955: Nested Learning, Spatial Intelligence and the AI Trends of 2026, with Sadie St. Lawrence

Episode Summary

Executive Summary: Sadie St. Lawrence and John Crone review 2025 AI trends, award the year’s biggest wins and misses, and forecast 2026. They argue intelligence is becoming radically cheaper, agents are overhyped in enterprise, Google resurged, NVIDIA remains a core winner, and the next wave will be specialized models, continual learning, spatial AI, research breakthroughs, and AI ops.

Main Topics: 2025 trend recap and award segment: They assess how Sadie’s 2025 predictions played out and then award the year’s wow moment, comeback, disappointment, and overall winner, using the recap as a bridge into 2026 forecasting. AI becoming radically cheaper: The hosts emphasize that the same capability now costs far less than it did a year ago, arguing that declining intelligence cost is the central force reshaping business models, work, and product strategy. AI product and platform adoption in daily life: They discuss integration into everyday devices and software, including AirPods translation, Google search/G Suite AI, Tesla Grok use, and consumer-facing image tools like Nano Banana. Enterprise AI, agents, and monetization: They debate the gap between agentic AI hype and production reality, noting that many enterprise agent projects remain stuck in MVPs and that ROI and implementation maturity are now critical. Future of AI research and specialized models: Sadie predicts the next breakthroughs will come from domain-specific models, continual learning, and renewed research emphasis rather than only scaling frontier general-purpose systems. Spatial intelligence, robotics, and physical AI: They expect 2026 to be a bigger year for spatial intelligence and simulation data, but Sadie says the true year of the robot will likely be 2027 rather than 2026. Jobs and operational roles in AI: They identify rising demand for AI engineering and forecast a new AI ops role focused on GPU infrastructure, model orchestration, and agent reliability.

Key Arguments: Intelligence is becoming dramatically cheaper, and that cost curve will force businesses to rethink strategy, margins, and product design. Agentic AI dominated 2025, but enterprise adoption lagged behind hype because many organizations are not set up with the right data, security, or workflow structures. AI integration into consumer devices is real, but quality and usefulness vary; not every product benefits from adding AI everywhere. AI has made meaningful progress in scientific domains, but some headline results were overstated or redacted, showing the field is still maturing. The data scientist role has splintered into a family of more specialized jobs, and AI engineer/prompting roles are rising faster than traditional data science roles. The next major leap is likely to come from specialized industry models and continual learning, not just bigger general models. Robotics and physical AI need new real-world and simulated datasets before they can scale meaningfully; 2027, not 2026, may be the breakout year. AI operations will become its own practical discipline as organizations manage GPUs, orchestration, and reliability at scale.

Data Points: AI capability cost reduction: $100 a year ago to $1 now - Used in the opening thesis that intelligence is becoming radically cheap. Workshop demo build time: 90 minutes - Ed Donner’s agent team built a functioning stock trading platform during a lunch break. AI agent team size: 4 agents - Front-end developer, back-end developer, tester, and project manager in the ODSC demo. HMCI research team size: 3 people - Sadie described HMCI as a small, AI-empowered team. Women in Data membership: 70,000+ individuals - Sadie reported the organization’s current membership scale. Women in Data chapters: 55 chapters - The organization’s chapter footprint. Women in Data geographic reach: 100+ countries - Membership representation across countries. Podcast episode: Episode 955 - This conversation is part of the Super Data Science Podcast. TEDx talk engagement: 300+ comments, 1,000 likes, 30,000+ views - John highlighted the response to Sadie’s TEDx talk on consciousness and human-machine collaboration. Reliability estimate for current AI tasks: ~50% - John said current agentic systems can handle some software/ML tasks about half the time. Task-duration growth: Doubles every 7 months - John described the rate at which agentic AI can replace longer human tasks. Projected task duration by end of 2026: ~8-hour workday - John extrapolated current scaling trends to end-of-2026 performance. Material science discovery lift from AI: 39% more - Mentioned as a paper that later got redacted. DeepMind genome/material acceleration: ~800 years compressed into a few hours and steps - Used as an example of AI accelerating scientific discovery. LinkedIn jobs report: AI prompting and AI engineering keywords rising - Sadie cited LinkedIn data to support rising demand for AI-focused skills. NVIDIA market performance: Share price described as "incredible" and "crazy" - Used as evidence for NVIDIA being an overall AI winner. Google documentary length: 1 hour 24 minutes - The Thinking Game documentary was noted as a free, full-length YouTube release. Thinker models benchmark note: LM Arena ELO difference not statistically significant - John explained why Gemini 3 Pro and Grok were effectively tied in one ranking context.

Pivotal Quotes: "Intelligence is becoming cheap, radically cheap, and that changes everything about how we should be thinking about the future." — John Crone: Opening framing for the episode’s central thesis. "The whole point is, we should ask hard questions that we normally don't ask, and our future is closer than we think." — Sadie St. Lawrence: Sadie explaining the theme of her TEDx talk and her approach to consciousness research. "For me, actually, agents were disappointing." — Sadie St. Lawrence: Her 2025 disappointment-of-the-year pick, focused on the gap between hype and enterprise reality.

Implications: Listeners should expect cheaper intelligence, more specialized AI systems, stronger AI ops and engineering demand, and slower but real progress in robotics and continual learning. Enterprises that do not adapt workflows, data, and security will fall behind.