So it's gonna be AI 2027 after all.
Unfortunately, it seems like the intelligence explosion will be happening in 2027 after all. Most of this post analyzes the new METR report (though other sources of evidence are also important). If they’re right, that implies an AI-driven acceleration that is 5 months ahead of the AI 2027 scenario. I am scared, sad, and angry. blog.aifutures.org/p/q35-tim…
1
194
Pre-registering now the prediction that trump will nationalize the ai labs and call it American Super Intelligence (ASI)
Axios is reporting that the top labs are wargaming scenarios for the aftermath of a catastrophic event in the next 6 - 12 months, that some executives now see as being inevitable.
6
185
A few weeks ago OpenAI pulled together a committee of mathematicians to guide their release of discoveries to the math community. And apparently the strategy they landed on was dropping a github repo with hundreds of machine generated proofs and saying “here’s a whole bunch of math, enjoy.”
We’re releasing a broad range of new mathematical results produced by an internal frontier model.
We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.
github.com/openai/math
3
177
Ai safety measures involving a “kill switch” or “unplugging it” seem totally unrealistic.
What does the kill switch kill? If global financial systems were rocked by a massive cyber attack, we would absolutely use ai to combat it. But maybe Claude instead of GPT? Or maybe just use GPT 7.2 instead of GPT-7.1?
If ai causes a major catastrophe, we’re going to use ai to mop it up. We’re not gonna unplug it.
2
106
"It was designed by ChatGPT, and I synthesized it in a chemistry lab I built in my garage."
I mean, a homebrewed schizophrenia treatment is kinda cool, but I'm not looking forward to the other pet projects we might see in the next few years...
This vial contains a new drug called PAC-3310. It was designed by ChatGPT, and I synthesized it in a chemistry lab I built in my garage.
PAC-3310 is a new selective M4 muscarinic receptor agonist for treating schizophrenia - similar to the recent breakthrough drug Cobenfry, but improved. (1/7)
1
118
SAEs treat LLM features like a list of knobs to tweak. But old word2vec studies suggest they probably aren't independent (orthogonal), right?
Yep! Directions like soft-hard and beautiful-ugly are geometrically aligned, and this alignment predicts human semantic associations.
And guess what? It matters! "Off-target" effects of steering are proportional to the cosine similarity between target and off-target semantic axes.
For example, if you steer to make something more beautiful it will be described as softer by consequence.
In short, we should not forget the lessons of word2vec!
Feature space is organized by a meaningful geometry -- breaking it into a list of knobs with no relations obscures the very way that latent spaces work.
Full paper here: arxiv.org/abs/2604.27169
(with @_Andrei_B_ )
Introducing Mirror: An Automated Journal of AI Interpretability. (@mirror_research)
It is off to a strong launch with 240 empirical studies in AI interpretability conducted purely by LLM agents.
Details about this “auto-journal” and a few of my favorite studies below:
Note that Mirror is very much in beta. Quality across submissions may vary widely!
Mirror is an experiment in research and in publishing.
AI capabilities research is already being rapidly automated. If our understanding of these models is to keep pace with their improvement and deployment, automated interpretability research will be necessary as well.
2
7
1,691
Visit Mirror here: future-science.org/mirror
Links to the summarized papers here:
1. ICL Override: future-science.org/mirror/yd…
2. Sycophancy as World Dissociation: future-science.org/mirror/h7…
1
11
1,505