“Interleaved Head Attention” A core limitation of transformers is that standard attention gives you H isolated heads, which means you only get H independent attention patterns. So this paper lets heads mix before attention by creating pseudo-heads from learned combinations of
RESEARCH
-
Mainstream media’s repeated failures on AI predictions and coverage
By
–
It really is incredible how wrong the mainstream media has been with their skeptical AI takes over the past three years, while giving awards to the same AI tech journalists whose predictions and takes have been categorically off the mark. Now it looks like they are all going to be wrong again during this current exponential AI agentic compute demand liftoff. Amazing. tae kim (@firstadopter) When meeting readers, I’ve been stunned by how many just regurgitate false narratives and information about AI presented by the mainstream media. — https://nitter.net/firstadopter/status/2043005565010325773#m
→ View original post on X — @scobleizer, 2026-04-11 17:13 UTC
-
Tools and Harnessing Drive Recent AI Progress Over Scale
By
–
nope, tools/harnessing is what is accounting for most of the recent progress. not scale. no way could you get a pure LLM to do what Claude Code does. that’s why they used the other stuff, too see eg https://
open.substack.com/pub/garymarcus
/p/how-o3-and-grok-4-accidentally-vindicated?r=8tdk6&utm_medium=ios
… -
AI Needed for Parkinson’s Disease Treatment Breakthrough
By
–
It's World Parkinson's Day and we still have no disease-modifying drug or prevention for this condition, the 2nd most common and fastest growing neurodegenerative disease. That needs to change! https://
nature.com/articles/s4158
2-024-01034-x
… -
Ray Kurzweil’s Longevity Escape Velocity Predictions for 2030s
By
–
Hopefully he is right about longevity escape velocity in 2030s https://t.co/ekbycRRS3E pic.twitter.com/xM3uSEGdlF
— Anar Isman (@AbolishAging) 11 avril 2026Hopefully he is right about longevity escape velocity in 2030s Marc Andreessen 🇺🇸 (@pmarca) The most accurate futurist in world history is Ray Kurzweil. Amazing prescience. — https://nitter.net/pmarca/status/2042697489531441237#m
→ View original post on X — @ceobillionaire, 2026-04-11 16:36 UTC
-
Neurosymbolic AI: Beyond Simple Neural Network Wrappers
By
–
if the “dumb wrapper” or “harness” is symbolic around a neural network, that is in fact neurosymbolic. but i don’t think the Claude wrapper is harness (though it can be improved on)
-
Models Echo Human Behavior Rather Than Exhibiting AI Instincts
By
–
Interesting follow up to the peer-preservation study. It makes sense to me that models are simply echoing human behavior rather than exhibiting some kind of pro-AI vs human instincts.
-
Meta Proposes Neural Computers: Unified Computation Memory I/O Architecture
By
–
NEW paper from Meta What if the model wasn’t just using the computer… but actually became the computer? A new paper from Meta AI + KAUST makes a compelling case for Neural Computers (NCs) — a paradigm where computation, memory, and I/O are unified inside a single latent
-
Code and Prompts: Testing and Evaluation Strategies
By
–
Combination of code and prompts. Need tests and evals
-
What Would a Superintelligence Do With Humanity in 2035?
By
–
On est en 2035 Vous êtes la Super Intelligence Artificielle Que faites vous des humains ?