Signal
40 articles covering "llm"
A two-year follow-up to the WorkBench benchmark finds the best workplace agent jumped from 43% to 89% task completion while unintended harmful actions fell from 26% to 2.5%.
Zafran Security disclosed four authorization flaws in the open-source LLMOps platform Dify, including two critical bugs that let an attacker redirect another tenant's AI conversations to an attacker-controlled endpoint.
Zhipu's GLM-5.2 ships under an MIT license as a frontier coding alternative just after the US barred foreign nationals from Anthropic's Fable 5 and Mythos 5.
A new MIT CSAIL inference method has a model inspect its prompt in a Python REPL and recursively call itself over snippets, processing inputs beyond its context window.
A DeepMind preprint reports an LLM-and-Lean agent that autonomously solved 9 of 353 open Erdős problems and proved 44 of 492 OEIS conjectures for a few hundred dollars each.
Moonshot AI released Kimi K2.7-Code, an open-weight 1-trillion-parameter MoE coding model under a Modified MIT License, claiming roughly 30% lower thinking-token usage than K2.6 on self-run benchmarks.
Anthropic launched Claude Fable 5 on June 9, a public version of its restricted Mythos model that falls back to Opus 4.8 on high-risk topics in under 5% of sessions.
Microsoft unveiled seven proprietary MAI models led by the 35B-active reasoning model MAI-Thinking-1, all trained from scratch without distillation as it reduces its dependence on OpenAI.
The 550-billion-parameter model activates 55 billion parameters per token, ships under the Linux Foundation's OpenMDW-1.1 license, and trades blows with China's Kimi-K2.6 on coding benchmarks.
Shanghai-based MiniMax launched M3 on June 1, pairing a 1-million-token context with a new sparse-attention design and company benchmarks that top GPT-5.5, with weights promised within 10 days.
Qwen3.7-Max debuts at Alibaba's Cloud Summit in Hangzhou as the company's most capable agent model, sustaining continuous autonomous execution for 35 hours in internal testing.
Google's new efficiency flagship outperforms Gemini 3.1 Pro on most evals while running 4x faster and costing 40% less.