Interview, Fireside Chat
Daphne Koller: Biomedicine and Machine Learning | Lex Fridman Podcast #93
- Cures for major diseases are expected eventually, though the timeline is described as "a very long time" with significant hurdles in regenerating entire body parts to their original state, and fundamental understanding for the majority of diseases is currently estimated as being "really close to zero," with Alzheimer's and schizophrenia specifically noted as likely closer to zero than 80.
- Specific disease understanding varies, with Type 2 diabetes presenting a mix of validated mechanisms like insulin resistance and many unknown factors, while autoimmune disease treatments have grown to include "probably four or five, maybe even more" drugs compared to only a few existing 12 years ago.
- Scientific efforts will increasingly overlap with longevity goals, as the risk of contracting most diseases (excluding childhood illnesses) "increases exponentially year on year, every year from the time you're about 40," driving aspirations for increased health spans rather than just extending the biblical age of 120.
- Machine learning is predicted to shift from playing a "significant role" to becoming central to drug discovery by prioritizing the creation of large-scale data sets to build predictive models, a transition driven by the recent existence of necessary data at scale which was previously unavailable.
- Traditional animal models are deemed "not very effective" for conditions like Alzheimer's, diabetes, and schizophrenia because these species do not naturally contract these diseases, whereas "disease in the dish models" enabled by technologies existing only in the last five to 10 years offer superior clinical relevance.
- Induced pluripotent stem cells are considered "almost certainly" possible to generate but currently limited to a global capacity of "somewhere between five and 10,000" cells, far below the tens of thousands or hundreds of thousands needed for scaling.
- Polygenic risk scores indicate that disease risk for individuals in the "highest decile" can be "a factor of 10 or 12 higher" than those in the lowest decile, with cellular models from diverse genetics expected to provide more signal than direct genetic analysis alone.
- Future advancements in bioengineering are projected to enable "disease models that we could make for things that we can't make today" within "three years," and "organoids" which did not "really exist" five years ago can now be derived "reasonably robustly" with further goals to connect them into "multi-organ system stuff" despite many challenges.
- Neural networks are predicted to find meaningful representations in "exceedingly high dimensional space" with large data, though a "completely knowledge-free approach" is not the current solution, and creating networks calibrated to admit uncertainty is deemed critical for mission-critical applications where human life is at stake.
- Safety concerns highlight that "gene editing and CRISPR" poses risks "at least as dangerous a technology if used badly than machine learning" due to potential viral creation, while machine learning itself carries risks of unpredictability in complex systems, though machines remain "nowhere close to the versatility and flexibility of even a human toddler."
- Societal expectations for education involve increased consumption of online content for continuing education, with a recommendation that machine learning practitioners not skip foundational mathematics to avoid errors, while MOOCs are not expected to replace face-to-face teaching due to the social nature of learning.
- Long-term predictions regarding artificial intelligence describe the protection of "human level or superhuman level intelligence" as "very speculative" given the lack of visible blueprints, emphasizing the need for social norms that positively correlate doing good with peer perception to prevent a degeneration into environments where harmful acts are accepted.