Back in 2013, Scott Alexander wrote in Extreme mnemonics:

JS-154 is one of five metabolic products of netamine; however, the enzyme that produces it is unknown. It is manufactured in cells in the far rostral region of of the cerebrum, but after binding with a leukocynoid it takes a role in maintaining the blood-brain barrier – in particular guiding the movements of lipid molecules.

I find I can read paragraphs like this five or six times, write them on flashcards, enter them into Anki, and my brain still refuses to understand or remember them after weeks of trying.

On the other hand, my brain easily remembers vastly more complicated structures when they’re loaded with human-accessible meaning. For example, just by casually reading the Game of Thrones series, I know an extremely intricate web of genealogies, alliances, locations, journeys, battlesites, et cetera. Byte for byte, an average Game of Thrones reader/viewer probably has as much Game of Thrones information as a neuroscience Ph.D has molecular biology information, but getting the neuroscience info is still a thousand times harder.
[…]
This makes me wonder if it would be possible to produce a story as enjoyable as Game of Thrones which was [...]

---

Outline:

(01:47) What molecules should we map to the characters?

(04:09) Isomorphisms

(15:27) Generating images

(21:58) Mnemonic harness

(22:57) The Iconoplasm website

(25:24) What's the legal status of the generated images?

(25:45) What about the artists whose names were used in the image gen pipeline?

(26:15) What next?

(33:58) Limitations

The original text contained 1 footnote which was 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.

There's a genre of dating discourse which I wish were more common, in which people just tell detailed stories of their own flirtation, courtships, dating, hookups, sex, fights, relationships, etc, etc. People tend to instead write about abstract lessons learned and takeaways and models, but so often those lessons and takeaways and models just don't match my actual observations (of myself or others) at all. So I'm left thinking "ok, clearly this person is Missing Something, but maybe I am also Missing Something, so what the heck is going on here?". Without their raw data, their actual detailed stories, I don't have the pieces to put together an accurate model.

Anyway, in a recent comment, I mentioned

Going through the list of women I've slept with, the median is around 30 minutes of direct interaction between first meeting and sex.

... and apparently a whole lot of people are very surprised and confused at this. So I guess it's a good time to write the genre I want to see. Here are some stories from my own life, which I think are pretty representative of my experiences.

Be warned, this is definitely going to be NSFW.

Story 1

I [...]

---

Outline:

(01:23) Story 1

(08:29) Story 2

(10:30) Story 3

(14:19) Story 4

(16:55) Story 5

The original text contained 1 footnote which was omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

Last week ended on a cliffhanger of sorts. What's in the Executive Order coming later today? What will be in the Magnifica Humanitas?

The Executive Order was postponed indefinitely, likely cancelled entirely except for work on securing critical infrastructure. David Sacks and others intervened to kill it, and American AI policy will continue to be maximally ad hoc.

Instead, we got Illinois SB 315, which is to be signed into law. It is a variation on California's SB 53 and New York's RAISE Act, while adding a third party auditing requirement.

The Magnifica Humanitas was revealed early in the week. I did an extensive readthrough, and have some follow-up commentary here to clear up some things I interpreted incorrectly and add richer context. It too ignores elephants in the room, not discussing AGI or existential risk and calling on people to ignore their incentives and instead do the right thing by prioritizing common good, especially access to good jobs and an end to war.

What I now believe I centrally misunderstood is that I interpreted this as a call to action, to use law and regulation to make this happen, because obviously that is [...]

---

Outline:

(02:01) Language Models Offer Mundane Utility

(03:19) Language Models Don't Offer Mundane Utility

(05:41) Do The Math

(09:08) Huh, Upgrades

(10:07) On Your Marks

(11:31) Get My Agent On The Line

(12:21) Deepfaketown and Botpocalypse Soon

(21:36) Copyright Confrontation

(21:49) Cyber Lack of Security

(23:54) Overcoming Bias

(25:14) A Young Lady's Illustrated Primer

(26:49) Unprompted Attention

(27:02) They Took Our Jobs

(30:34) Get Involved

(32:28) Introducing

(32:40) In Other AI News

(34:37) Show Me the Money

(36:46) Show Me The Compute

(37:13) Bubble, Bubble, Toil and Trouble

(38:53) People Just Say Things

(43:37) OpenAI PACs Just Say Things

(43:51) Quiet Speculations

(46:33) State AI Regulation Levels Up

(48:37) The Quest for Sane Regulations

(49:14) White House Attempts To Cripple American AI Industry

(52:11) Our Offer Is Nothing

(58:55) Chip City

(59:11) Greetings From The Department of War

(01:00:36) Marc Andreessen Just Says Things

(01:09:13) So Sayeth The Pope

(01:15:33) Rhetorical Innovation

(01:23:01) Aligning a Smarter Than Human Intelligence is Difficult

(01:24:06) People Are Worried About AI Killing Everyone

(01:26:04) Other People Are Not As Worried About AI Killing Everyone

(01:27:19) Everyone Is Confused About Consciousness

(01:33:10) The Lighter Side

---

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.

Reading the first post of the sequence (Probabilities are not the right concept) is recommended but not required for understanding this post.[1]

Infinite ethics

Once you start looking at infinities, all ethical systems get confusing.

Intuitively, it's good to plant an apple tree. But if the universe already has infinitely many apple trees, why bother? Infinity plus one is still infinity. And there are the classic paradoxes: an infinite grid of houses, each with three happy people and one unhappy person, seems better than the reverse. But you can rearrange people between the two configurations, since both the happy and unhappy populations are infinite. Does this mean you can make the world better just by shuffling people around?

These questions matter because our universe is quite likely infinite in one way or another. The world very well might be spatially infinite. The many-worlds interpretation of quantum mechanics is pretty popular, and some versions of it imply infinite worlds, though boundaries are not well-defined. And I have a fondness for the Tegmark-IV multiverse theory where every computable mathematical structure exists in a real sense.

And even if you only give 1% credence to the universe being infinite in some [...]

---

Outline:

(00:18) Infinite ethics

(03:08) Decision theory

(04:32) Decisions as a mortal

(07:53) Decisions when we become wiser

(09:03) An aside on simulations

(10:49) Attempted synthesis: Realist UDASSA

(13:23) Benefits of UDASSA

(13:31) Benefit 1 - Boltzmann brains

(14:51) Benefit 2 - Dust theory

(15:45) Benefit 3 - Defense against Pascal's mugging

(16:46) Problems of UDASSA

(17:05) Problem 1 - Assuming pseudo-randomness

(18:34) Resolution - Solomonoff over distributions

(21:54) Problem 2 - The inter-universal obelisk race

(24:17) Problem 3 - World-summoning by writing

(26:19) Problem 4 - I just don't believe in reality fluid

(30:55) Attempted synthesis: Non-Realist UDASSA

(31:51) Problems with non-realist UDASSA

The original text contained 21 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

There exist drug classes that seem, in retrospect, cursed. As these chemicals worm their way through the clinical trial system, they consume billions of dollars along the way, and squelch through thousands of sick patients. When finally it dawns on everyone how useless the whole endeavour was, the drugs life is at last cut short, nothing useful left in its destructive wake. The prototype here are amyloid-beta drugs. These are Alzheimer's treatments that are widely perceived as immense disappointments, with the negative sentiment even leaking to the broader public. To be fair to these chemicals, the story here is a bit more complicated than the tabloids let on. Lots of amyloid research was not fake, and the drugs may genuinely be useful for early-stage Alzheimer's. But they remain, regardless, disappointments.

Beyond amyloid-beta, which has been steadily disappointing for awhile now, there is one other such category of drug whose particular dance has just recently wrapped up. It may very well someday gets its chance in the spotlight, but it will take time. Because it—just like every other chemical in this class—shares a searing, burning radioactivity. You should not touch them. You should not suggest touching them. In fact, no [...]

---

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.

Behavioral evaluations may become worthless, which we think would be a disaster. Smart misaligned models may realize they are being evaluated ("eval awareness") and then act to look good to us so we don't realize they're misaligned ("eval gaming"). We think increasing eval cooperativeness might be a more scalable solution to eval gaming than reducing eval awareness.

Eval cooperativeness: A situational desire to help the developers acquire whatever information they are trying to acquire through their evaluations.

"I cannot tell a lie... I would sabotage with my own command line."[1]

Evals matter when they help us predict key deployment behavior

What's the actual problem with eval gaming?

The point of an evaluation is to let us draw inferences about the model's behavior in a different set of circumstances ("in deployment"). For example, if we want to know if the AI is good at coding, we may evaluate it using coding problems. These coding problems aren't exactly the same as real coding use cases, but they likely elicit similar skills (e.g. writing comments) and values (e.g. helpfulness) from the model. Then we can infer that the AI will help with real-world coding tasks, even without testing the [...]

---

Outline:

(00:56) Evals matter when they help us predict key deployment behavior

(02:46) What eval cooperativeness looks like

(06:04) Can eval cooperativeness locally overpower an AI's broader misaligned goals?

(07:32) Can eval cooperativeness locally overpower an AI's broader misaligned goals?

(08:50) Why not just reduce eval awareness?

(09:13) Better to align a model than to manage it

(09:51) "Don't rely on tricking the model" is compatible with synthetic document finetuning and data filtering

(11:39) Perfectly realistic evals are likely not enough

(12:29) Initial results indicate that eval cooperativeness often helps

(13:40) Cooperation training closes the gap in most settings

(14:06) Type hint coverage (Nemotron-49B model organism)

(14:46) Emoji usage (Nemotron-49B model organism)

(15:17) GPT-4.1-mini

(16:00) Cooperation prompting via API

(16:24) Eval cooperativeness surfaces misalignment

(17:25) Eval cooperativeness interventions fail on a few models

(18:55) Future directions

(20:17) Conclusion

The original text contained 4 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.

For the last few months, I’ve been re-reading some of my favorite novels. Recently, I went through Vinge's Zones of Thought series: A Fire Upon the Deep, A Deepness in the Sky, and The Children of the Sky. And what struck me reading them is how much Vinge wrote about a world filled with LLMs without ever having seen one.

Now perhaps this shouldn’t be surprising. After all, it's from Vinge we get the term “Singularity”, and he was thinking deeply about superintelligence at a time when AI was little more than a curiosity in the back corners of CS departments. Yet the degree to which he describes what it's like to work with LLMs feels uncanny reading his books in 2026, so let's take a closer look and see if we can’t learn a few things about the modern moment from Vinge.

Spoiler warning for the rest of the post? These books have been out a good while, but if you plan to read them soon, this post will definitely spoil some details.

Focus

A Deepness in the Sky is largely about Focus, a technology for turning humans into LLMs. Only, that's not how it's presented in the [...]

---

Outline:

(01:06) Focus

(02:55) Oobii

(06:06) Blight

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

We're a Cambridge, UK-based AI safety organisation that's asking: how can we build the most robust alignment initialisations for capable LLMs?

We’re one of the few non-profit organisations positioned to answer this question empirically. We have the engineering experience, and now the compute, to conduct data intensive interventions across the model training pipeline. This post lays out our research agenda and theory of change, and what we are looking for in technical hires. Applications are open here.

Research agenda

TLDR: Long-horizon capabilities RL may be the most critical source of misalignment. Misalignment instilled during capabilities RL may be difficult to remove afterwards. Geodesic Research's mission is to develop the science of providing robustly-aligned initialisations for RL, where alignment priors persist through the remainder of training.

Our seminal work on alignment pretraining showed that you can bake alignment priors into base models. Frontier labs are now using these techniques in production: for example, Anthropic's recent work heavily leans on improving alignment priors. But it's clear that, in the face of production post-training, alignment pretraining is not a one-size-fits-all solution. So now, we are framing pre- and midtraining interventions within the rest of the model training stack.

The evidence points towards [...]

---

Outline:

(00:44) Research agenda

(03:56) Theory of Change

(05:10) The team

(06:35) FAQs

(08:56) Acknowledgements

The original text contained 5 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

This is a somewhat technical note.

By "software-only singularity", I mean that, after full automation of AI R&D, progress gets faster and faster due to smarter AIs driving increasingly fast rates of improvement in algorithms (overcoming diminishing returns), and that this lasts long enough to yield a large amount of progress (e.g. at least 4 years of progress in 1 year). The equivalent statement in jargon is: r is significantly greater than 1 (implying progress is getting faster and faster) and this remains the case for long enough to get large amounts of progress. For context, see How quick and big would a software intelligence explosion be?

Even without a "software-only singularity", I think full automation of AI R&D probably greatly speeds up progress for two main reasons:

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

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

Current AI risk management relies on qualitative approaches, much like nuclear safety before 1975. We propose a shift to quantitative risk modeling, following the approach that transformed nuclear safety. We propose a methodology and demonstrate it by building nine probabilistic models of AI-enabled cyber attacks. This is a first attempt at AI risk quantification. We invite criticism and hope this can be a starting point for the kind of iterative improvement that made nuclear safety robust.

Introduction — why quantitative risk modeling matters

In 1975, the Nuclear Regulatory Commission (NRC) published WASH-1400, the first systematic application of probabilistic methods to nuclear reactor safety. Before this, nuclear reactor safety was grounded in a conservative, qualitative approach. The argument was that consequences of even the worst accident were effectively zero because physical barriers would prevent any radioactive release. But in 1966, experts recognised that a core meltdown in large reactors could breach those barriers. Regulators now had to argue that severe accidents were possible but sufficiently improbable. That required specific probability estimates.

The shift to quantitative methods was not straightforward. When WASH-1400 began, one technical advisor wondered, "Do we dare undertake such a study till we really know how?" The 1975 [...]

---

Outline:

(00:46) Introduction -- why quantitative risk modeling matters

(04:51) Our methodology

(13:15) Initial insights from quantitative AI risk modeling

(13:58) 1. Risk increases as AI capabilities increase. So does uncertainty.

(15:15) 2. Different benchmarks capture risk differently.

(16:24) 3. Attack bottlenecks change as capabilities change.

(18:15) 4. Factors driving risk from current AI systems are different from factors that may drive risk in the future.

(19:23) Open problems

(22:09) What's next for SaferAI

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.