Interview, Podcast
a16z Podcast | AI, from 'Toy' Problems to Practical Application
- Immediate real business impact is achievable through AI go-to-market approaches, with Fortune 100 companies transitioning systems from R&D experimentation to production where optimization is critical for profitability.
- Financial, healthcare, and life sciences sectors are actively applying AI techniques for disruption and preventative maintenance, utilizing longer time series sensor data to extend predictive model horizons for machinery failures.
- Over 2 million customers are currently on the Amazon Web Services platform, supporting an infrastructure layer expected to be free or the lowest common denominator compared to past IT models.
- Five to six hundred internal use cases are being tracked to identify those salient enough to drive ROI, while startups focusing on product-market fit without defined company building are deemed to be engaging in wishful thinking.
- Complex AI systems are growing exponentially in size and computational intensity, becoming twice as large as previous iterations and requiring custom, non-transferable optimization tricks for specific cases.
- Zero-to-one AI solutions are powerful for small and medium-sized businesses, but transitioning to a one-to-two stage requires bespoke knowledge and custom datasets, whereas untuned sophisticated systems may underperform tuned simple ones.
- Deep reinforcement learning for autonomous driving is not tractable as a standalone algorithm due to the infeasibility of a "crash a million times" approach, and simulating driving in the cloud to transfer domains remains an unsolved problem.
- Successful AI startups demonstrate traction by applying AI to vertical problems with proprietary data, necessitating a combination of research expertise and true domain experts, while reinforcement learning algorithms can generate cumulative metrics from scratch.
- One researcher can now achieve work previously requiring a team of a decade, and basic NLP tasks like sentiment analysis can be completed via RESTful API calls without building custom corpora, allowing companies to outsource everything except core differentiators.
- While simple algorithms like random forests consistently provide a "B minus answer," tuned sophisticated deep learning algorithms can beat humans in practical applications, and startups must leverage proprietary data within specific verticals to succeed.