849: 2025 AI and Data Science Predictions, with Sadie St. Lawrence

Episode Summary

Executive Summary: Sadie Saint Lawrence and John Crohn review 2024 AI/data science trends, naming Google as comeback of the year, OpenAI as both disappointment and wow-moment, and NotebookLM/Waymo as standout experiences. They then predict 2025 will be dominated by agentic AI, AI in everyday devices, AI-powered scientific discovery, enterprise monetization pressure, and a shift toward AI engineering skills over traditional data science roles.

Main Topics: 2024 trend retrospective: They revisit Sadie’s 2024 predictions and assess accuracy across GPU demand, slow-thinking models, tool consolidation, and workplace upheaval. Comback, disappointment, and wow moments: Google is chosen as comeback of the year; OpenAI and Apple Intelligence are flagged as disappointments; O1, NotebookLM, Tesla FSD, and Waymo are highlighted as wow moments. Agentic AI as the dominant 2025 theme: The hosts predict autonomous, tool-using AI systems will dominate conferences, startups, and product development, with major security and permissions challenges. AI in everyday devices: They expect AI to spread into glasses, phones, computers, and consumer hardware, though some integrations may feel gimmicky or intrusive. AI-driven scientific discovery: AI is expected to accelerate research, especially in domains where models can run experiments or analyze broad cross-domain knowledge faster than humans. Enterprise AI monetization and infrastructure: Massive GPU, energy, and nuclear investments will force companies to prove ROI; profitable AI deployments and security remain hard. Career shift toward AI engineering: Demand is expected to grow faster for AI engineering skills than for traditional data science skills, but as an evolution of the role rather than replacement.

Key Arguments: GPU demand remained strong in 2024 and diversified beyond NVIDIA, validating the prediction that more hardware players would emerge. Slow-thinking/inference-scaling models like OpenAI o1 represent a real shift beyond brute-force scaling, especially for tasks that can be broken into verifiable steps. LLM tool use evolved into agentic AI frameworks, but true tool consolidation is still incomplete because specialized products remain strong. Workplace upheaval is already visible in software and tech hiring, with AI contributing to layoffs and slower postings, especially for entry-level roles. Agentic AI will become the central 2025 narrative, but widespread adoption depends on trust, accuracy, permissions, and security across platforms. AI will increasingly be embedded into consumer devices and wearables, but some product placements may be superficial or unnecessary. AI-assisted science could materially raise discovery rates, patent output, and prototyping speed, especially when experiments can be run in silico or in automated labs. Enterprises will need to justify large GPU and infrastructure spending with measurable revenue or cost savings, otherwise projects will remain stuck in proof-of-concept purgatory. Data science careers are not disappearing, but the market will increasingly value AI engineering abilities such as API orchestration, prototyping, and fine-tuning. Open-source AI and cheaper models benefit consumers and startups by lowering access barriers and expanding competition.

Data Points: Podcast episode: 849 - This annual trend prediction special for 2025. Sadie’s predictions reviewed: 5 - Five predictions were made for 2024 and then evaluated. Women in Data community: 60,000+ women across 55 countries - Sadie’s broader professional work. Students reached by Sadie’s content: 600,000+ - Her instructional content audience. NVIDIA employee millionaire share: 75% - Claim mentioned during discussion of stock appreciation. Software development job postings: Sharp decline since 2022 - Used as evidence of workplace upheaval; peak coincided with ChatGPT’s release. OpenAI o1: Inference scaling / slow-thinking model - Cited as exemplifying the move beyond pure training scaling. AI scientist impact on materials discovery: 44% more new material discoveries - MIT study referenced in the scientific discovery discussion. AI scientist impact on patents: 39% more patents - MIT study referenced in the scientific discovery discussion. AI scientist impact on prototypes: 70% increase in new product prototypes - MIT study referenced in the scientific discovery discussion. Meta GPU target: Nearly 600,000 NVIDIA H100 GPUs by end of 2024 - Used to illustrate infrastructure investment scale. XAI cluster: 100,000-GPU AI cluster - Discussed as the largest single cluster mentioned in the episode. NeurIPS conference year: 38th year - Used to underscore the prominence of agentic AI interest in academia. NotebookLM episode reference: Episode 822 - John referenced a prior show featuring a NotebookLM-generated podcast about his PhD dissertation. AI engineer episode reference: Episode 847 - Mentioned as a resource for listeners wanting to become AI engineers.

Pivotal Quotes: "We are building the single best driver." — Waymo co-CEO (as relayed by Sadie): Used to explain the idea of specialized AI “brains” for one task that can be scaled widely. "I wanted automation to be taking away the mundane tasks from me ... but instead, AI's taken that away, and I'm left folding the laundry." — Referenced viral post discussed by John and Sadie: Illustrates the tension behind generative AI and the rise of agentic/physical automation. "The single best driver brain" — Sadie Saint Lawrence: Her framing of how focused machine intelligence can transform an industry at scale.

Implications: 2025 is likely to be defined by agentic systems, embedded AI, and ROI pressure. Listeners should build AI engineering fluency, think about security and monetization early, and expect faster career and product shifts across the data/AI stack.