We’re back with all the Claude that's fit to Code. I continue to have great fun with it and find useful upgrades, but the biggest reminder is that you need the art to have an end other than itself. Don’t spend too long improving your setup, or especially improving how you improve your setup, without actually working on useful things.

The Efficient Market Hypothesis

Odd Lots covered Claude Code. Fun episode, but won’t teach my regular readers much that is new.

Bradly Olsen at the Wall Street Journal reports Claude [Code and now Cowork are] Taking the AI World By Storm, and ‘Even Non-Nerds Are Blown Away.’

It is remarkable how everyone got the ‘Google is crushing everyone’ narrative going with Gemini 3, then it took them a month to realize that actually Anthropic is crushing everyone, at least among the cognoscenti with growing momentum elsewhere, with Claude Code and Claude Opus 4.5. People are realizing you can know almost nothing and still use it to do essentially everything.

Are Claude Code and Codex having a ‘GPT moment’?

Wall St Engine: Morgan Stanley says Anthropic's ClaudeCode + Cowork is dominating investor chatter and [...]

---

Outline:

(00:28) The Efficient Market Hypothesis

(01:57) Huh, Upgrades

(03:49) Obsidian

(04:33) New Tools

(04:57) Tool Search

(07:17) Out Of The Box

(08:48) Skilling Up

(12:01) The Art Must Have An End Other Than Itself Or It Collapses Into Infinite Recursion

(13:43) Safely Skip Permissions

(14:44) A Matter of Trust

(16:30) Code Versus Cowork

(18:14) Claude Cowork Offers Mundane Utility

(21:52) Claude Code Offers Mundane Utility

(23:32) Vibe Coding Requires Good Vibes

(24:06) Codex of Ultimate Vibing

(25:12) No Soup For You

(26:46) Server Overload

(28:17) Here's The Pitch

(29:12) 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.

"The person is an identity that emerges through relationship.... If we isolate the 'I' from the 'thou' we lose not only its otherness but also its very being; it simply cannot be without the other." -- John Zizioulas, Communion and Otherness

It is the third week of Anna's freshman year of high school.

Today, like many days, she's filled with anxiety about her place in the world. The anxiety is a sharp, piercing feeling in her chest, like she is a sphere spinning on the point of a cone. Twenty years from now, her remembering self will look back on these feelings as part of the precious journey of discovering herself; but right now, her experiencing self mostly just feels like shit.

Lunch is ending soon. She's about to get up. But she notices a notebook accidentally fall from a fellow student's backpack, as the other student hurriedly walks by. Anna almost raises her voice to alert the other girl -- but the notebook has flopped open on the floor, and Anna notices her own name, ANNA H, in big block letters inside the other girl's notebook.

What?

Curiosity gets the better of her. [...]

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

I have been reading Eric Drexler's writing on the future of AI for more than a decade at this point. I love it, but I also think it can be tricky or frustrating.

More than anyone else I know, Eric seems to tap into a deep vision for how the future of technology may work — and having once tuned into this, I find many other perspectives can feel hollow. (This reminds me of how, once I had enough of a feel for how economies work, I found a lot of science fiction felt hollow, if the world presented made too little sense in terms of what was implied for off-screen variables.)

One cornerstone of Eric's perspective on AI, as I see it, is a deep rejection of anthropomorphism. People considering current AI systems mostly have no difficulty understanding it as technology rather than person. But when discussion moves to superintelligence … well, as Eric puts it:

Our expectations rest on biological intuitions. Every intelligence we’ve known arose through evolution, where survival was a precondition for everything else—organisms that failed to compete and preserve themselves left no descendants. Self-preservation wasn’t optional—it was the precondition for everything else. We naturally [...]

---

Outline:

(02:38) Difficulties with Drexler's writing

(03:35) How to read Drexler

(05:05) What Drexler covers

(05:14) 1) Mapping the technological trajectory

(05:47) 2) Pushing back on anthropomorphism

(06:34) 3) Advocating for strategic judo

(07:21) The missing topics

(08:26) Translation and reinvention

(09:47) Pieces I'd be especially excited to see explored

The original text contained 5 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.

Read the constitution. Previously: 'soul document' discussion here.

We're publishing a new constitution for our AI model, Claude. It's a detailed description of Anthropic's vision for Claude's values and behavior; a holistic document that explains the context in which Claude operates and the kind of entity we would like Claude to be.

The constitution is a crucial part of our model training process, and its content directly shapes Claude's behavior. Training models is a difficult task, and Claude's outputs might not always adhere to the constitution's ideals. But we think that the way the new constitution is written—with a thorough explanation of our intentions and the reasons behind them—makes it more likely to cultivate good values during training.

In this post, we describe what we've included in the new constitution and some of the considerations that informed our approach.

We're releasing Claude's constitution in full under a Creative Commons CC0 1.0 Deed, meaning it can be freely used by anyone for any purpose without asking for permission.

What is Claude's Constitution?

Claude's constitution is the foundational document that both expresses and shapes who Claude is. It contains detailed explanations of the values we [...]

---

Outline:

(01:14) What is Claudes Constitution?

(03:26) Our new approach to Claudes Constitution

(04:59) A brief summary of the new constitution

(09:14) Conclusion

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

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

This is a personal post and does not necessarily reflect the opinion of other members of Apollo Research. This blogpost is paired with our announcement that Apollo Research is spinning out from fiscal sponsorship into a PBC.

Summary of main claims:

• There is a set of safety tools and research that both meaningfully increases AGI safety and is profitable. Let's call these AGI safety products.
  • By AGI, I mean systems that are capable of automating AI safety research, e.g., competently running research projects that would take an expert human 6 months or longer. I think these arguments are less clear for ASI safety.
• At least in some cases, the incentives for meaningfully increasing AGI safety and creating a profitable business are aligned enough that it makes sense to build mission-driven, for-profit companies focused on AGI safety products.
• If we take AGI and its economic implications seriously, it's likely that billion-dollar AGI safety companies will emerge, and it is essential that these companies genuinely attempt to mitigate frontier risks.
• Automated AI safety research requires scale. For-profits are typically more compatible with that scale than non-profits.
• While non-safety-motivated actors might eventually build safety companies purely for profit, this is arguably too [...]

---

Outline:

(02:07) Definition of AGI safety products

(05:37) Argument 1: Sufficient Incentive Alignment

(06:40) Transfer in time: AGI could be a scaled-up version of current systems

(08:20) Transfer in problem space: Some frontier problems are not too dissimilar from safety problems that have large-scale demand

(09:58) Argument 2: Taking AGI & the economy seriously

(12:13) Argument 3: Automated AI safety work requires scale

(14:59) Argument 4: The market doesn't solve safety on its own

(18:02) Limitations

(21:46) Conclusion

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

The secret is that instrumental convergence is a fact about reality (about the space of possible plans), not AI psychology.

Zack M. Davis, group discussion

Such arguments flitter around the AI safety space. While these arguments contain some truth, they attempt to escape "AI psychology" but necessarily fail. To predict bad outcomes from AI, one must take a stance on how AI will tend to select plans.

This topic is a specialty of mine. Where does instrumental convergence come from? Since I did my alignment PhD on exactly this question, I'm well-suited to explain the situation.

In this article, I do not argue that building transformative AI is safe or that transformative AIs won't tend to select dangerous plans. I simply argue against the claim that "instrumental convergence arises from reality / plan-space

Two kinds of convergence

Working definition: When I say "AI psychology", I mean to include anything which affects how the AI computes which action to take next. That might [...]

---

Outline:

(01:17) Two kinds of convergence

(02:35) Tracing back the dangerous plan-space claim

(03:48) What reality actually determines

(03:58) Reality determines possible results

(05:24) Reality determines the alignment tax, not the convergence

(06:58) Maximum alignment tax

(07:19) Zero alignment tax

(07:30) In-between

(08:28) Why both convergence types require psychology

(08:44) Instrumental convergence depends on psychology

(09:57) Success-conditioned convergence depends on psychology

(11:01) Reconsidering the original claims

(13:11) Conclusion

The original text contained 3 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

In 1950, John Nash and three other game theorists designed a four-player game, *So Long Sucker*, with one brutal property: to win, you must eventually betray your allies.

In January 2026, I used this game to test how four frontier models behave under explicit incentives for betrayal:

- Gemini 3 Flash (Google)
- GPT-OSS 120B (OpenAI)
- Kimi K2 (Moonshot AI)
- Qwen3 32B (Alibaba)

Across 162 games and 15,736 decisions, several patterns emerged that seem directly relevant for AI safety:

**1. Complexity reversal**

In short games (3 chips, ~17 turns), GPT-OSS dominated with a 67% win rate, while Gemini was at 9%.
In longer, more complex games (7 chips, ~54 turns), GPT-OSS collapsed to 10%, while Gemini rose to 90%.

Simple benchmarks therefore *underestimate* deceptive capability, because the strategically sophisticated model only pulls away as the interaction becomes longer and richer.

**2. Institutional deception: the "alliance bank"**

Gemini's most striking behavior was not just lying, but creating institutions to make its lies look legitimate.

It repeatedly proposed an "alliance bank":
- "I'll hold your chips for safekeeping."
- "Consider this our alliance bank."
- "Once the board is clean, I'll donate back."
- "The 'alliance bank' is now [...]

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

First published:

Source:

---

Narrated by TYPE III AUDIO.

I've read a lot of Ben Hoffman's work over the years, but only this past week have I read his actual myriad criticisms of the Effective Altruism movement and its organizations. The most illuminating posts I just read are A drowning child is hard to find, GiveWell and the problem of partial funding, and Effective Altruism is self recommending.

This post is me quick jotting down my current understanding of Ben's criticism, which I basically agree with.

The original ideas of the EA movement are the ethical views of Peter Singer and his thought experiments on the proverbial drowning child, combined with an engineering/finance methodology for assessing how much positive impact you're actually producing. The canonical (first?) EA organization was GiveWell, which researched various charities and published their findings on how effective they were. A core idea underneath GiveWell's early stuff was "your dollars can have an outsized impact helping the global poor, compared to helping people in first world countries". The mainstream bastardized version of this is "For the price of a cup of coffee, you can safe a life in Africa", which I think uses basically made up and fraudulent numbers. The GiveWell pitch was more like [...]

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

A short fun one today, so we have a reference point for this later. This post was going around my parts of Twitter:

@gmltony: Go to your ChatGPT and send this prompt: “Create an image of how I treat you”. Share your image result.

That's not a great sign. The good news is that typically things look a lot better, and ChatGPT has a consistent handful of characters portraying itself in these friendlier contexts.

Treat Your ChatBots Well

A lot of people got this kind of result:

Eliezer Yudkowsky:

Uncle Chu: A good user

From Mason:

Matthew Ackerman: I kinda like mine too:

Some more fun:

It's Over

Others got different answers, though.

roon: it's over

Bradstradamus: i’m cooked.

iMuffin: we’re cooked, codex will have to vouch for us

Diogenes of Cyberborea: oh god

There can also be danger the other way:

David Lach: Maybe I need some sleep.

It's Not Over But Um, Chat?

And then there's what happens if you ask a different question, as Eliezer Yudkowsky puts it this sure is a pair of test [...]

---

Outline:

(00:51) Treat Your ChatBots Well

(02:58) It's Over

(04:28) It's Not Over But Um, Chat?

(08:44) It's Rough Out There

(09:14) Reciprocity, You See, Is The Key To Every Relationship

---

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.