Large Language Models

OpenAI is testing SearchGPT, a prototype that combines its AI models with web information to give quick answers with clear source links.

OpenAI’s Rule-Based Rewards method uses simple rules to grade model outputs, improving safety behaviour without collecting large amounts of human data.

NVIDIA has launched four NeMo Retriever NIM microservices that connect LLMs to business data for RAG, giving 30% fewer inaccurate answers.

Mistral and NVIDIA released Mistral NeMo, a 12-billion-parameter model with a 128k context window under Apache 2.0, trained with quantisation awareness for FP8 inference.

Mistral AI has released Mathstral, a STEM-focused model built on Mistral 7B that leads its size class on maths reasoning benchmarks.

OpenAI’s Prover-Verifier Games train strong models to produce answers that weaker models, and so people, can check more easily, such as maths solutions.

Mistral AI has released Codestral Mamba, an open-source code model with linear-time inference that can in theory handle infinitely long sequences.

Anthropic has added tools to its Developer Console, powered by Claude 3.5 Sonnet, that help developers generate, test and grade the quality of prompts.

Anthropic has launched a programme to fund third-party evaluations of advanced AI capabilities and safety, from cybersecurity to societal impacts.

OpenAI’s CriticGPT, based on GPT-4, finds mistakes in ChatGPT’s code, and reviewers using it outperformed those without help 60% of the time.

OpenAI has described a content moderation approach that combines content taxonomies, data quality control and active learning to detect undesired content.

Mistral AI has introduced tools and services to make customising its models more efficient and affordable, with three entry points for developers.

OpenAI researchers have developed new methods for interpreting neural activity in language models, giving a view of the concepts behind GPT-4.

Reddit's deals with AI companies for training data highlight the growing value of user-generated content and raise questions about data ownership.

Niantic, the company behind Pokémon GO, uses Meta’s Llama 2 in its AR game Peridot so its virtual pets can react and converse with players in real time.