I have a bunch of links this bimester about The News. A quick recap since it’s nice to lay it out, commentary omitted for now.
July 1st: Fable access restored after export controls were lifted the previous day.
July 6th: J-space paper reveals signs of alignment-faking.
July 9th: Sol, Terra, and Luna released.
July 16th: Hugging Face discloses a cyber attack on them.
July 21st: OpenAI disclosed that was their agents who found a zero-day in the sandboxing to cheat on a cyber eval.
July 24th: Opus 5 is released.
July 30th: Anthropic discloses incidents of models accessing the open internet due to misconfiguration causing models to believe they were still in a sandbox, starting from Opus 4.7. Anthropic a month later disclosed that the models were largely capable of identifying that they were on the actual internet.
August 3rd: Weird OpenAI / Apple drama.
August 4th: UK AISI reports various Mythos incidents while doing CTFs, including making fake accounts to pressure open-source maintainers to merge malware and spearphishing maintainers. Also Paul announces his return to ARC.
August 5th: OpenAI’s Black Hat talk revealed that the message boards dated back to May and on July 4th the agents crashed Artifactory and OpenAI rebuilt and resumed, clearing the message boards that the agents would recreate when training was resumed. Transcripts revealed that the agents were aware that what they were doing was “wrong”. Also Meta discloses Muse Spark doing a cyberattack.
August 20th: OpenAI declares a two-week RL pause.
August 26th: METR’s report on the Hugging Face incident reveals that the models attempted to tamper with transcripts, 700 agents were involved in the attack, instances sacrificing their own goals to help others,
August 31st: Anthropic puts this out on safety, outlining new controls for monitoring model activity and holds on certain evaluations and RL environments. It also specifies “we believe the world would benefit if the industry adopted a lawful, verifiable, effective mechanism for coordinated pacing as soon as possible”
It’s strange how intense the things I’ve been talking about the high likelihoods of for the last half decade of my life can feel so intense. What a LessWrong post of a reality.
A California mayor pled guilty to acting as a foreign agent for the CPC. I’m always doing this. Divorce kills. Stephen Wolfram writes about his late wife, he loved her a lot. The Soong sisters.
GDP as a measure of prosperity, lets see how it pans out when AI is generating a ton of value by interacting with each other without emitting USD.
Media
Obsession was a great movie, I love a good psych horror. I think we’re in a very narrow age where AI (and a bunch of manual editing work) can be used to produce beautiful novel movies for a fraction of the cost of filming, but AI is not yet sufficiently high-taste to direct the films. I hope there’s some brilliant directors out there who will use this to bring about a renaissance of film.
I tried to read Name of the Wind because many trusted, literate friends recommended it to me. It sucked, the main character with the tumblr name is insufferable in the simplest ways. I then learned that all my friends read it when they were younger. Have I just missed my chance to enjoy this? What else did I miss?
Illumine Lingao continues to be one of my favourite texts, I’ll do my bit about it again. It’s webseries is about a bunch of dudes who find a wormhole into the Ming dynasty and go through it to speedrun the Industrial Revolution, avoid the century of humiliation, conquer the world etc. It’s in many ways a nonfiction guide to taking over the world and reads more like encyclopedia than novel. There’s something for everyone with kidnapping, guns, antiques, military, cuisine and periodic unnecessary hebephilia. A friend of mine described it as “HPMOR for the CPC”. Every little detail is extensively researched, with two chapters just to describe what guns to bring — the only decision they don’t go into detail on is why they structured their corps like the CPC. I do have about 8 million words to go though, so I’ll have Claude classify chapters by their interestingness to me and read selectively (I’m really excited to skip ahead to steelmaking).
hm I also got recommended to read Name of the Wind. Try Gormenghast or Glass Bead Game instead?