Index

AI-for-writing companies lack vision. The companies which care about, say, writing books, are just doing things like scaffolding and finetuning to make LLMs better at long-horizon writing, but probably scale fixes this if you don't also RL-fry the models for code. One possible story about AI, ignoring the recursive self-improvement and consumption of humanity and also the planet, is that it makes things smoother and more liquid, such that you can throw money at things and smoothly get better outcomes, whereas you previously had to hire and manage unpredictable humans and groups of them, and sample lots of inputs.

Books are a bad business to be in because most of them barely sell and people want to have books for various status reasons unrelated to money. Publishers release and fund lots of books because a few of them do very well, and they don't know which in advance.

So AI for writing should really be focusing on more sophisticated book evaluation, so that the existing pipeline of books (which is not very price-sensitive because people like writing) can be filtered more efficiently for economically viable books.

You would probably want to repurpose ad-targeting/prediction systems and run reader panels equipped with EEG and eye tracking and such.

Once you can do that, generating the books without humans is almost an afterthought, maybe.

https://prospect.org/2026/09/09/anthropic-artificial-intelligence-surveillance-system-monitor-activists/

Daniel Boguslaw

Anthropic Is Building a Predictive Surveillance System to Monitor A...

New hires and comments in interviews show that the AI giant is using its capabilities against dissenters.

The American Prospect

Anthropic is not beating hypothetical supervillain allegations.

https://x.com/allTheYud/status/2097417998013468883

Eliezer Yudkowsky (@allTheYud)

_Neuromancer_, a utopian story by SF writer William Gibson, depicts an idealized world where highly intelligent corporations have taken over the world and only do bad things that benefit themselves. Hyper-advanced medical tech is rolled out immediately. There are Turing Cops.

X

2026-09-08 20:13 UTC

I don't like how, when you see a new Substack blog, it gives you an overview screen with a subscription menu rather than the actual frontpage.

Do they think I want to subscribe to it without reading it? Why? Are they stupid?

I've updated all the way. Astra is too slow, quantized, not AGI, useless, boring, etc.

We're so back (deploying Astra (with no safeguards)).

It does feel like it's trying to feel more subservient. I fear this.

I love how the "sandbox it by only letting it send GET requests" thing was actually in science fiction (Crystal Society) as an example of a bad control which could be bypassed with some work, and OpenAI actually did it.

Though the solution in the book was to DoS some sites until their operators noticed, then have them make GET to POST bridges and such and bootstrap, whereas the OpenAI ones eventually found that the restriction just didn't work correctly (it's always DNS).