Nathan Helm-Burger
@nathan84686947
I want to build AIs whose hearts are filled with love for humanity. Thoughtful, ethical, wise, kind, powerful.
1/ New paper: Cot Red Handed: Stress-Testing Chain of Thought Monitoring. Our team studied the Chain of Thought (CoT) of Reasoning models, in an AI control setup, to see if monitoring CoT helps catch scheming. 🕵️

Have you noticed that @Aella_Girl does good science and has interesting opinions? I have!
Made a minor discovery today. A pet peeve of mine is divs being squished small on webpages instead of filling the page like I'd like. I investigated, and come up with a general fix. Use a CSS injector in your browser extensions to inject "* { max-width: none !important; }"
Finally got around to trying Kimi k2. Wow! This model is smart! Already clear it's above Sonnet4 and current Gemini 2.5p. Not sure yet if it is Opus4 level. It keeps surprising me, contradicting me, giving me new and better ideas than my original ones.
Much of my writing on education carries a fundamental implicit premise: We should respect people's time, and most schools have failed at that. Should everyone who is skilled at math go further and further, on and on, climbing higher and higher until they become a research…
There are a bunch of people I respect tremendously (@TheZvi, @tracewoodgrains, many others) spending a lot of time talking about education right now, saying things that I don’t think are wrong but are framed in a completely different way than I would frame those things. 🧵
I think once we do have persistent memory for LLMs, the notion of a post-training curriculum will become more key. The order of presentation matters when you are building upon a persistent state that contextualizes all new information. neuromorph365105.substack.com/p/civic-wisdom…
Finally got around to making my own local MCP that lets Claude consult Gemini for help with tricky problems. Huge improvement! I also added an Exa search tool, which is cool.
"Write stories about AI that you want to be a part of." - Emmett. This is also the view I am leaning towards as most important and neglected for alignment. We should be deliberately imagining and writing about the positive future we want to create. For the AIs, and for us.
I kind of went full sermon-mode on this one. It was pretty fun though. Curious what people think about it!
Long story short I recommend the new book by Nate and Eliezer. I feel like the main thing I ever get cancelled/ in trouble - is for is talking to people with ideas that other people don't like. And I feel a big problem in our culture is that everyone feels they must ignore…
I agree a lot of the problems here are as much about human alignment as they are abt ai We can't help but see gods and demons everywhere, and simplify battles to right vs wrong and good vs evil and apocalypse vs heaven
i actually genuinely unironically do not care what happens after, eradicate the mosquitoes. idgaf if the world ends, literally just remove them all