Latest Interviews
Showing 1–15 of 18 interview transcripts.
Clear all filters- Stanford Online1h 1m
Stanford CS153 Frontier Systems | Andreas Blattmann from Black Forest Labs on Visual Intelligence
Andreas Blattmann, Anjney Midha
Black Forest Labs, a Freiburg-based team of former Stability AI researchers, has scaled a 25-person operation to a $3 billion valuation by bootstrapping the Flux family of multimodal generative models. The company distinguishes itself through an open-weight commercial strategy and a strict adherence to EU AI Act compliance, maintaining identical safety guardrails for all partners including Meta and XAI. Looking forward, the organization is shifting its research focus from image synthesis to physical AI and robotics, aiming to validate model intelligence through real-world causal interactions rather than subjective aesthetic metrics.
- RAISE Summit29 min
Anjney Midha of a16z: Algorithmic Independence The Next Frontier in National Infrastructure
The event defines Sovereign AI as a four-tiered strategy for achieving technical and cultural independence, utilizing methods like mid-training to correct deep-seated biases rather than simple fine-tuning. As nations transition from closed-source prototypes to open-source models, a geopolitical divide is emerging where "hypercenters" like the US and China face competition from Europe's "Mistral Compute" initiative while non-hypercent nations adopt joint-venture models to retain value. The discussion concludes with predictions that general-purpose robotics and universal computer use will mature within two years, leveraging Europe's specific strengths in computer vision to secure its strategic position in the global AI landscape.
- a16z52 min
Building an AI Physicist: ChatGPT Co-Creator’s Next Venture
Anjney Midha, Liam Fedus, Ekin Dogus Cubuk
Founded by co-creators of ChatGPT and DeepMind physicists, Periodic Labs operates a frontier AI research facility dedicated to advancing physical science by coupling Large Language Models with automated high-throughput experimentation. The organization deploys a unique "mid-training" methodology and physically grounded reward functions to overcome the epistemic limits of current models, specifically targeting the discovery of high-temperature superconductors exceeding 200 Kelvin. By integrating ML scientists, experimentalists, and simulators into a unified workflow, the company aims to replace theoretical predictions with real-world data loops, eventually expanding its "AI physicist" capabilities to serve aerospace, defense, and semiconductor industries.
- a16z53 min
From Vibe Coding to Vibe Researching: OpenAI’s Mark Chen and Jakub Pachocki
Mark Chen, Jakub Pachocki, Anjney Midha, Sarah Wang
OpenAI researchers Mark and Jakob Sutskever outline a strategic roadmap centered on GPT-5, which aims to mainstream advanced reasoning and automate scientific discovery by merging the capabilities of instant-response and deep-thought models. This approach shifts evaluation metrics from solving static competition problems to generating economically relevant insights and extending autonomous time horizons to several hours through reinforcement learning. The organization distinguishes itself by balancing protected fundamental research teams with product accountability, prioritizing talent that persists through failure to overcome current limitations in coding autonomy and physical robotics.
- a16z42 min
Google DeepMind Lead Researchers on Genie 3 & the Future of World-Building
Jack Parker-Holder, Shlomi Fruchter, Anjney Midha, Marco Mascorro, Justine Moore, Erik Torenberg
Google DeepMind has released Genie 3, a research preview that generates interactive, photorealistic 3D worlds in real-time from text prompts to support navigation and control. Built by integrating insights from three internal projects, the model introduces spatial memory for one-minute object persistence and emergent physical reasoning to distinguish it from previous video generation systems. While currently limited to visual simulation without audio, Genie 3 aims to bridge the sim-to-real gap for robotics and agent training by providing diverse, high-fidelity environments free from physical data collection risks.
- a16z42 min
The Current Reality of American AI Policy: From ‘Pause AI’ to ‘Build’
Martin Casado, Anjney Midha, Erik Torenberg
Driven by the rapid rise of open-source models from competitors like DeepSeek, US policy has pivoted from existential risk narratives to a 2024 Innovation Action Plan co-authored by technologists to prioritize scientific discovery over restrictive liability frameworks. This new strategy replaces theoretical safety concerns with an empirical evaluation ecosystem and predicts a market split where open weights serve sovereign entities while closed-source models power frontier applications. By rejecting historical precedents of technology lock-downs, the plan aims to maintain global leadership through open collaboration and rapid iteration despite acknowledging a lack of direct academic funding.
- a16z1h 45m
Beyond Leaderboards: LMArena’s Mission to Make AI Reliable
Anjney Midha, Anastasios N. Angelopoulos, Wei-Lin Chiang, Ion Stoica
LM Arena has transformed from a static benchmark into a dynamic "humanity's exam" that evaluates over 280 AI models through real-time feedback from one million monthly users, effectively eliminating data contamination through fresh prompt generation. By treating evaluation as Reinforcement Learning rather than Supervised Learning, the platform utilizes techniques like "style control" and the open-sourced "Prompt-to-Leaderboard" router to achieve twice the performance-per-cost while maintaining academic neutrality. Looking forward, the organization plans to expand into private industry-specific Arenas and multi-modal agent testing while remaining committed to open-sourcing all data and research to preserve ecosystem trust.
- a16z1h 18m
Rick Rubin: Vibe Coding is the Punk Rock of Software
Rick Rubin, Marc Andreessen, Ben Horowitz, Anjney Midha, Erik Torenberg
Rick Rubin introduces "vibe coding" as a methodology merging the ancient spiritual principles of the Tao Te Ching with modern AI to democratize creation for non-technical users. The discussion outlines how this approach treats AI as a tool for human expression rather than an autonomous creator, aiming to counteract the homogenization of global culture and narrow demographic biases in current tech development. Ultimately, the event advocates for a future of education focused on cultivating taste and self-knowledge, allowing artists to leverage AI to raise creative ceilings while maintaining authentic individual agency.
- a16z16 min
Sovereign AI: Why Nations Are Building Their Own Models
Anjney Midha, Guido Appenzeller
Saudi Arabia has announced the construction of a $100 billion to $250 billion local hyperscaler named "Humane" to establish sovereign AI infrastructure capable of running 500-megawatt clusters that prioritize national control over cultural and informational output. This strategic pivot distinguishes itself from traditional cloud computing by treating AI as a critical cultural asset, requiring nations to build independent "AI Factories" to prevent foreign entities from dictating model values and societal narratives. The resulting geopolitical landscape favors a competitive market ecosystem where nations secure their own inference capabilities, potentially avoiding total centralization while mitigating risks associated with reliance on foreign foundation models.
- a16z1h 36m
Building the Next Generation of Conversational AI
Ankit Kumar, Anjney Midha, Maya
Sesame is developing a voice-first "companion" interface using a talent-dense team of fewer than fifteen engineers to prioritize natural conversational dynamics over general-purpose utility. The company has open-sourced its Conversational Speech Model base weights while withholding character-specific implementations, aiming to evolve toward a full duplex architecture capable of native audio understanding and real-time interruption handling. By targeting smart glasses as the optimal hardware form factor and employing qualitative human evaluation rather than standard metrics, Sesame seeks to build a long-term memory layer that acts as an emotionally resonant mediator for multi-step tasks.
- a16z18 min
AI Is Becoming a Regional Race
Modern AI is classified as a General Purpose Technology diffusing faster than previous innovations, forcing nation-states to prioritize the strategic choice of building or buying compute infrastructure over the next 24 months. The resulting global landscape is bifurcating into "hypercenters" that own the full stack and "compute deserts," compelling smaller nations to form value-aligned joint ventures with major powers rather than attempting infeasible total vertical sovereignty. True national autonomy now depends on securing critical components like energy and data while navigating divergent regulatory regimes that determine whether a country becomes a leader or a dependent in the new AI economy.
- a16z1h 17m
The Quest for Community-Trained Open Source AI Models
Bowen Peng, Jeffrey Quesnelle, Anjney Midha
News Research has unveiled the Distro method, a decentralized training framework that enables the creation of state-of-the-art "Hermes" language models using only standard internet connections and consumer-grade hardware. This breakthrough achieves an estimated 857-fold reduction in bandwidth requirements by allowing individual nodes to train independently and exchange only high-value insights rather than full model weights. By demonstrating that global AI development can be replicated without reliance on high-end data centers or single-entity resources, the project aims to democratize access to foundational models while preserving the neutrality and open nature of the technology.
- a16z30 min
Luma's Dream Machine and Reasoning in Video Models
Luma released Dream Machine, a foundational video generative model that leverages massive 2D data scaling to achieve robust text-to-video and image-to-video synthesis with emergent 3D structural reasoning. The model implicitly simulates complex physical phenomena, such as depth perception, light transport, and causal character interactions, enabling consistent scene reconstruction from single inputs without native 3D priors. While currently classified as a research preview, the development roadmap aims to evolve the technology into a 4D spatiotemporal simulator and multimodal agent capable of handling intricate narrative and interactive requirements.
- a16z46 min
How Discord Became a Developer Platform
Jason Citron, Anjney Midha, Mark Mandelmann, David Malani
Discord has grown to serve over 200 million monthly active users, leveraging a recent shift in developer activity that generated more than 20,000 new activities via its Embeddable Apps SDK. This platform evolution, driven by CEO Jason Citron's strategy to prioritize community feedback and open architecture, now allows startups to deploy rich HTML5 applications directly within the ecosystem while utilizing new one-click payment features for monetization. As a result, development friction has decreased significantly, enabling a projected surge from 20,000 to 200,000 apps within a single year as the platform transitions from a gaming chat tool into a comprehensive hub for the generative AI and interactive metaverse.
- a16z39 min
Text to Video: The Next Leap in AI Generation
Anjney Midha, Andreas Blattmann, Robin Rombach
Released on November 21st, Stable Video Diffusion is a state-of-the-art open-source model that generates short video clips from single input images by prioritizing the learning of complex physical properties like 3D consistency and camera movement. The architecture employs diffusion methodology over autoregressive methods to optimize perceptual details and utilizes LoRA adapters for scalable control of camera motion, while training strategies focused on temporal dynamics and specific 3D orbit refinement to achieve surprising reasoning capabilities in as few as 2,000 iterations. This release continues the team's philosophy of driving innovation through algorithmic efficiency rather than sheer compute volume, having already catalyzed a rapid ecosystem of community experimentation and set a roadmap for longer sequences and future audio integration.