October 2022 Summaries

Humanloop is collaborating with Carper AI, a Stability AI company, to develop a groundbreaking 70 billion parameter open-source large language model (LLM) that employs Reinforcement Learning from Human Feedback (RLHF) to enhance safety and usability in AI systems. This initiative aims to democratize the "instruction-tuning" of LLMs by adapting them for specific tasks through direct human feedback, making AI interactions as seamless as instructing a colleague. The project involves partnerships with Scale and Hugging Face, with the latter hosting the final model to make it widely accessible. Although traditional LLMs excel in tasks like code generation and writing assistance, their reliance on next word prediction often leads to inaccurate outputs and potential misuse. Training with RLHF addresses these issues by aligning models more closely with human feedback, thereby reducing risks such as misinformation and social bias while enhancing the models' practical utility. This open-source release, a pioneering effort in the field, is expected to drive extensive research and innovation, paving the way for new applications and companies to explore state-of-the-art AI systems.

Humanloop, co-founded by Raza Habib, is opening access to its platform designed to assist developers in creating applications using large language models like GPT-3, which is seen as a revolutionary computing platform comparable to the internet and smartphones. The platform aims to bridge the gap between prototyping and full production applications by addressing challenges such as subjective evaluation, complex prompt engineering, and the customization of models. Humanloop offers tools to capture user feedback, experiment with prompts, and fine-tune custom models, allowing developers to optimize application performance and build defensible, AI-first products beyond the base capabilities of models like GPT-3. The initiative seeks to empower developers to harness the potential of large language models, fostering innovation in various domains such as intelligent assistants and interactive design tools.