Greetings,

There are some minor signs of an AI investment bubble, and building products with Large Language Models isn’t as easy as it would seem. But we’re still headed in a great direction because the pace of progress is unrelenting and we keep seeing exciting new models like Meta’s Voicebox, which can quickly imitate your voice.

• Mistral is a Paris-based AI startup that raised €105m without a product just four weeks after its launch – the founders want to be AI leaders in Europe and have solid experience in big-name companies, but some people see this as a sign of an AI investment bubble. - Sifted
  
  
• Product development with LLMs brings unique challenges, including limited input capacity, slow response times, unpredictable results from complex prompts, security risks from prompt injection, and legal compliance concerns. - Honeycomb.io 
  
  
• I-JEPA (Image Joint Embedding Predictive Architecture) is a new AI model proposed by Meta's Chief AI Scientist, Yann LeCun, and it can learn about the world, plan complex tasks, and adapt to unfamiliar situations – so far it performs well on computer vision tasks, and it’s open-source so you can play around with it. - Meta AI
  
  
• Meta researchers also unveiled a speech-generating model, Voicebox, which can speak in your voice based on a recording and outperforms similar models by up to 20 times – this one isn’t open-source though, because of the obvious risk of being used for fraud. - Meta AI
  
  
• Most of 10 tested foundational models failed to meet even 50% of the EU’s AI Act requirements in a Stanford study – researchers believe that enforcing these requirements will significantly benefit the AI ecosystem. - Stanford
  
  
• Self-healing code is a new concept emerging at the intersection of LLMs and programming – as models are fine-tuned to become better at programming tasks, it seems that AI could potentially review and fix code in real-time, saving enormous amounts of time for human reviewers. - Stack Overflow
  
  
• a16z has produced a good reference for understanding the complexity of the AI app development stack, which already has 12 categories of tools (data pipelines, embedding models, vector databases, playgrounds, orchestration, APIs/plugins, LLM cache, logging / LLMops, validation, app hosting, LLM APIs, cloud providers). - a16z

----

Found this helpful? Forward this email to a colleague.

Was this email forwarded to you? Sign up here to stay on top of the AI news.