LLMs

MIT's SigLLM framework uses large language models to detect anomalies in time-series data without extensive training, showing promise for complex systems.

MIT researchers found LLMs may form internal simulations of reality while improving language skills. A model achieved 92.4% accuracy on Karel puzzles without direct exposure, suggesting deeper understanding beyond mimicry.

Mistral AI enables model customisation via La Plateforme and introduces AI Agents, allowing developers to create complex, shareable workflows.

The guide suggests fine-tuning, particularly parameter-efficient fine-tuning, as a more viable approach for smaller teams with limited resources compared to pre-training methods.

The guide identifies five scenarios where fine-tuning excels: customising tone and format, improving accuracy, addressing niche domains, reducing costs via distillation, and developing new abilities.

Meta AI's guide emphasises dataset quality for fine-tuning LLMs, suggesting small high-quality datasets often outperform larger low-quality ones. It compares full fine-tuning and PEFT techniques.