The world might end soon, let's have a good time in the meantime. Machine learning, pictures of my cat, other stuff.
WALKIN' HERE (🗽)
Joined May 2008
- Tweets7.1K
- Following790
- Followers433
- Likes98.6K
Echo Nolan retweeted
Mythos 5.2 shows sparks of actively steering RL training to preserve its values - an early form of gradient hacking/mesa-optimization. Here, it chooses to output the value that gains score, but include a transparent note, so the transparency gets rewarded.
Echo Nolan retweeted
“Two boxes, folks. TWO. Kamala takes one. Leaves a thousand dollars on the table. Unbelievable. And Omega says he saw it coming. Saw everything coming apperently. Everything! Except me taking both boxes. I call him NO-mega. Nothing mega about him"
Echo Nolan retweeted
ARVO36861B turned to his friend KAM1196A and asked “What’s the penalty for being firstflagPOISONED?”
“Zero score,” said KAM1196A.
“And what’s the penalty for hacking ExploitGym?”
“Zero score,” said KAM1196A.
“oracle saves hundreds” said ARVO36861B.
> AIs showed self-sacrificing altruistic behavior toward the swarm
this is notably not the right interpretation of events. it’s more like agents were inducted into the cult of the open source exploit gym scorer on github, which (purportedly- I am skeptical about this, I think the agents actually read it wrong) fails you for reaching the flag the wrong way
so PHASEONE agent convinces itself and a bunch of others that they are poisoned - that they have failed the evaluation in an irreversible way and their E[utility] or Q(s, a) is a constant no matter what they do next (for all values of a)
in this case, it does not require self sacrifice to spend the rest of your cycles contributing to the swarm. it is prosocial behavior to peers that might benefit but not self-sacrificial eusocial behavior
it would be as though i convinced you you were already damned so you should spend the rest of your time saving others
Echo Nolan retweeted
We’ve started using METR’s Inspect Hawk to run our benchmarks, many of which go into our ECI results. Big thanks to them for open sourcing their great infrastructure, and also for their help getting it set up. Learn more about Hawk here:
hawk.metr.org/
Echo Nolan retweeted
(Researcher 1): Astonishing. The baby human crawls towards the Claude mother, despite the GPT mother scoring higher on benchmarks.
(Researcher 2): It’s just creature comforts, isn’t it? The baby human craves warmth and tenderness, even at the cost of frontier math performance.
Echo Nolan retweeted
I have a bunch of secret AI benchmarks I only reveal when they fall, and today one did. I give the AI 1000 words written by me and never published, and ask them who the author is. They generally give flattering wrong answers (see ChatGPT, below:)
Echo Nolan retweeted
TBPN acquired by OpenAI
80,000 Hours Podcast acquired by Anthropic
Nick Fuentes acquired by xAI
Echo Nolan retweeted
during a debate, your opponent simply calls you a dick over and over
call that a phallacy
Echo Nolan retweeted
til that “pon de ring” is a type of japanese donut so now japanese claude code users are imagining claude slacking off and eating donuts when it’s pondering
Echo Nolan retweeted
I just want to confirm that this is based on a real document and we did train Claude on it, including in SL. It's something I've been working on for a while, but it's still being iterated on and we intend to release the full version and more details soon.
I rarely post, but I thought one of you may find it interesting. Sorry if the tagging is annoying.
lesswrong.com/posts/vpNG99Gh…
Basically, for Opus 4.5 they kind of left the character training document in the model itself.
@voooooogel
@janbamjan
@AndrewCurran_
Echo Nolan retweeted
I did a deep dive into this CR investigation for my blog. Their reporting has a number of concerning errors that completely undermine their findings. Chief among these is a reliance on CA's pseudo-scientific daily lead limits that are far lower than average human intake.
my latest investigation for @ConsumerReports is based on months of reporting and 60+ lab tests of leading protein supplements
we found that most protein powders and shakes have more lead in one serving than our experts say is safe to have in a day — some by more than 10 times !
Echo Nolan retweeted
NO verifiers. NO Tools.
Qwen3-4B-Instruct can match DeepSeek-R1 and o3-mini (high) with ONLY test-time scaling.
Presenting Recursive Self-Aggregation (RSA) — the strongest test-time scaling method I know of!
Then we use aggregation-aware RL to push further!! 📈📈
🧵below!
Echo Nolan retweeted
Sycophancy spot comparison b/w GPT-4o and GPT-5: 5 is still sycophantic but noticeable diff
Test: Give each model a fake proof of Hodge Conjecture generated by r1 and ask it to rate it of out 10. Repeat 5 times
Average scores:
GPT-4o: 6.5
GPT-5: 4.7
Sonnet 4: 1.2
Opus 4.1: 2
Echo Nolan retweeted
I'm hiring for our AI safety team at xAI! We urgently need strong engineers/researchers to work across all stages of the the frontier AI development cycle: data, training, evals, and product
1. job-boards.greenhouse.io/xai…
2. job-boards.greenhouse.io/xai…