Large Language Models

Anthropic’s new Message Batches API lets developers send up to 10,000 queries per batch, processed within 24 hours at half the cost of standard calls.

MIT study: AI models in home surveillance inconsistently detect crime, recommend police calls, and show demographic biases in decision-making

MIT CSAIL’s Co-LLM algorithm lets a general-purpose language model hand individual words to an expert model when needed, improving factual accuracy.

OpenAI has released o1, models trained to spend more time thinking before answering, scoring 83% on an IMO qualifying exam against 13% for GPT-4.

Google has released DataGemma, open models that use statistics from Google’s Data Commons, through RIG and RAG methods, to reduce LLM hallucination.

Anthropic has launched Claude Enterprise, with a 500K-token context window, a native GitHub integration and enterprise security controls.

Meta says its Llama models are nearing 350 million downloads, a tenfold rise in a year, with more than 20 million on Hugging Face in the past month.

Anthropic has made Artifacts, which lets users view and iterate on work created with Claude, available to Free, Pro and Team plan users.

MIT researchers have built SigLLM, a framework that uses large language models to detect anomalies in time-series data without extensive training.

MIT researchers found an LLM trained on Karel puzzle solutions reached 92.4% accuracy, suggesting it built its own internal simulation of the task.

Mistral AI now lets developers customise models such as Mistral Large 2 and Codestral on La Plateforme, and has released an early version of AI Agents.

Meta AI has published a guide to adapting LLMs, recommending parameter-efficient fine-tuning as a more viable route than pre-training for smaller teams.

Meta AI’s guide sets out five cases where fine-tuning helps, including tone and format, accuracy, niche domains, distillation and new abilities.

Meta AI’s fine-tuning guide says dataset quality matters most, with small high-quality datasets often beating large low-quality ones.

Stanford researchers found that current chatbot alignment methods carry a Western bias that weakens performance for users in other languages and regions.