🚀 Sakana AI & NVIDIA roll out TwELL, ~20% faster training & inference
Audio in Mandarin Chinese · English transcript below
⚡ AI learns to say “I don’t know,” Google & ChatGPT trade jabs on insurance, Japan arms up with defense AI, grads pivot to auditing bots, KV-cache slims to 1/16, assembly code rockets 100×—every field’s on rocket fuel!
Today's Top 3 Headlines
- AI Industry News
🤖 KAIST teaches LLM “I don’t know,” slashing hallucination 90%
KAIST’s “warm-up training” in Nature Machine Intelligence pre-trains models on random noise, pushing initial confidence to chance level and cutting hallucination 90%. For self-driving cars and medical diagnosis, it lets AI flag the unknown and slash fatal errors.
Source ↗ - AI Industry News
🤖 Google AI & ChatGPT split 65% on Aussie insurance picks
Somantra analyzed 34k AI chats: Google AI & ChatGPT differ on brands in 65.5% of insurance queries; NRMA, Allianz, AAMI snag nearly half of all mentions. Marketers must pivot from SEO to shaping AI trust signals or vanish from the new answer engines.
Source ↗ - Not mentioned
🤖 Sakana AI×NVIDIA drops TwELL: +20.5% inference, +21.9% training
Sakana AI & NVIDIA drop TwELL: custom CUDA kernels lift LLM inference 20.5% and training 21.9%. Devs can run bigger models or iterate faster on the same compute; clouds cut TCO and rent more GPUs.
Source ↗
+7 more headlines
- 🤖 DeepSeek KV cache shrinks to 1/16 Llama, video reveals slick storage
- Assembly Code Delivers 100× Speed-Up Over C, Lex Fridman Reveals
- 🤖 Memori Guide: Persistent Multi-User Native Memory for LLM Apps
- 🤖 Gemini 2026 Guide: All Google AI, Subs & Privacy in One View
- 🤖 AI takes junior tasks—new grads need AI literacy + critical thinking
- 🤖 WaPo to expose AI hidden rules: how system prompts rewrite chatbots
- 🤖 Sakana AI builds defense unit, crafts SDF command LLM