September 2023 Summaries

Mistral recently launched its first model, Mistral 7B, which has shown promising initial results by outperforming Llama 2 7B on academic NLP benchmarks, particularly in legal reasoning tasks where it achieved a 59% pass rate compared to Llama 2's 50%. However, Llama 2 excelled in creative writing tasks with an 83% pass rate, suggesting its suitability for open-ended and conversational settings. While Mistral demonstrated strong reasoning capabilities, it generated toxic outputs in 14% of safety tests, highlighting the need for improved guardrails. The Mistral team encourages the community to implement safety measures, and Patronus AI is actively engaged in developing scalable evaluation methods for large language models (LLMs) to assess their real-world applicability.

Patronus AI is addressing the challenges associated with testing and evaluating large language models (LLMs), which are known for their unpredictability despite showing impressive capabilities. Traditional methods of manual inspection for model evaluation are insufficient for high-stakes applications, as they rely on subjective judgment and limited test coverage. The complexity of potential input combinations makes exhaustive testing impractical, especially with multimodal models. While human evaluation is valuable, it is often costly and time-consuming. Patronus AI offers scalable automated techniques for creating challenging test cases and evaluating results, thus enabling enterprises to effectively assess the performance of LLMs and trust their outputs for critical tasks.

Generative AI has seen rapid consumer adoption, exemplified by ChatGPT's record-breaking growth, but enterprises remain cautious due to concerns over language model (LLM) errors and the challenges of scalable evaluation. Patronus AI, founded by Rebecca and Anand, aims to address these issues by providing an automated evaluation and security platform for LLMs, helping enterprises detect errors, generate tests, and benchmark models for safe deployment. With a $3M seed round led by Lightspeed Venture Partners, Patronus AI is poised to enhance enterprise confidence in AI adoption by acting as a trusted third-party evaluator, akin to Moody's for AI, and is already partnering with leading AI and traditional companies. The team, with strong backgrounds in applied ML from institutions like Meta and Airbnb, is committed to advancing LLM evaluation and robustness, addressing gaps in current testing frameworks, and overcoming the long tail of AI failures to ensure production-ready AI usage.