#LLM
Curated coverage on «LLM»: releases, research, benchmarks, open-source.
AIRI Exposes the Weakness of Token Cramming: 1568 Tokens in One Vector, but Comprehension Is Lost
FusionBrain AIRI researchers show at ICML 2026 that an LLM can restore 1568 tokens word for word from a single vector, yet completely loses the ability to reason…
China's AI Security Market: A Sovereign Stack, Party Standards, and the Same Threats as Everywhere Else
Raft dissected China's AI security market: opaque from the outside but mature inside, with guardrails, agent protection, and Party-defined standards for "safe AI."
PrismML's Bonsai 27B: Qwen3-27B Runs on a Smartphone for the First Time
Caltech researchers' startup PrismML compressed Qwen3.6-27B 14x with 1-bit weights. The 27-billion-parameter Bonsai 27B now runs directly on a smartphone, retaining…
The Vibe Coder's Cookbook: One LLM as Architect, Reviewer and Tester
Chinese Models Sweep OpenRouter's Top 5: MiMo-V2.5 Leads With 10.5 Trillion Tokens
Viterbi, beam search and LLM reranking: how an algorithm picks one meaning out of many
A Quadrillion-Parameter LLM: When to Expect One — and Why the Answer Is "Never"
MIT Sloan: AI Gives Good Financial Advice If You Ask the Right Questions
MIT Sloan researchers found that AI chatbots advise on personal finance at an expert level, but only when prompts are detailed and specific.
OpenAI Shuts Down Cambodian Fraud Network That Exploited ChatGPT
OpenAI blocked a Cambodian criminal network that used ChatGPT for investment fraud, romance scams, illegal gambling and impersonation.
avatarin's GPT-Realtime AI Agent Served 30,000 Yamada Denki Customers
Dutch insurer Univé rolls out ChatGPT Enterprise to its entire workforce
Peter Denning: Turing's Core Assumption About AI Has Been Wrong Since 1950
Microsoft Research's EvoLib Lets Language Models Learn From Experience Without Retraining
Tencent Open-Sources AngelSpec, a Framework for Speeding Up LLM Inference With Speculative Decoding
Tencent has open-sourced AngelSpec, a torch-native framework for training speculative decoding draft models that speeds up LLM generation without lowering answer…
Vercel's DeepsecBench: GPT-5.6 Leads at Finding Vulnerabilities in Application Code
Vercel has released DeepsecBench, an open benchmark that measures how accurately AI models find vulnerabilities in code. GPT-5.6 Sol leads, but cheaper models are…