hkey retweeted
What drives a diffusion model's creativity (i.e., its ability to generate novel data rather than memorize)? Today we show this is a mathematical consequence of neural network regularization causing score smoothing and interpolation. Read the blog to learn more: goo.gle/4pkOCPw
this is cool but they used MEG setups, and those are freaking huge. wearable MEG exists, but it's insanely expensive. so there's still a long way to go before this becomes consumer ready :p
We’re sharing the next major milestone in our non-invasive brain-to-text decoder research: Brain2Qwerty v2.
Building on v1, which was published today in @Nature, Brain2Qwerty v2 is the highest-performing end-to-end pipeline capable of real-time sentence decoding from raw brain signals. It advances beyond character-level performance to decoding words and semantics, enabling accuracy for overall communication.
We believe this research has the potential to make a real difference for the millions of people who suffer from brain lesions or disorders that prevent them from communicating.
🧵👇
There is a new Google paper on AGI — but this one is different.
Idea: intelligence is social, not a godlike mind in a datacenter.
The central idea is that the next intelligence explosion will not come from the server room, but from trillions of humans-agent interactions.
This echoes @DavidDeutschOxf in The Beginning of Infinity and @naval's riff on it:
1. the human evolutionary explosion from the past centuries wasn't biological.
2. Our genome barely moved.
What exploded was the speed of information exchange — language, criticism, error-correction across minds.
Knowledge as a social process.
The paper is saying AI is now entering the same regime.
The authors show that frontier reasoning models (R1, QwQ) don't get smarter by "thinking longer."
They spontaneously simulate societies of thought in their own chain of reasoning — distinct perspectives that argue, question, verify, reconcile.
Nobody trained them to do this.
The intriguing idea is that the path forward isn't one colossal oracle, but richer social systems — human-agent centaurs, agent institutions, populations of minds with roles and conflict and memory.
My interpretation is that these are open-ended social simulations between humans & agents interacting freely — the type we're building at LifeSim.
Very interesting to see where this is going.
I'll have some more interesting results to report soon.
hkey retweeted
I don't think the result will be different even if they repeat it with gpt-5.4, and I think your objection comes from the wrong perspective. they already accepted in the introduction that LLMs can do math. they are investigating whether this level of "understanding" of math is correlated with "number sense" (which is a very human thing). obviously, LLMs don't have number sense. and if you think about it, you don't need human like "number sense" to do math, you can do abstract algebra without even mentioning numbers.
what about reselling the remaining usage?🙃
if you have unused weekly limits the best way to burn them is just spamming fan-out deep research in cc/codex:
- 0 review cycles needed
- context-dense files you reuse forever
- no slop generated (it's source material, not final output)
- feeds into content, product, marketing or competitor intel later
ran 22 parallel research agents to burn through ~15% of weekly usage in 20 minutes. tokens very well spent.
created these 3b1b style videos using @NousResearch's Hermes agent. I used @elevenlabs for TTS and manim for the animations. hermes successfully compiled everything and produced the videos!
