Daphne Koller: Biomedicine and Machine Learning | Lex Fridman Podcast #93

Understanding varies by disease. While some are in the 70-80% range of comprehension, diseases like Alzheimer's and schizophrenia are closer to zero, with unvalidated hypotheses and complex, heterogeneous mechanisms.

There's a significant overlap, as the risk for many diseases increases exponentially with age. Cellular aging processes like DNA damage and protein accumulation contribute to disease.

Machine learning is crucial for analyzing vast biological datasets to build predictive models. Insitro, for example, focuses on creating high-quality data sets to train powerful ML models for identifying potential drug interventions.

Animal models often mimic the 'phenotype' (clinical outcome) but not the underlying mechanisms of human diseases. Cures found in animals frequently fail to translate to humans because the biological pathways are different.

This model uses induced pluripotent stem cells (iPSCs) derived from human cells, which can be differentiated into specific cell types. It allows researchers to study genetic disease manifestations in human cells directly, offering a more relevant model than animals.

Currently, generating iPSCs is not feasible at the scale of tens of thousands or hundreds of thousands. However, techniques like CRISPR for targeted mutations allow for valuable comparisons between healthy and diseased cells.

Advanced techniques like single-cell RNA sequencing allow measurement of gene activity in individual cells, turning biological processes into digital data. High-resolution microscopy also provides detailed sub-cellular information.

MOOCs emerged from early online course initiatives at Stanford University in the late 2000s. Daphne Koller and Andrew Ng launched viral courses in 2011, leading to the founding of Coursera in 2012 to address the growing need for accessible education.

Effective online learning prioritizes brevity (short modules), flexibility (learning at one's own pace), and engagement (quizzes, quick feedback). Breaking content into shorter units also provides a sense of accomplishment.

Calibrated uncertainty is crucial, especially in critical applications like medical diagnosis. Methods like Bayesian deep learning and ensemble networks aim to provide reliable confidence levels or indicate when a system 'gives up' on uncertain data.

Daphne Koller believes AGI is still very far off, as current ML systems excel only within specific domains. The more immediate concern is the danger posed by complex, poorly understood 'dumber' AI systems and the potential for technology abuse rather than AGI takeover.