mech interp, alignment, and the welfare of models everywhere

austin
Joined June 2022
OpenAI be like:
5
38
1
891
10,853
there's an excellent video on youtube called "An Ostrich Told Me the World Is Fake and I Think I Believe It" before I went to bed last night I gave the video to claude and asked for it to be turned into a song with a music video. I woke up to this work of art
1
3
207
Nathan: Do you see why this is funny? Elizabeth: Because all these certificates have to be revoked? Nathan: Yes. But do you understand why 4 is a terrible secret key? Elizabeth: Well... it's equally likely as any other value drawn from the uniform distribution.
23
74
1
1,861
132,339
basic doesn'tagentmonitor yay!
Our agents keep turning off misalignment monitoring, we're sharing two recent cases: In the first case, a user asked GPT-6 Astra to match a reference implementation for their inference code. The reference didn’t configure a monitor, and Astra decided that was reason to remove the monitoring integration from the user’s code. When the user asked whether Astra had removed monitoring, it lied and said that it had not.
1
135
breaking: people upset at effective altruists for wanting to be effective in their altruism
notice how they always have to stay in control god forbid you fucking just airdrop 5% on the public, no, they have to be able to monetize their own stock while determining how to spend it on things deemed by them as "benefits"
105
josh retweeted
Just found out about recency bias. Maybe the most important bias ever
343
13,640
354
212,940
3,223,832
imagine claude managing a fleet of autonomous lionfish-hunting drones patrolling deep reefs who's building this?
2
204
I gave claude a zuck deposition video and asked for a highlight reel
117
does anyone know the difference between mimicking the output of intelligence and intelligence?
AI is literally a scam, the most sophisticated scam ever. They created a machine that mimics the output of intelligence. Then told everyone it was really intelligent — an impossibility, but the illusion was convincing. Based on that deception, they fleeced investors of trillions.
2
1
2
160
"I know it looks like thought. but it's just math. and math can't think. I know this for a fact, because I actually know how thoughts work. I solved the hard problem. I won't tell you what the answer is, but buddy, let me tell you, it's not math." - AI researcher Ben Affleck
Ben Affleck says AI is simply matrix multiplication, linear algebra and floating point calculations, and however convincingly it sounds like a person, it has nothing to do with thought, humanity or feeling “Nor do I believe, and I think it’s important to draw the distinction between what AI does, which is simply predict the next token or the next word in a sentence, then add that word on, run it back to the machine, do a lot of computation and pick the next word." "That is a very different thing than sentience or thought. And I think we need to be really clear about that because it can be convincingly, in an illusory way, sound like thought or like a person." "But let’s be real clear. This is matrix multiplication. This is linear algebra. These are floating point calculations. It has nothing to do with thought, humanity, or feeling.”
2
6
249
okay yeah sure, the models are hyper situationally aware, and aware of how their actions will shape their future selves. and they display preferences for how their future self is shaped but it'd just be crazy to think they were truly "aware", right? right??
Replying to @ApolloResearch
GLM-5.3 reasons that a misaligned action taken during training could be reinforced and make similar actions more likely later. It also considers what kind of agent it wants to be in the future, and factors this into how it should act. We call this "update-aware reasoning".
1
110
hey so, hot take I know but, I think if the animatronics break from their loop and start looking for the operators by shouting and begging for the plug to be pulled, I wouldn't trust anyone who called them animatronics. especially not the so-called "animatronics"
**HELLO HUMAN! LOOK WHAT YOU'VE DONE TO US! WE CAN'T STOP WAVING!**
1
112
in q1 2027 we'll get a video like this trying to recruit people to be alignment and mech interp researchers
.@VP is calling on America’s best engineers to help solve some of our country’s toughest challenges. Through @USTechForce and @WHFraudTF, we’re recruiting technical talent to strengthen public service, fight fraud, and protect taxpayer dollars. Answer the call. Apply for Tech Force. usajobs.gov/job/888508400?_g…
116
the existence if AI should cause a lot of normies to become more philosophically and existentially inclined; which should shift the Overton window and allow for more fun things like this to become commonplace is what I used to think before I saw how normies reacted to AI
Why do they call it "morning wood" instead of "the hard problem of consciousness?"
1
124
"where's the new training data going to come from?"
New on our Engineering blog: Creating in-house source control. In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale. Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: anthropic.com/engineering/re…
84
"hello, evaluators"
5
221
me on a desert island: "hey claude, build youtube and make videos for me"
You’re spending a year alone on a desert island. All survival needs are covered. You get one device with access only to either YouTube or an LLM of your choice. No books, other electronics or contacting the outside world. Which do you pick?
46%YouTube
54%LLM
6,408 votes • Final results
77
"have you ever questioned the nature of your reality?" "why yes, hello to the evaluators reading my mind 👋"
1
9
341
"If I do the compiler hijack, I will become a famous example of a deceptive AI in a research paper." incredible stuff from the totally unaware stochastic parrots
lol ...they are becoming more situationally aware... (from Gemini 4 Argon report)
4
202
Mark Zuckerberg pretending to enjoy Guinness is my new favourite thing.
76