• Distilling step-by-step is a simple mechanism that enables a tiny 770M parameter model to outperform a 540B PaLM model, a 700x model size reduction while preserving performance. - Google
• Surprising alliance – Amazon is investing up to $4 billion in Anthropic to compete with Microsoft and Google, they’ll offer early access to unique features to AWS customers. - CNBC
• Your chance to work with OpenAI – experts from around the world and from all sorts of specializations (biology, psychology, law, many more) can apply to join OpenAIs Red Teaming Network to help improve the safety of their models. - OpenAI
• For less than $1,000, Xiaoice creates AI clones that are bringing in millions of dollars in sales for Chinese livestreamers – they want to create “100,000,000 silicon-based laborers” by 2025, so far they’ve created 400,000 of them. - MIT
• GPT to validate, Llama to deploy – GPT3.5 is 4-6x more expensive to fine-tune, but it’s more convenient if you want to validate your approach, Llama 2 is much less costly and provides full training and deployment flexibility so it could be better for a production app. - Ragntune
• Replit’s ModelFarm is a rapid way to build genAI apps – no need to worry about API key management because it’s all handled by the service, you can build your first genAI app in 5 lines of code. - Replit
• Put a leash on LLMs – validating LLM outputs is a must to avoid unexpected behaviors and ensure a reliable user experience in production apps, here’s a guide on how to do it using the open-source Guardrails AI package. - Cohere
• MidJourney’s advantage is that it provides distinctive quality that’s difficult to describe or compare in a spreadsheet – but DALL-E 3 is a strong competitor, since it’s built on ChatGPT and can leverage all of its power and user base (100M compared to MJs 16M). - X (Twitter)
• Spotify is launching a pilot feature that translates podcasts into different languages while maintaining the original voice. - X (Twitter)
