: Superhuman coder persuasion mitigations can become more important when you get good alignment news

(11:05) Potential implication 3: Risk management advice might apply less than you think

(12:34) Potential implication 4: You might in principle not want to make AI control mitigations too good

(15:31) More generally, unconditional payments to AIs might be more tractable than you think

(16:57) Should you ignore negative net-sabotage-value vulnerabilities? (No)

(18:38) Toy case study: Advanced mitigations are still useful against non-scheming superhuman coder persuasion

(19:31) Final note: don't use this argument for wildly superintelligent AIs

---

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.

Since artificial superintelligence has never existed, claims that it poses a serious risk of global catastrophe can be easy to dismiss as fearmongering. Yet many of the specific worries about such systems are not free-floating fantasies but extensions of patterns we already see. This essay examines thirteen distinct ways artificial superintelligence could go wrong and, for each, pairs the abstract failure mode with concrete precedents where a similar pattern has already caused serious harm. By assembling a broad cross-domain catalog of such precedents, I aim to show that concerns about artificial superintelligence track recurring failure modes in our world.

This essay is also an experiment in writing with extensive assistance from artificial intelligence, producing work I couldn’t have written without it. That a current system can help articulate a case for the catastrophic potential of its own lineage is itself a significant fact; we have already left the realm of speculative fiction and begun to build the very agents that constitute the risk. On a personal note, this collaboration with artificial intelligence is part of my effort to rebuild the intellectual life that my stroke disrupted and hopefully push it beyond where it stood before.

Section 1: Power Asymmetry [...]

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

A multitude of forecasts discuss how powerful AIs might quickly arise and influence the world within the coming decades. I’ve run a variety of tabletop exercises created by the authors of AI 2027, the most famous such forecast, which aim to help participants understand the dynamics of worlds similar to AI 2027's. At one point in the exercise participants must decide how persuasive the AI models will be, during a time when AIs outperform humans at every remote work task and accelerate AI R&D by 100x. I think most participants underestimate how persuasive these AIs are. By default, I think powerful misaligned AIs will be extremely persuasive, especially absent mitigations.

Imagine such an AI wants to convince you, a busy politician, that your longtime advisor is secretly undermining you. Will you catch the lie out of all the other surprisingly correct advice the AI has given you?

The AI tells you directly about it, of course, in that helpful tone it always uses. But it's not just the AI. Your chief of staff mentioned that the advisor has been surprisingly absent for the last week, spending a surprisingly small amount of time at work and deferring quite a lot [...]

---

Outline:

(06:58) AI is being rapidly adopted, and people are already believing the AIs

(10:12) Humans believe what they are incentivized to, and the incentives will be to believe the AIs.

(13:52) Will you hear the truth?

(18:54) What about mitigations?

(21:01) What if the AI is not worst-case?

(22:36) So what?

The original text contained 3 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.

And once they take our jobs, will we be able to find new ones? Will AI take those too?

Seb Krier recently wrote an unusually good take on that, which will center this post.

I believe that Seb is being too optimistic on several fronts, but in a considered and highly reasonable way. The key is to understand the assumptions being made, and also to understand that he is only predicting that the era of employment optimism will last for 10-20 years.

By contrast, there are others that expect human employment and even human labor share of income to remain robust indefinitely, no matter the advances of AI capabilities, even if AI can do a superior job on all tasks, often citing comparative advantage. I will centrally respond to such claims in a different future post.

So to disambiguate, this post is about point #2 here, but I also assert #1 and #3:

---

Outline:

(01:39) Normality Is Magic

(03:20) Welcome To The Cyborg Era

(06:09) The Cyborg Era Will End Within Our Lifetimes

(07:46) Transitional Unemployment In The Cyborg Era Will Be High

(08:51) Congratulations, Many Of You Are Literally A Human

(09:34) You've Got Taste

(11:33) The Revolution of Rising AI Capabilities

(11:56) The Revolution of Rising Expectations

(12:25) Proceed With Caution

(13:13) The Ostrich Take

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

Audio note: this article contains 37 uses of latex notation, so the narration may be difficult to follow. There's a link to the original text in the episode description.

Authors: Riya Tyagi, Daria Ivanova, Arthur Conmy, Neel Nanda

Riya and Daria are co-first authors. This work was largely done during a research sprint for Neel Nanda's MATS 9.0 training phase.

🖥️ Deployment code

⚙️ Interactive demo

TL;DR

• We believe that more research effort should go into studying many chains of thought collectively and identifying global patterns across the collection.
• We argue this direction is important and tractable; to start, we focus on a simpler setting: analyzing a collection of CoTs generated with a fixed model on a fixed prompt. We introduce two methods—semantic step clustering and algorithmic step clustering—that group recurring patterns across reasoning traces at different levels of abstraction.
• Semantic step clustering leverages the fact that traces for the same prompt often repeat each other in slightly different ways. By clustering similar sentences like “1+2=3” and “2+1=3”, we compress a large set of CoTs into a detailed but messy reasoning graph.
  • We evaluate semantic step clustering on a downstream task: predicting a model's output [...]

---

Outline:

(00:49) TL;DR

(03:03) Introduction

(06:59) Method 1: Semantic step clustering

(07:04) Building the semantic graph

(08:47) Chunking

(09:31) Clustering

(11:11) The issue of semantic drift

(12:30) Interlude: CoTs are chaotic!

(14:25) Sanity check: viewing our clusters

(15:29) Downstream task: predictive power

(16:54) Deriving predictions from our graph

(19:30) Predictive power baselines

(22:44) Results

(27:00) Method 2: Algorithmic step clustering

(27:58) Building the algorithm graph

(29:35) Sanity check: studying sample problems

(29:48) Tree structure

(30:45) Convergent structure

(31:37) Qualitative insights

(31:41) Models reuse strategies frequently

(33:24) Different models solve problems differently

(34:31) Takeaways and future work

(37:26) Acknowledgements

(37:41) Contribution Statement

(38:55) Appendix

(38:58) Chunking

(39:47) Clustering

(39:55) Stage 1: Agglomerative Clustering with Semantic Embedding Model

(41:01) Stage 2: LLM refinement via the Leiden Algorithm

(42:28) Predictive power

(42:31) Semantic graph prompts

(43:00) Method ablation study

(44:31) A comparison of three Gemini baselines:

(45:31) Gemini predictive power baseline prompts

(45:51) Gemini no rollout baseline (0 rollouts)

(46:13) Gemini 20 rollout baseline

(46:35) Gemini full context baseline (1000 rollouts)

(46:58) Footnotes

The original text contained 15 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.