The podcast by and for AI Engineers! In 2025, over 10 million readers and listeners came to Latent Space to hear about news, papers and interviews in Software 3.0. We cover Foundation Models changing every domain in Code Generation, Multimodality, AI Agents, GPU Infra and more, directly from the founders, builders, and thinkers involved in pushing the cutting edge. Striving to give you both the definitive take on the Current Thing down to the first introduction to the tech you'll be using in the next 3 months! We break news and exclusive interviews from OpenAI, Anthropic, Gemini, Meta (Soumith Chintala), Sierra (Bret Taylor), tiny (George Hotz), Databricks/MosaicML (Jon Frankle), Modular (Chris Lattner), Answer.ai (Jeremy Howard), et al. Full show notes always on https://latent.space Sponsorship and business inquiries: [email protected] www.latent.space
Modal provides critical AI infrastructure, offering compute and storage platforms that enable developers to build and scale AI applications efficiently. The platform’s growth is anchored in its ability to manage complex GPU capacity, shifting from reliance on spot instances to securing multi-year commitments across twe...
Moshi founder Anekha Sokhal leverages her background in mathematics and energy derivatives trading to build AI-driven software that provides real-time, actionable insights for financial professionals. By processing vast quantities of structured and unstructured data—ranging from market time series to alternative source...
Tenex functions as an AI-native partner for large enterprises, transitioning organizations from simple application usage to fully integrated AI workflows. The company utilizes a performance-based compensation model, rewarding engineers for output measured in story points rather than hours, which has driven a roughly 2x...
The compute market requires physical settlement to achieve true liquidity and standardization, addressing the inherent heterogeneity of hardware and the complexities of cluster management. By treating compute as a commodity with a forward curve, market participants can hedge risks and optimize utilization across divers...
Small molecule drug discovery faces significant hurdles in data quality and scalability, prompting Leash Bio to leverage DNA Encoded Libraries (DELs) for high-throughput measurement. By attaching DNA barcodes to millions of small molecules, researchers can physically screen vast chemical libraries against hundreds of p...
World-action models represent a paradigm shift in robotics by integrating future state prediction with physical action generation. Dyna-2 utilizes a diffusion transformer architecture trained on over one million hours of data to enable precise, dexterous motion in real-world environments. This approach emphasizes the i...
The "bitter lesson"—the principle that scalable methods eventually outperform handcrafted ones—requires a critical re-evaluation when applied to data-driven bioscience. While scaling compute and data is essential, success depends on identifying the right problem and prioritizing data quality over mere quantity. Modelin...
Accelerating materials discovery requires moving beyond theoretical simulation to a "synthesis superintelligence" that integrates AI with high-throughput physical experimentation. Because materials science involves complex, noisy environments—where variables like furnace degradation or mechanical vibrations introduce u...
Recursive Language Models (RLMs) and advanced agent harness design represent a fundamental shift in AI development, moving beyond standard autoregressive decoders toward compositional systems that leverage code-based sub-agent calls. Alex Zhang, a researcher and contributor to GPU Mode, emphasizes that current AI progr...
OpenAI’s latest developments in computer use agents and API infrastructure mark a shift toward highly efficient, autonomous software interaction. By providing agents with dedicated Linux virtual environments, these tools now execute complex, multi-step tasks—such as meal planning or YouTube management—faster than human...
Agentic coding has evolved from a niche experiment to a standard engineering workflow, driven by the rapid development of Claude Code and its associated harnesses. Effective interaction with these agents requires building a precise mental model of their capabilities, particularly in handling ambiguity and verifying com...
OpenRouter functions as a critical infrastructure layer for the AI ecosystem, bridging the gap between model labs and developers by providing a neutral, scalable marketplace for LLM inference. By treating models as consumable SKUs, the platform enables developers to navigate the rapidly evolving landscape of frontier m...
Runway co-founder Anastasis Germanidis details the evolution of generative video technology, tracing the company's path from early creative tools to the development of general-purpose world models. The discussion highlights the shift from simple content generation to simulating physics and human dynamics, emphasizing t...
Generative genomics leverages large language models trained on DNA to read and write the fundamental fabric of life, enabling breakthroughs in scientific discovery and human health. Unlike traditional bioinformatics, models like Omni utilize unsupervised pre-training to predict regulatory functions and identify causal ...
AI-driven scientific discovery centers on the AERA framework, which transforms complex scientific problems into "scorable tasks" optimized by AI agents. By leveraging large language models like Gemini, researchers can iteratively generate and refine code to maximize specific scientific objectives, effectively offloadin...
Jev, a newly launched "System One" model, represents a shift toward machine-native AI designed for programmatic consumption rather than human-facing chat. By prioritizing intelligence-per-dollar and reliability, Jev functions as a foundational infrastructure layer for software developers, moving away from the mode-drop...
Risk, rather than technical capability, remains the primary bottleneck for widespread AI adoption in critical sectors like banking and healthcare. AIUC addresses this by building "confidence infrastructure" through a standardized, quarterly-updated certification framework for AI agents. This process combines rigorous t...
The "Eureka Machine" represents a vision for superintelligence designed to automate scientific and technological invention across fields like physics, chemistry, and biology. Rather than fearing a rapid "hard takeoff," the development of such systems faces significant physical and economic constraints, necessitating a ...
Integrating AI with physical world modeling through neural operators enables high-fidelity simulations that outperform traditional numerical methods. Neural operators generalize neural networks by learning mappings between function spaces, allowing for multi-scale resolution and efficient modeling of complex physical p...
Simulation technology serves as a powerful tool for human decision-making by modeling complex social behaviors rather than merely predicting future outcomes. Joon Sung Park, co-founder of Simile, notes that current frontier models often lack the deep, behavioral "dark knowledge" of humanity, which requires specialized ...
Follow this podcast in Podwise
Sign in to get AI summaries, transcripts and mind maps for any episode, including new ones.
