Lesswrongers go full panpsychism: https://www.lesswrong.com/posts/wCY4oNHpH2HGYGSoP/the-user-in-the-llm-might-be-a-moral-patient-too-a-strange

To try to TLDR this multilayered bit of crazy... lesswrongers already theorize LLMs might be sapient and/or have morally meaningful internal experiences. Except... LLMs obviously "behave" wildly differently when prompted differently and especially when put into different "agent harnesses"... so the big brain follow-up thought is what if the different personas an LLM can exhibit are all different "people" (for the purposes of utilitarian accounting and/or what their internal experiences are like... the exact "theory" can vary by lesswronger). The even more insane full galaxy brain follow-up thought: what if the user's input to the LLM is forming yet another "person", like a copy/reflection/simulacrum of the user!

TLDR for the TLDR: What if not only the LLM was sentient/sapient, but also held a sentient/sapient simulacrum of the user inside!

Jonny took a look at Meta's "Wonder" model they just released. Here's how the thread starts:

The hourly history backup just failed in a way you should hear about plainly: the backup worker accidentally deleted the backup repo itself mid-run. The repo's entire git history and all the DB sidecar files up to watermark 330145 are gone — that deletion was the worker's own mistake, not a platform event.

Is there a history of how the cult of prediction markets spread to Silicon Valley? Its an old idea but since the 1990s is especially popular with Libertarians and hard neoliberals like Robin Hanson. When did it get on LessWrong? By 2013 or so, rationalists were telling strangers at social events that the only way to know if a healthcare system worked was to bet on it, and not just, you know, risk your life and health on it.

guys what if AI makes something really really good won't that make people not want to kill it????

https://www.lesswrong.com/posts/Ht4JZtxngKwuQ7cDC/tsvibt-s-shortform?commentId=5GqafvboceJ7phYLj

So I got bored and for some reason decided to look at the AI 2027 prediction grading doc again…and its Claude Mythos rank is just straight up wrong?

To quote the doc: “ The current best 80% time horizon is Mythos Preview at 186 minutes (version 1.1; there is no measurement for version 1.0). The methodology change from TH1.0 to TH1.1 drops the best model as of AI 2027 publication (Claude 3.7 Sonnet) by 2/3 from 15 to 10 minutes. Because we care more about the pace of progress rather than the absolute values, we look at when the AI 2027 trajectory reaches 186*(3/2)=279 minutes. In a central AI-2027-speed trajectory from our AI 2027 timelines model, 279 minutes is achieved in May 2026.”

Except Claude Sonnet wasn’t dropped to 10 minutes, it was dropped to 12. to play devil’s advocate maybe it was updated between then and now? but it doesnt seem like it.

Feel like this blog must have made the rounds but today is the first time I’m actually reading it and it’s sharp enough to cut with:

Every death on my corner is offset, with a third of a toddler left over, and I’d like that number in the conversation the next time somebody calls me a murderer out the window of a passing Kia.

.

Previously, on Awful, Gutsick Gibbon spent an hour ripping apart T. Beale's (VD's) anti-evolution book. She was still angry though, so she got her prior collaborator and four friends to join her for a complete dissection of VD's bullshit claims. Content warning for literally everything awful we discuss here, including sex crimes, neo-Nazis, Silicon Valley bullshit, and remarkably bad arithmetic. My choice sneer is her chosen video title, "Six Scientists Roast One Racist Loser (Heartwarming) (Four Hours) (Hopecore)". GG's requested that we not give VD any attention other than reminding him that he's such a pathetic loser that even white nationalists disavow him as too boring and insipid, so I'm not posting this to a top-level topic, but I'm getting a lot of household chores done while listening to it in the background and maybe you will too.

More news about the character dot ai bankruptcy: as it turns out, the true economics of the company are disastrous. At their peak, their revenue was around $1M a month while their AI compute expenses alone were $14M a month! Their total expenses in 2025 were around $18.5M a month before counting legal fees. All of this is for a pure text-based low power LLM (none of the "reasoning" business).

Before AI boosters crow about how much cheaper AI is currently, they tried every trick in the book to enshittify: open models, ads, usage limits, etc. This allowed them to cut down their losses to a mere $9M a month.

Surely this is all one big exception, everyone else in the AI bubble is doing just fine!

(h/t david gerard)

How much slop is in Linux now? You're no longer allowed to know, so its probably all slop

(why did I jump ship to Linux at this exact moment)

https://sfstandard.com/2026/10/13/openai-engineer-arrested-alleged-bugging-executive-conference-room/ h/t Naked Capitalism

A chatbot account tied to Alfozan’s spouse, which shared identifiers with his own account, submitted prompts between Sept. 24 and Sept. 30 about building, installing, obscuring, and remotely accessing a recording device, prosecutors said. Queries included how to “build a continuous audio recorder” and “encrypt the SD card,” prosecutors claimed. On Sept. 27 and 28, the account asked about the specific device that was later found, including its unique MAC address.

lol asking ChatGPT how to bug OpenAI from your wife's account.