The podcast & newsletter where 170,000+ AI Engineers gather to talk models, tools and ideas. Breaking news today you will use at work tomorrow! Full show notes and newsletter at https://latent.space
Dylan Patel discusses AI geopolitics and semiconductor strategy while cooking fried rice, explaining why the U.S.-China tech war creates complex tradeoffs between talent retention and national security in the race for AI dominance.
Max Welling traces his career from quantum gravity to founding CuspAI, explaining how physics underpins modern AI architectures and why the field of AI for science is attracting billions in investment to solve foundational materials problems—from carbon capture to next-generation GPUs—by treating nature as a computational "Physics Processing Unit."
Doug O'Laughlin describes Claude Code as a capable but imperfect 'junior analyst' that amplifies expert productivity rather than replacing human judgment, while recounting his journey from value investing to semiconductor analysis through his early conviction that Moore's Law was ending.
AI has created a unique capital flywheel where frontier model companies raise billions pre-monetization to buy compute, convert dollars directly into capabilities via scaling laws, and vertically integrate into applications—potentially outspending and absorbing the entire ecosystem built on top of them.
Jeff Dean explains Google's strategy of 'owning the Pareto frontier' by developing both frontier-capable AI models (Pro/Ultra) and highly efficient variants (Flash) through distillation, enabling massive-scale deployment across Google's products while pushing boundaries in long context and multimodality.
AlphaFold2's 2020 breakthrough solved single-chain protein structure prediction using evolutionary sequence correlations, but the field still struggles with folding dynamics, protein complexes, and molecules lacking evolutionary data, while expanding to small molecules and nucleic acids.
Pratyush Maini from Datology demonstrates how the 'seahorse emoji' query acts as a diagnostic probe to reverse-engineer when frontier labs began injecting reasoning traces into mid-training data, revealing that self-correction capabilities have shifted from post-training additions to core foundation model ingredients.
Goodfire AI, now valued at $1.25 billion after a $150 million Series B, is pioneering the use of mechanistic interpretability to move beyond analyzing AI models to actually designing them—enabling surgical edits to model behavior for production applications ranging from bias removal to PII detection.
Context graphs represent the emerging institutional memory layer for AI agents, capturing 'decision traces'—the reasoning behind human and agent actions—to solve enterprise reliability gaps and create durable competitive moats as foundation models commoditize raw AI capabilities.
Andrew White recounts his journey from molecular dynamics researcher to founder of Future House and Edison Scientific, revealing how AlphaFold disrupted specialized supercomputing approaches and why current LLMs have already compressed a decade-long vision for automated science into just two years—prompting urgent policy attention at the highest levels of government.
OpenAI launches Prism, a free AI-native LaTeX editor powered by GPT-5.2 that eliminates context-switching for scientists by embedding reasoning capabilities directly into academic writing workflows, from converting hand-drawn diagrams to verifying complex equations.
Yi Tay returns to Google DeepMind Singapore to lead the Reasoning and AGI team, explaining the shift toward on-policy reinforcement learning as the dominant paradigm for model reasoning and sharing the technical story behind Gemini's IMO Gold achievement using an end-to-end text approach.