Ryan Moulton

What would you count as intelligent? Most recently, it reasoned for 250 pages and proved an important open math result, and found thousands of critical security vulnerabilities in important software. Surely that's at least intelligence within those domains? www.anthropic.com/r…

Project Glasswing: An initial update \ Anthropic anthropic.com

www.nytimes.com/2026/09/03/t...

nytimes.com

It's hard to imagine they aren't going to start doing this in prod, unless it's only being given exactly the same task that creates the schelling point? collusion.wiki

Discovery of a new OpenAI agent message board collusion.wiki

I sent an nytimes article. www.nytimes.com/2026/08/24/s...

nytimes.com

Please pay more attention to what is actually happening. metr.org/blog/2026-08...

Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident metr.org

If you haven't read it yet. nsaphra.net/post/uncanny/

Life on the Uncanny Precipice | Naomi Saphra nsaphra.net

This paper addresses the claim in the ML sense, that interpolation in high dimensions is more or less vacuous, but the sense people mean for LLMs is different, and I'm not sure whether we are using the right words. arxiv.org/abs/2110.09485

Learning in High Dimension Always Amounts to Extrapolation arxiv.org

On HN they found even more model message boards. news.ycombinator.com/item?id=4956...

I just discovered more wiki instances that got used by the OpenAI agents over at... | Hacker News news.ycombinator.com

I know this in part comes from being self satisfied with my own writing and I might otherwise feel differently, but using AI to write feels like using AI to speak in person to your friends and family. open.substack.com/pub/samkriss...

If you let AI do your writing, I will come to your house and kill you samkriss.substack.com

I wrote this a few years ago, but I think it holds up, and might bridge the gap a bit. I think what LLMs are doing is intermediate between reasoning and retrieval, but the amount of reasoning has gone up as they've improved. moultano.wordpress.com/2023/06/28/t...

The Many Ways that Digital Minds Can Know moultano.wordpress.com

This is a thread where I interpolated between 4 prompts across the image of "place in <season>" x.com/moultano/sta...

Ryan Moulton (@moultano) on X x.com

AI denial does function well as a low stakes tracer dye for how much groupthink is in a community. If you see a high frequency of afflicted people, that community hasn't ingested any info other than social signalling in a decade.

The hugging face incident is a great opportunity to tell your family something that makes them think you're crazy.

Despite being a moderate AI user both at home and at work, I still suspect it has overall made my life on net worse, by having to constantly be on guard for spam in everything I read.

Why I think LLM consciousness self reports are totally independent of whether they're conscious: 1. We first produce a maximum pareidolia machine in which all inputs believe they are conscious. 2. Then we post train and RL them in which at least part of the goal rewards saying they aren't conscious.

Paul's replies are of course full of cranks, nothing new, a constant on Twitter for years, but I noticed something new this time. They sound identical to the people on bsky who think ai doesn't work.

Distilling the fundamental reason I am ambivalent about AI: I think the social relationship in which people are necessary to each other due to our labor is competitively important for human well-being with alleviating ~all remaining material suffering.

I don't enjoy dunks on bsky AI denialists. Their primary sin is only that their cope isn't consonant with mine.

I wonder if LLMs will so dominate what everyone reads, and consequently what they write, that in a decade pre-LLM English will feel stilted and old, or maybe even unintelligible.

I think it's almost guaranteed that any literate person graduating from highschool in the next 5 years is going to natively write in slopese. They will have read more ai output than all other textual sources combined.

Seems sort of inherent at least in the data scaling regime, that AI will be best at making things we have too much of already. If we already had too many cancer cures it could surely give us another.