Last Week in AI AI for COVID detection fails, GitHub's Copilot can code, GAN Theft Auto is fun
Jul 1, 2021
This week, Abigail C, a Stanford PhD and friend of the hosts, shares insights on the latest AI news. They dive into why certain AI models for COVID detection are deemed unfit for clinical use and discuss GitHub’s new AI tool that can write code on its own. The conversation takes a fun turn with a look at GAN Theft Auto, where AI hilariously recreates elements of GTA 5. The team also critiques Amazon’s warehouse robots and the biases plaguing LinkedIn’s job-matching AI, emphasizing the need for ethical scrutiny in AI development.
AI Snips
Chapters
Transcript
Episode notes
Model Optimization Trends
- Google's survey emphasizes optimizing deep learning models for size, speed, and quality.
- This signifies a shift beyond solely focusing on accuracy metrics.
Model Selection Strategy
- Choose efficient architectures for resource-constrained training.
- Compress large, high-quality models post-training.
GPT Model Size in Practice
- Abby used smaller GPT models for the Alexa Prize chatbot due to latency constraints.
- GPT-2 medium offered a balance between size and speed for practical deployment.