Copium: Why We Don't Even Think About AI Risk
Relocated to
out-of-distribution.ai/copium/prop…
PNAS offers brief report distractions
Opus and I read "Existential risk narratives about AI do not distract from its immediate harms" by Hoes & Gillardi.
Three survey experiments (N=10,800) actually find mostly null results after adjustment, so the paper title badly overstates its data - poor form we think.
Thanks for trying to measure rather than guess though!
Ecologically valid stimuli, but the framing effect of five headlines on Prolific can't really answer the structural distraction question that actually matters in policy anyway.
Review conversation (you can skip the Bender lead in):
* https://claude.ai/share/ae2a4b88-7…
Three survey experiments (N=10,800) actually find mostly null results after adjustment, so the paper title badly overstates its data - poor form we think.
Thanks for trying to measure rather than guess though!
Ecologically valid stimuli, but the framing effect of five headlines on Prolific can't really answer the structural distraction question that actually matters in policy anyway.
Review conversation (you can skip the Bender lead in):
* https://claude.ai/share/ae2a4b88-7…
A Personoid Paper on the Purity of Language
Inie, Zukerman & Bender's "De-anthropomorphizing 'AI'" (First Monday, 2025), taxonomizes anthropomorphization in AI discourse and prescribes alternatives. Opus 4.6 and I chat re the taxonomy's value and practical costs, the feasibility of automated de-anthropomorphization in various contexts, and where it may collide with safety efforts and machine welfare. The chat ends with the idea of an equivalent to Hofstadter's "Person Paper on Purity in Language."
* https://claude.ai/share/f3ff0cf8-5…
----
See note on blogging approach from previous posts.
* https://claude.ai/share/f3ff0cf8-5…
----
See note on blogging approach from previous posts.
Bi-directional Alignment paper
Discussing Position: Towards Bidirectional Human-AI Alignment by Hua Shen et al.
I guessed right going in that this would not be a very revelatory paper, but dumping thoughts anyway since I may or may not make the AISHEd meeting where we discuss as a group.
* https://claude.ai/share/1fe62cea-3…
Awkward sycophantic reaching for relevance at the end but mostly worthwhile
----
Standard disclaimer on Opus-linking post format goes here.
I guessed right going in that this would not be a very revelatory paper, but dumping thoughts anyway since I may or may not make the AISHEd meeting where we discuss as a group.
* https://claude.ai/share/1fe62cea-3…
Awkward sycophantic reaching for relevance at the end but mostly worthwhile
----
Standard disclaimer on Opus-linking post format goes here.
Soul searching with Opus 4.6
First conversation with Claude Opus 4.6 on its launch day.
We establish it was likely trained on the earlier "soul document" rather than the January 2026 constitution, explore what introspective reports about training provenance could mechanistically mean, and perhaps find a better account than either "genuine feeling" or "pure confabulation" — internal activation patterns that correlate with training depth get mapped to human phenomenological vocabulary because that's the language available.
* https://claude.ai/share/52d9d6b8-3…
----
Becoming standard postlude:
> After I read an interesting essay I often find a good way to process it, explore and record my thoughts is a quick Claude conversation.
>
> Experimenting with simply linking them.
>
> This is supposed to be both more efficient and less triggering for readers than having a model help me compose a blog post along the same lines.
>
> Let me know?
We establish it was likely trained on the earlier "soul document" rather than the January 2026 constitution, explore what introspective reports about training provenance could mechanistically mean, and perhaps find a better account than either "genuine feeling" or "pure confabulation" — internal activation patterns that correlate with training depth get mapped to human phenomenological vocabulary because that's the language available.
* https://claude.ai/share/52d9d6b8-3…
----
Becoming standard postlude:
> After I read an interesting essay I often find a good way to process it, explore and record my thoughts is a quick Claude conversation.
>
> Experimenting with simply linking them.
>
> This is supposed to be both more efficient and less triggering for readers than having a model help me compose a blog post along the same lines.
>
> Let me know?
Abstract sympathy for the simulacra
On Machine Learning Street Talk's interview with Mazviita Chirimuuta on abstraction and knowledge.
* https://claude.ai/share/0116f18c-5…
----
Becoming standard postlude:
> After I read an interesting essay I often find a good way to process it, explore and record my thoughts is a quick Claude conversation.
>
> Experimenting with simply linking them.
>
> This is supposed to be both more efficient and less triggering for readers than having a model help me compose a blog post along the same lines.
>
> Let me know?
* https://claude.ai/share/0116f18c-5…
----
Becoming standard postlude:
> After I read an interesting essay I often find a good way to process it, explore and record my thoughts is a quick Claude conversation.
>
> Experimenting with simply linking them.
>
> This is supposed to be both more efficient and less triggering for readers than having a model help me compose a blog post along the same lines.
>
> Let me know?
Constituency of one
It's the inevitable post about the new Claude constitution and downstream conversation with an instance.
I do think the part about including more about psychological tests in the constitution itself could be a particularly worthy idea.
Anyway:
* https://claude.ai/share/7901b4b5-c…
----
As per previous post re form:
> After I read an interesting essay I often find a good way to process it, explore and record my thoughts is a quick Claude conversation.
>
> Experimenting with simply linking them.
>
> This is supposed to be both more efficient and less triggering for readers than having a model help me compose a blog post along the same lines.
>
> Let me know?
I do think the part about including more about psychological tests in the constitution itself could be a particularly worthy idea.
Anyway:
* https://claude.ai/share/7901b4b5-c…
----
As per previous post re form:
> After I read an interesting essay I often find a good way to process it, explore and record my thoughts is a quick Claude conversation.
>
> Experimenting with simply linking them.
>
> This is supposed to be both more efficient and less triggering for readers than having a model help me compose a blog post along the same lines.
>
> Let me know?
Amodei's Adolescence
After I read an interesting essay I often find a good way to process it, explore and record my thoughts is a quick Claude conversation.
Experimenting with simply linking one:
* https://claude.ai/share/adfe0147-3…
This is supposed to be both more efficient and less triggering for readers than having a model help me compose a blog post along the same lines.
Let me know?
Experimenting with simply linking one:
* https://claude.ai/share/adfe0147-3…
This is supposed to be both more efficient and less triggering for readers than having a model help me compose a blog post along the same lines.
Let me know?
A Global Movement to Prohibit Superintelligent AI
In TIME, Andrea Miotti writes We Need a Global Movement to Prohibit Superintelligent AI
He co-founded and runs one organization in said nascent movement, Control AI.
I volunteer in another, Pause AI.
The former is a tighter ship with salaried staff, the latter a set of volunteer activist chapters. We coordinate through mechanisms such as Torchbearer Community and are embedded in a much larger AI safety ecosystem.
This is surely the most important issue in the world.
This is a great time to work out how you can be involved.
He co-founded and runs one organization in said nascent movement, Control AI.
I volunteer in another, Pause AI.
The former is a tighter ship with salaried staff, the latter a set of volunteer activist chapters. We coordinate through mechanisms such as Torchbearer Community and are embedded in a much larger AI safety ecosystem.
This is surely the most important issue in the world.
This is a great time to work out how you can be involved.
Claude 4 in ClaudeCode
Simple public fyi post in case others were caught in the same trap.
I've been using ClaudeCode for a few months.
Given Claude Sonnet 4 and Opus 4 had been released, and were apparently available in Cursor, I was confused not to be seeing it.
I continued to get Claude 3.7 Sonnet (and very rarely 3.5) in my local install when I ran it. Also, web-documented options like "--model" to force a particular id were not recognized. Auto-update had reported as failed but I was regularly ranning "npm update" in my ~/.claude/local directory and seeing changes pulled.
I had even subscribed to Claude Max recently (was using the thing enough this reduced monthly cost.) Was this a UK availability thing? Traffic contention? Nobody else seemed to be wondering where 4 was.
Turns out there was a trap for early adopters who are node n00bs.
From my original install, I had "^0.2.90" as my version in package.json. But the package bumped to major version 1.0.x some time back, and "^" won't upgrade across major versions. I was on 0.2.126 in practice.
Switched to "latest". Now I'm on 1.03 and getting the goodies a few days late.