August 2024 Summaries

The exploration of using Large Language Models (LLMs) to create a web-based version of the game show Family Feud, dubbed "Featherless Feud," showcases both the potential and challenges of integrating AI into game development. The project highlights two primary challenges: generating relevant question and answer sets and implementing fuzzy answer matching to accommodate variations in player responses. By employing LLMs to prompt for question and answer generation in structured formats like JSON, the developers were able to automate a process traditionally reliant on human input and creativity. Despite the structured approach, achieving the game's trademark comedic subtleties proved elusive even with advanced models. The developers experimented with various models and inference providers, including featherless.ai, which offers serverless access to a wide range of models, allowing for dynamic updates and customization. While larger models provided nuanced judgment, their slow response times were less conducive to gameplay, leading to the selection of models like Sao10K/L3-8B-Stheno-v3.2 for their balance of speed and quality. This endeavor underscores the complexity of capturing human-like creativity and humor in AI applications, inviting further exploration and feedback from users.

Featherless is an innovative serverless inference provider for Hugging Face, gradually expanding its support for various model architectures since its launch in June. The platform recently included Mistral Nemo 12B and Qwen2 32B, bringing the total number of supported architectures to 12 and nearly 2,000 inferencible models. Smaller models, such as the 7B and 8B, are more prevalent due to their lower resource demands for fine-tuning, whereas older models have more fine-tuning iterations, explaining the numerous Llama 2 adaptations compared to fewer Nemos and Qwens. The company encourages community involvement through Discord for feedback and suggestions on future model support, and it plans to introduce new features beneficial to model creators involved in fine-tuning.