972: In Case You Missed It in February 2026

Episode Summary

Executive Summary: This February 'In Case You Missed It' episode recaps three major AI conversations: Will Falcon on turning PyTorch Lightning into Lightning AI and the shift from research to startup scaling; Tom Griffiths on how human intelligence differs from AI under constraints and why inductive bias matters; and Amazon’s Nova Act plus Samsara’s quantum computing discussion, highlighting the push toward more reliable, production-ready, and eventually more autonomous AI systems.

Main Topics: From open-source project to startup at scale: Will Falcon explains how PyTorch Lightning unexpectedly became a company, driven by investor interest, heavy user demand, and the difficulty of balancing research with support requests and productization. Human vs. AI intelligence and constraint-based design: Tom Griffiths argues that human cognition is shaped by tight limits on time, compute, communication, and lifespan, while AI systems are not, leading to different forms of intelligence and raising design choices around human-like inductive bias. Nova Act and reliable agentic UI automation: Amazon’s Nova Act is presented as a developer-friendly service for building and deploying UI automation agents, emphasizing quick prototyping, playground testing, trace visibility, code generation, and AWS production deployment. Training agents with reinforcement learning web gyms: Nova Act’s reliability is attributed to reinforcement learning in simulated web environments, where agents self-play, fail, correct, and generalize across changing interfaces rather than relying only on imitation learning. Quantum computing for routing and optimization: Praveen Murugesan discusses how quantum computing could improve complex logistics optimization such as routing, vehicle scheduling, and real-time decision-making beyond today’s approximation-based methods. Toward self-operating systems: Across the clips, the episode returns to the idea that today’s AI mostly assists humans, but future systems may become more autonomous if reliability, reasoning, and compute constraints continue improving.

Key Arguments: Will Falcon argues that PyTorch Lightning became a company not by design but because the open-source project gained enough traction that investors and users forced a startup path. Falcon says his research slowed because he kept being pulled into coding for users and maintaining features, creating a conflict between academic work and product demands. Tom Griffiths argues that human intelligence is shaped by scarcity: short lifespans, limited neurons, and low-bandwidth communication force efficient learning and coordination strategies. Griffiths suggests AI systems can outperform humans because they are not bound by those same constraints, but this may also make their solutions less interpretable or human-like. Griffiths emphasizes that if we want AI to think more like humans, we need to engineer inductive bias, but that is harder than simply scaling models. Amazon’s Nova Act is positioned as a practical developer tool because it lets users prototype in a playground, inspect reasoning traces, generate Python code, and then deploy to AWS. Nova Act’s reliability focus is based on the claim that a 60% success rate is effectively useless in production, so the goal is 90%+ completion reliability. Praveen Murugesan argues that quantum computing could substantially improve route optimization and scheduling by solving much larger combinatorial problems with higher precision than classical approximation methods. Murugesan believes the near-term value of AI is still mostly assistive, and full self-operating systems remain limited by compute, data access, and real-time decision constraints.

Data Points: Episode number: 972 - Super Data Science Podcast 'In Case You Missed It' for February Original interview episode: 965 - Will Falcon / Lightning AI excerpt Original interview episode: 969 - Tom Griffiths excerpt Original interview episode: 963 - Nova Act excerpt Original interview episode: 967 - Praveen Murugesan excerpt Company revenue: over $500 million annually recurring revenue - Will Falcon describes Lightning AI’s scale after converting PyTorch Lightning into a startup VC term sheets: 10 - Falcon says he received about 10 term sheets by Monday after flying out on Thursday Research scale: 4 to 8,000 GPUs - Falcon describes pre-training work on very large GPU clusters Reliability threshold: 60% - Nova Act team says many flashy agents work only about 60% of the time Reliability goal: 90%+ - Nova Act is designed to reliably complete workflows at production quality Quantum time comparison: 10 septillion years to minutes - Murugesan cites the potential speedup of quantum computing for some problems Vehicles in routing example: 15 vehicles - Murugesan discusses the vehicle routing problem Stops in routing example: 20 stops - Murugesan uses a route-planning example Research preview start: March last year - Nova Act had been in research preview since March before launch Launch event: AWS re:Invent in December - Nova Act was launched as a GA service at AWS re:Invent

Pivotal Quotes: "I did not want to start a company, first of all." — Will Falcon: Explaining that Lightning AI emerged from an open-source project he initially did not intend to commercialize "A reliable 60% is 0% useful." — Anche Bhart: Why Nova Act focused on production reliability for agents "Scaling is much easier than engineering inductive bias." — Tom Griffiths: Describing why building human-like AI behavior is harder than simply making models bigger

Implications: The episode highlights a broad industry shift: from research tools becoming startups, to agents needing production-grade reliability, to longer-term bets on quantum and self-operating systems. The common theme is that better AI will depend not just on scale, but on engineering, robustness, and better problem framing.