Essential AI News from 7/1 - 7/5
It was a shortened and light news week. While no articles warranted an "essential" rating, there were a number of important articles. Below are some of the most important as well as a "bonus" article that didn't make the prior summaries.
- ElevenLabs adds AI voice of celebs to new digital narrator — but is it safe? Companies need to start staking out the voices they want to own for their brand. The voice element of interactions are least developed. And voices are to the coming audio wave as generic URLs (eg, cars.com which sold for millions) were to the early internet. Listen to John Sviokla's interesting points on this at minute 24:23 of our morning briefing.
- Scaling Synthetic Data Creation with 1,000,000,000 Personas The TenCent researchers demonstrate that you can train models with more diverse synthetic data than otherwise possible using this method. Everything is done by LLMs here - the creation of the personas, the testing, etc. A 7B model that matched the performance of GPT-4 Turbo performed this work. Important to know that there's a method to scale synthetic data creation.
BONUS: Potential New Benchmark for Agentless AI Software Development A team from the University of Illinois built an agentless autonomous software development AI that beat other state-of-the-art, open-source models for performance and cost on the SWE-bench Lite benchmark. Notably, this reinforces the importance of matching problem complexity with model complexity.