Takeru Miyato

@takeru_miyato

MTS @ElorianAI

takerum.github.io

Joined June 2017

Replies Reposts Media

Pinned
Takeru Miyato
@takeru_miyato
Feb 23, 2025
Synchronization is ubiquitous in nature and is a key mechanism for information processing in the brain. We introduce AKOrN as a dynamical alternative to threshold units, which can be combined with MLPs, CNNs or Transformers. ICLR'25 Oral. To feel the vibe: takerum.github.io/akorn_project_…
GIF
GIF
6
Takeru Miyato
@takeru_miyato
Sep 12
when you gaze long into the abyss, the abyss gazes also into you
Anthropic
@AnthropicAI
Sep 10
We're publishing our most detailed threat intelligence report to date. It covers how people tried to misuse Claude—for cyberattacks, influence operations, surveillance, biology, and building weapons—and how we found and stopped them. We disrupted every operation in the report,
Takeru Miyato
@takeru_miyato
Sep 2
what's this ceiling? is it just model fail, or did they find alternative plausible answers that humans wouldn't come up with?
ARC Prize
@arcprize
Sep 1
Fable 5.1 from @AnthropicAI on ARC-AGI (Verified): - ARC-AGI-2: 90.0%, $3.12/task - ARC-AGI-1: 97.5%, $1.40/task Across the two benchmarks, Fable 5.1's average cost per task is about 32% lower than Fable 5's, driven by better token efficiency.
Takeru Miyato
@takeru_miyato
May 8
In 2023 we already experimented with DeepSeek-v4-style output rotation but in vision tasks (diffusion and novel view synthesis) and it works better than QK-only rope. arxiv.org/abs/2310.10375. Also it's shown that QKVO-rope gives better scaling in arxiv.org/abs/2602.21341
You Jiacheng
@YouJiacheng
May 7
> Directly applying RoPE rotation to KV will leak positional information into value matrix V 科学空间亦有记载 kexue.fm/archives/10862
Takeru Miyato
@takeru_miyato
May 7
think deep but in time direction