EP 140 : Healthcare AI Breakthrough: Cancer Prediction, Brain-Like Models & HealthBench

We dive into the latest developments in artificial intelligence starting with today's main points. Discover how Mass General Brigham’s FaceAge AI is revolutionizing cancer survival predictions by analyzing facial photographs. This tool translates subtle facial characteristics into a biological age estimate, finding that cancer patients often appear older, which correlates with worse survival rates. Adding FaceAge risk scores improved doctors' accuracy in predicting 6-month survival, and the AI's predictions aligned with a gene associated with cellular aging.Next, explore Sakana AI’s Continuous Thought Machines (CTMs), a nature-inspired approach teaching AI to "think" step-by-step over time, much like our brains. This differs from current AI systems' instant decisions and draws inspiration from how neuron timing is crucial for intelligence. Sakana demonstrated CTMs solving complex mazes and adapting processing time based on task difficulty in image recognition.We also look at OpenAI’s HealthBench, a new benchmark created with 262 physicians to evaluate AI systems in health conversations. HealthBench tests models across various themes and behaviors like accuracy and communication quality. Recent models performed significantly better, with OpenAI's o3 scoring 60% compared to GPT-3.5 Turbo's 16%, and smaller models like GPT-4.1 Nano showing improved capability and cost-effectiveness. OpenAI has open-sourced the evaluations and a dataset of 5,000 realistic health conversations. Having physician-validated benchmarks is crucial for measuring AI performance and deciding deployment in healthcare.Following these updates, we cover "Everything else in AI today", including news about Manus AI agent access, Google DeepMind's AI Futures Fund, Softbank's Stargate investment being stalled, Perplexity's reported $500M funding round, Carnegie Mellon's LegoGPT, Saudi Arabia's new AI venture, Humain, aiming to become an AI hub, and the US FDA's plans to deploy AI agency-wide. We also touch on OpenAI allowing subscribers to export deep research reports as PDFs, Saudi Arabia potentially receiving advanced Nvidia chips, and Meta's experimental byte-based LLM.Finally, we highlight Today’s Trending AI Tools. Examples include HunyuanCustom for generating custom videos, Granola iOS as an AI notepad for back-to-back meetings, Deep Research connecting GitHub repos to OpenAI, Deer Flow as an open-source research tool, Mendel for automating code reviews, Seveum for job matching, Clockwise as an AI calendar assistant, DeckSpeed for creating presentation slides, and LLMRefs for tracking keyword rankings and optimizing AI SEO.