Meta has released Code LLaMA 70B, its most powerful coding model to date, and it is likely the most powerful coding model available. Matthew Berman explores this new model, highlighting its capabilities and testing it to see if it can build the Snake game in one go. The video includes the following key points:
1. Announcement and Access: Meta’s AI at Meta team announced the release of Code LLaMA 70B, available under the same license as previous Code LLaMA models. The model can be requested for access through a form, and access is typically granted quickly.
2. Versions: There are three versions of Code LLaMA 70B – the base model, a version specifically trained for Python, and an instruct version fine-tuned for understanding natural language instructions.
3. Performance: The Code LLaMA 70B instruct model achieves 67.8 on human eval, making it one of the highest-performing open models available. It supports both research and commercial use.
4. Community and Fine-Tuning: The community has already started fine-tuning models based on Code LLaMA 70B, such as SQLCoder 70B, which outperforms GPT-4 in SQL generation tasks.
5. Testing: Matthew tests Code LLaMA 70B by attempting to build the Snake game using a virtual machine provided by MassedCompute. The model generates a significant amount of code, and while the initial attempt to run the game encounters issues, it demonstrates the model’s capabilities.
Matthew also mentions his investment in LM Studio and his use of MassedCompute to power the virtual machine for testing. He emphasizes the potential of Code LLaMA 70B in advancing AI-driven coding and invites viewers to suggest further tests.