Large Language Models

Mistral AI explains how to score retrieval-augmented generation systems using an LLM as judge with structured outputs, covering retrieval and answer quality.

Alibaba has released Qwen 2.5-Max, claiming it outperforms DeepSeek-V3, as Chinese AI developers respond to DeepSeek’s V3 and R1 releases.

Google has unveiled Gemini 2.0, a model designed to understand and act on information across formats, as Sundar Pichai described a new agentic era of AI.

OpenAI actively seeks legal protection for innovations while building tech using copyrighted materials, highlighting contradictions in AI's intellectual property landscape.

Stanford researchers tested language models with 8,000 questions on 300 topics and found answers less consistent on controversial subjects.

Mistral AI has launched a batch API on La Plateforme that processes high-volume requests at half the cost of standard synchronous API calls.

"New Grok-beta model features 128k token context and function calling, with REST API compatibility with OpenAI/Anthropic. Multi-modal version coming soon."

OpenAI has upgraded ChatGPT with web search that returns real-time sports scores, news, stock quotes and weather, with links to sources.

"AI systems went from solving 2% of real coding problems to 49% in one year. Anthropic warns the window for safe regulation is closing fast."

OpenAI has released SimpleQA, a benchmark of 4,326 fact-seeking questions that measures how accurately language models answer and how often they hallucinate.

A Stanford-led study compared 50 physicians with and without ChatGPT-4 on real patient cases, and found the model alone scored higher on diagnostic reasoning.

Mistral AI has released Ministral 3B and 8B, known as les Ministraux, small models for on-device and edge use with context windows of up to 128k tokens.

Stanford researchers found GPT-3.5 and GPT-4 can assess how well instructional materials work and produce lesson content that human teachers preferred.

NVIDIA says software work cut Llama 70B latency 3.5 times in under a year, and Blackwell delivered a fourfold gain and the first FP4 result in MLPerf.

Anthropic implements measures to prevent AI misuse in 2024 US elections, including policy updates, detection systems, and redirects to voting information.