#LLM
Curated coverage on «LLM»: releases, research, benchmarks, open-source.
Google launches Gemini 3.5 Live Translate: speech-to-speech model for 70+ languages
Gemini 3.5 Live Translate translates speech in real time, synthesizing audio with a lag of several seconds. The model is available in Gemini Live API, Google Meet…
USAF: even a novice can write an AI application for combat use
MIT researchers found that AI chatbots help non-technical military personnel write applications for their tasks. The barrier to entry in development has dropped to…
Anthropic Opened a Window Into Claude's Thinking: Jacobian Lens Tool Catches Hidden Concepts
Anthropic developed a method that shows exactly how Claude reasons when answering a question. Hidden concepts in intermediate layers of the model are visible.
OpenAI Launched ChatGPT Work — Agent for Taking Actions in Applications
FINESSE-Bench: Evaluating LLMs in Financial Domain Without Deception
OpenAI Releases GPT-5.6 and ChatGPT Work: Agent for Multi-Stage Tasks
The Error Isn't in the LLM: How Architecture Masks Problems as Model Quality
Tencent Open-Sourced LLM Hy3 with 295 Billion Parameters
Tencent released LLM Hy3 under Apache 2.0 license — a model with 295 billion parameters (21 billion active) with 256K context, competing with much larger open…
Raiffeisenbank deployed AI agents to 500 engineers, but speed didn't increase. Why
The German bank found that local acceleration of engineering work via AI doesn't transform the system without restructuring processes overall.
The New York Times demands sanctions against OpenAI for evidence concealment
OpenAI Introduced GPT-Live — Updated Voice ChatGPT with GPT-5.5 Support
SageMath Improved LLM-Agents in Mathematics by 9.7pp on Average
DeepSeek V3.2 Achieves 67% Accuracy on the ARC-AGI Abstract Reasoning Benchmark
SynthAVE: when 21 models labeled 12 thousand products instead of humans
The SynthAVE dataset: 12,726 products in 4 languages annotated by an LLM arena of 21 judges. Results match manual annotation at 95.2%.
DeLS-Spec accelerated LLM inference without retraining models
Researchers proposed DeLS-Spec—a method for accelerating LLM inference by adding a lightweight adapter trained separately from the base model.