Interview, Fireside Chat, Podcast
a16z Podcast | Revenge of the Algorithms (Over Data)... Go! No?
- Historical patterns suggest board game achievements may trigger "AI winter" fears by fostering fallacious assumptions that general intelligence is imminent, while current capabilities are expected to remain discrete without immediate cross-domain generalization.
- AlphaGo Zero's techniques are projected to apply to drug discovery, protein folding, quantum chemistry, and material science, though their efficacy remains uncertain in domains lacking defined rules or loss functions, with no clear consensus on whether reinforcement learning can generalize to such unstructured problems.
- Future solutions are anticipated to be hybrid models combining established methods with new approaches, including machine learning coexisting with traditional natural language processing for tasks like spell-checking, while simulating millions of interactions will drive benefits in sales forecasting, cybersecurity, and robotics provided constraints can be codified.
- Practical success will likely rely on the combination of data and algorithms, such as in ride-sharing navigation, mirroring Google's historical reliance on web crawling barriers rather than unique algorithmic secrets, as data remains an inescapable input despite occasional rule-based victories.
- Societal and commercial perceptions are expected to shift toward accepting machine error similar to human error, though current low tolerance for algorithmic mistakes creates friction, with bias primarily stemming from incomplete human-selected data sets and cultural labeling choices rather than inherent neural network architecture.
- The AI landscape is described as a "treadmill" of constant advancement where founders must navigate a lack of a "meta algorithm" for strategy selection over the next two years, requiring them to draw specific lines from problem to solution rather than chasing investor hype, while opportunities exist in the non-standardized "art" of configuring neural network layers and parameters.