If you want to look under the hood at the actual custom CUDA kernels and see exactly how we implemented the TwELL format for H100 GPUs, we’ve released the reference code. GitHub: https://
github.com/SakanaAI/spars
er-faster-llms
…
Blog: https://
pub.sakana.ai/sparser-faster
-llms/
…
MACHINE LEARNING
-
Custom CUDA Kernels for TwELL Format on H100 GPUs Released
By
–
-
Sakana AI Accelerates Sparse LLMs with NVIDIA
By
–
Sakana AIは、@NVIDIAとの共同研究で、スパースなTransformer言語モデルの推論・学習を高速化する新しいGPUカーネルとデータ形式を開発しました。
— Sakana AI (@SakanaAILabs) 9 mai 2026
ブログ：https://t.co/fMARMRFsJJ
LLMのコストの大部分を占めるフィードフォワード層では、実は各トークンに対して大半の活性がほぼゼロで無駄な計算に… https://t.co/nTMg0QgdSrSakana AI has developed new GPU kernels and data formats that accelerate inference and training of sparse Transformer language models through joint research with @NVIDIA
. Blog: https://
pub.sakana.ai/sparser-faster
-llms/
… In the feedforward layers, which account for the majority of LLM costs, most -
Claude Mythos Preview snapshot outperforms next best model by 2x
By
–
An early Claude Mythos Preview snapshot we provided METR has a time horizon of more than 2x the next best model on their 80% success rate benchmark
-
UFOs on HF: who will train first computer vision model?
By
–
The UFOs are on HF thanks to @MTSlive
! Who’s going to train the first computer vision model? https://
huggingface.co/MTSlive/datase
ts
… -
Sony and Bandai Namco launch generative AI pilot for game development
By
–
It was just a matter of time: Sony and Bandai Namco are launching a collaborative pilot around generative AI, positioning the tech as a way to speed up game development. Sony says AI is already helping with facial animation, QA, payments, visual fidelity, and future
-
Domino platform integrates AI lifecycle management from training to deployment
By
–
84% of developers use AI to code (Stack Overflow, 2025). Most stop at autocomplete. Inside Domino, your assistant learns the rest of the lifecycle. Training. Deployment. Governance. Monitoring. Same prompt. Whole platform. Blueprint: https://
hubs.ly/Q04fnbm-0 -
Architecting Infrastructure for Long-Horizon AI Agents
By
–
Join our live session and learn: How to think about the agent harness vs. the runtime layer underneath it What infrastructure long-horizon agents need for durable, stateful execution How to plan for memory, recovery, human intervention, observability, and scale Why
-
ByteDance AI Video Model Sparks China’s Next Big Moment
By
–
ByteDance’s new AI video model goes viral as China looks for second DeepSeek moment
#AI #AIio #AIInnovation #ML #DataScience #Futureofwork @timnitgebru @oriolvinyalsml @ceobillionaire @soumithchintala @waitin4agi_ @sallyeaves @bernardmarr -
DeepMind AI scores 48% on research-level math problems
By
–
DeepMind's AI co-mathematician scored 48% on FrontierMath Tier 4-research-level math problems that professional mathematicians need weeks to solve. The base model (Gemini 3.1 Pro) scores 19% alone. The entire jump comes from agentic scaffolding, parallel agents reviewing each
-
Hardening agentic stacks against reasoning and tool parsing drift
By
–
Most agentic stacks run into the same problems pretty quickly: reasoning and tool parsing drift across turns, KV cache reuse falls apart, or tools fire too late. We’ve been hardening Dynamo’s harness-facing path so @Claudeai Code, @OpenClaw
, and @openai Codex-style agent