Forecaster perspectives

Sentinel forecasters in aggregate assess as “83% true” (65% to 100%) the statement that the Trump administration has disobeyed the Supreme Court so far. On the one hand, the Trump administration has done nothing to “facilitate” bringing Abrego García back to the US, and this White House's tweet shown below clearly indicates that it won’t try to; on the other hand, it could yet still do so when the current news cycle subsides. On April 17th, forecasters estimated a 66% chance (50% to ~100%) that Abrego García will not be brought back to the US within the next 60 days (by June 17).

Data on the number of deportations in past administrations are available but can be difficult to compare between administrations; some up-to-date sources include refusals at the border due to Covid (“Title 42 expulsions”) under Biden, but other data have been discontinued. Perhaps a good [...]

---

Outline:

(00:12) Forecaster perspectives

(03:16) Summary of events

(03:31) The Abrego García US Supreme Court case.

(05:41) Developments since the Abrego García Supreme Court case

(08:07) Next steps for the court case

(09:32) Plans for further deportations

(11:29) The Insurrection Act

(12:43) Implications and recommendations

---

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.

Back in the 1990s, ground squirrels were briefly fashionable pets, but their popularity came to an abrupt end after an incident at Schiphol Airport on the outskirts of Amsterdam. In April 1999, a cargo of 440 of the rodents arrived on a KLM flight from Beijing, without the necessary import papers. Because of this, they could not be forwarded on to the customer in Athens. But nobody was able to correct the error and send them back either. What could be done with them? It's hard to think there wasn’t a better solution than the one that was carried out; faced with the paperwork issue, airport staff threw all 440 squirrels into an industrial shredder.

[...]

It turned out that the order to destroy the squirrels had come from the Dutch government's Department of Agriculture, Environment Management and Fishing. However, KLM's management, with the benefit of hindsight, said that [...]

---

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 Culture Expects Self-Justification

I really like David Chapman's explication of what he calls “reasonableness” and “accountability.”

At least in the culture he and I live in, one is constantly called to account for one's behavior. At any moment, one may be asked “what are you doing?” or “why did you do that?” And one is expected to provide a reasonable answer.

What's a reasonable answer?

It's not a rigorous, beyond-a-shadow-of-a-doubt proof that at this very moment one is engaging in “optimal” or “morally correct” or “maximally virtuous” behavior. That would indeed be an unfair expectation! Nobody could possibly provide such a justification on demand.

What people actually expect is a more-or-less plausible-sounding account that makes your actions sound understandable, relatable, and okay.

You have a lot of leeway to present yourself in a favorable light. Not infinite leeway — if you’re caught shoplifting, it probably won’t be [...]

---

Outline:

(00:14) Our Culture Expects Self-Justification

(01:43) Asking for favors

(02:52) Denying requests

(04:09) Defending Your Reasoning

(05:24) Demanding Justification is a Cheap Heuristic

(06:24) It's Good To Have A Self-Serving Narrative

(07:51) Isn't This Just Rationalization?

(10:23) Everybody Gets To Stop Somewhere

(11:52) Don't Knock Mere Feeling Good

---

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.

Audio note: this article contains 36 uses of latex notation, so the narration may be difficult to follow. There's a link to the original text in the episode description.

Our posts on natural latents have involved two distinct definitions, which we call "stochastic" and "deterministic" natural latents. We conjecture that, whenever there exists a stochastic natural latent (to within some approximation), there also exists a deterministic natural latent (to within a comparable approximation). We are offering a $500 bounty to prove this conjecture.

Some Intuition From The Exact Case

In the exact case, in order for a natural latent to exist over random variables _X_1, X_2_, the distribution has to look roughly like this:

Each value of _X_1_ and each value of _X_2_ occurs in only one "block", and within the "blocks", _X_1_ and _X_2_ are independent. In that case, we can take the (exact) natural latent [...]

---

Outline:

(00:51) Some Intuition From The Exact Case

(02:16) Approximation Adds Qualitatively New Behavior

(02:59) The Problem

(03:02) Stochastic Natural Latents

(04:04) Deterministic Natural Latents

(05:26) What We Want For The Bounty

(06:43) Why We Want This

---

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.

This seemed like a good next topic to spin off from monthlies and make into its own occasional series. There's certainly a lot to discuss regarding crime.

What I don’t include here, the same way I excluded it from the monthly, are the various crimes and other related activities that may or may not be taking place by the Trump administration or its allies. As I’ve said elsewhere, all of that is important, but I’ve made a decision not to cover it. This is about Ordinary Decent Crime.

Table of Contents

---

Outline:

(00:38) Perception Versus Reality

(04:21) The Case Violent Crime is Up Actually

(05:31) Threats of Punishment

(06:23) Property Crime Enforcement is Broken

(11:33) The Problem of Disorder

(13:59) Extreme Speeding as Disorder

(15:17) The Fall of Extralegal and Illegible Enforcement

(16:42) In America You Can Usually Just Keep Their Money

(18:53) Police

(25:56) Probation

(28:46) Genetic Databases

(30:56) Marijuana

(36:20) The Economics of Fentanyl

(37:26) Enforcement and the Lack Thereof

(41:20) Jails

(45:20) Criminals

(45:55) Causes of Crime

(46:33) Causes of Violence

(47:52) Homelessness

(48:44) Yay Trivial Inconveniences

(49:25) San Francisco

(54:24) Closing Down San Francisco

(55:47) A San Francisco Dispute

(59:31) Cleaning Up San Francisco

(01:03:20) Portland

(01:03:30) Those Who Do Not Help Themselves

(01:05:30) Solving for the Equilibrium (1)

(01:10:30) Solving for the Equilibrium (2)

(01:10:57) Lead

(01:12:36) Law & Order

(01:13:16) Look Out

---

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.

AI 2027 lies at a Pareto frontier – it contains the best researched argument for short timelines, or the shortest timeline backed by thorough research[1]. My own timelines are substantially longer, and there are credible researchers whose timelines are longer still. For this reason, I thought it would be interesting to explore the key load-bearing arguments AI 2027 presents for short timelines. This, in turn, allows for some discussion of signs we can watch for to see whether those load-bearing assumptions are bearing out.

To be clear, while the authors have short timelines, they do not claim that ASI is likely to arrive in 2027[2]. But the fact remains that AI 2027 is a well researched argument for short timelines. Let's explore that argument.

(In what follows, I will mostly ignore confidence intervals and present only median estimates; this is a gross oversimplification of the results presented in the [...]

---

Outline:

(01:17) Timeline to ASI

(06:57) AI 2027 Is Not Strong Evidence for AI in 2027

(08:12) Human-only timeline from SIAR to ASI

(10:46) Reasons The AI 2027 Forecast May Be Too Aggressive

(10:52) #1: Simplified Model of AI R&D

(12:23) #2: Amdahls Law

(14:51) #3: Dependence on Narrow Data Sets

(15:55) #4: Hofstadters Law As Prior

(16:46) What To Watch For

The original text contained 10 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.

Disclaimer: this post was not written by me, but by a friend who wishes to remain anonymous. I did some editing, however.

So a recent post my friend wrote has made the point quite clearly (I hope) that LLM performance on the simple task of playing and winning a game of Pokémon Red is highly dependent on the scaffold and tooling provided. In a way, this is not surprising—the scaffold is there to address limitations in what the model can do, and paper over things like lack of long-term context, executive function, etc.

But the thing is, I thought I knew that, and then I actually tried to run Pokémon Red.

A Casual Research Narrative

The underlying code is the basic Claude scaffold provided by David Hershey of Anthropic.[1] I first simply let Claude 3.7 run on it for a bit, making observations about what I thought might generate [...]

---

Outline:

(01:04) A Casual Research Narrative

(09:10) An only somewhat sorted list of observations about all this

(09:22) Model Vision of Pokémon Red is Bad. Really Bad.

(12:58) Models Cant Remember

(14:34) Spatial Reasoning? Whats that?

(15:22) A Grasp on Reality

(18:40) Costs

(19:13) Why do this at all?

(19:47) So which model is better?

(22:36) Miscellanea: ClaudePlaysPokemon Derp Anecdotes

The original text contained 6 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.

This post summarizes some of the research I have been doing for

TL;DR

Epigenetics and imprinting

This post assumes basic familiarity with epigenetics. The next paragraph

---

Outline:

(00:17) TL;DR

(00:56) Epigenetics and imprinting

(02:20) Why we need to know all imprinting regions and genes

(04:48) How imprinted genes work and how we measure them

(07:43) Have we identified all imprinted genes?

(12:39) Fixing paternal imprinting might be easier than fixing maternal imprinting

(13:15) PCR for methylated DNA

(17:11) Thoughts on the role of RNA in imprinting

(18:47) Some useful datasets and resources if someone else wants to look further into imprinting and epigenetics in the future

(20:57) References

The original text contained 11 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.

I’ve been thinking recently about what sets apart the people who’ve done the best work at Anthropic.

You might think that the main thing that makes people really effective at research or engineering is technical ability, and among the general population that's true. Among people hired at Anthropic, though, we’ve restricted the range by screening for extremely high-percentile technical ability, so the remaining differences, while they still matter, aren’t quite as critical. Instead, people's biggest bottleneck eventually becomes their ability to get leverage—i.e., to find and execute work that has a big impact-per-hour multiplier.

For example, here are some types of work at Anthropic that tend to have high impact-per-hour, or a high impact-per-hour ceiling when done well (of course this list is extremely non-exhaustive!):

---

Outline:

(03:28) 1. Agency

(03:31) Understand and work backwards from the root goal

(05:02) Don't rely too much on permission or encouragement

(07:49) Make success inevitable

(09:28) 2. Taste

(09:31) Find your angle

(11:03) Think real hard

(13:03) Reflect on your thinking

---

First published:

Source:

Narrated by TYPE III AUDIO.

Background: With the release of Claude 3.7 Sonnet, Anthropic promoted a new benchmark: beating Pokémon. Now, Google claims Gemini 2.5 Pro has substantially surpassed Claude's progress on that benchmark.

TL:DR: We don't know if Gemini is better at Pokémon than Claude because their playthroughs can't be directly compared.

The Metrics

Here are Anthropic's and Google's charts:

[1]Unfortunately these are using different x and y axes, but it's roughly accurate to say that Gemini has made it nearly twice as far in the game[2] now:

And moreover, Gemini has gotten there using approximately 1/3rd the effort! As of writing, Gemini's current run is at ~68,000 actions, while Claude's current run is at ~215,000 actions.[3][4]

So, sounds definitive, right? Gemini blows Claude out of the water.

The Agents' Harnesses

Well, when Logan Kilpatrick (product lead for Google's AI studio) posted his tweet, he gave an important caveat:

"next best model only [...]

---

Outline:

(01:13) The Metrics

(02:19) The Agents Harnesses

(05:13) The Agents Masters

(07:46) The Agents Vibes

(11:27) Conclusion

The original text contained 14 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.