Viewing Podcast: Podcast
AI
Arts
Business
Crypto
Finance
Health
History
Interviews
Investing
Macro
Misc
News
Politics
Product
Programming
Science
Social
Startups
Technology
VC
Autoformalization and Verifiable Superintelligence with Christian Szegedy - #745
Duration: 01:11:48
September 2, 2025
- Christian believes AI will soon achieve "mathematical super intelligence," exceeding human capabilities in specific domains with demonstrably testable results.
- A key focus of Christian's current research is auto formalization, converting mathematical knowledge into formal languages that AI can verify and build upon.
- Christian emphasizes the importance of formal verification in AI development to ensure outputs are guaranteed to meet specifications, thereby mitigating potential risks of AI subversion or unintended behaviors, especially about AI safety.
Genie 3: A New Frontier for World Models with Jack Parker-Holder and Shlomi Fruchter - #743
Duration: 01:01:01
August 19, 2025
- Genie 3 represents a significant leap forward in world model technology, achieving approximately a 100x improvement by pushing the limits across dimensions like generation quality, resolution, interaction duration, and frame generation speed.
- The project evolved from Genie 1 and 2 by prioritizing real-time interaction, enabling the model to respond to user actions in real time, a key design decision that influenced the entire system architecture.
- A key limitation of Genie 3 is the lack of complex multi-agent interactions within generated worlds, although future development may focus on teaching agents to interact in visually realistic, embodied worlds with humans.
Closing the Loop Between AI Training and Inference with Lin Qiao - #742
Duration: 01:01:11
August 12, 2025
- The "fast iteration experimentation loop" needs to combine both training and inference, as product AB testing is the ultimate judge of whether a model investment is successful.
- A key lesson learned from PyTorch is that a cohesive system is needed to ensure training and inference alignment for quick cross-deployment, and the inference system for experimentation should be the same as that for large-scale production.
- The most exciting trend is making AI model customization accessible to all developers, so they can leverage their production data in a closed loop to gain a competitive edge.
Context Engineering for Productive AI Agents with Filip Kozera - #741
Duration: 00:46:01
July 29, 2025
- Wordware is simplifying agent creation by allowing users to define agent tasks using natural language documents, which are then executed by React agents.
- The discussion highlighted the importance of incorporating human feedback into agent reflection loops to handle situations where the agent lacks knowledge or requires creativity.
- A key challenge is balancing data access and privacy, specifically concerning how AI agents can leverage user data in silos like Slack and Notion without compromising data ownership.
Infrastructure Scaling and Compound AI Systems with Jared Quincy Davis - #740
Duration: 01:13:02
July 22, 2025
- Jared Quincy Davis discusses compound AI systems, highlighting their potential to improve efficiency and reduce costs by strategically combining different AI models.
- He introduces the concept of "networks of networks," architectures that compose multiple AI models, enabling significant performance gains on verifiable tasks like code generation and math.
- Foundry is developing infrastructure to support these compound AI systems, focusing on intelligent scheduling and resource allocation to optimize compute utilization and reduce inference costs.
Building Voice AI Agents That Don’t Suck with Kwindla Kramer - #739
Duration: 01:13:03
July 15, 2025
- Enterprise voice AI is demonstrating flexible LLM conversation capabilities, unlike current consumer demos which are demos, not products due to structural constraints.
- The stack for building voice AI applications includes models, APIs, orchestration, and application code, with Pipcat offering a vendor-neutral, open-source orchestration layer.
- Challenges in voice AI development involve managing low latency for UDP networking, handling cold starts, and addressing the multi-model nature of most production voice applications.
Distilling Transformers and Diffusion Models for Robust Edge Use Cases with Fatih Porikli - #738
Duration: 01:00:29
July 9, 2025
- DIMA, Qualcomm's new state-of-the-art AI system for autonomous driving, achieves significant reductions in collision rates and trajectory error by using an end-to-end approach.
- LLMs can enhance autonomous driving by leveraging their "world knowledge," acting as a regularizer to improve generalization and handle long-tail scenarios more effectively.
- Sharp Depth bridges the gap between generative and discriminative approaches to monocular depth estimation, producing detailed and accurate depth maps by intelligently fusing the strengths of both methodologies.
Building the Internet of Agents with Vijoy Pandey - #737
Duration: 00:56:13
June 24, 2025
- The "Internet of Agents" is envisioned as a platform for agent-to-agent collaboration, enabling subject matter expert agents to seamlessly work together to solve complex business problems.
- A key challenge lies in transitioning from deterministic computing to probabilistic agentic systems, balancing the agility and potential of AI agents with the need for deterministic APIs and human oversight.
- Cisco's open-source project "agency" addresses the communication and transport layers for agent collaboration, introducing "slim" as a secure, real-time messaging layer that enhances existing protocols like A2A and MCP.
LLMs for Equities Feature Forecasting at Two Sigma with Ben Wellington - #736
Duration: 00:59:31
June 17, 2025
- The field of NLP has shifted from a syntax-focused approach to an empirical, data-driven approach, which has dramatically changed how language is processed and understood by machines.
- Feature forecasting at Two Sigma involves identifying and quantifying interesting, observable facts about companies and other entities to predict their future performance, exemplified by using satellite imagery.
- Large Language Models (LLMs) have revolutionized feature creation by significantly reducing the time and resources required to test hypotheses and develop new predictive signals, potentially enabling a "renaissance of feature creation."
Zero-Shot Auto-Labeling: The End of Annotation for Computer Vision with Jason Corso - #735
Duration: 00:56:15
June 10, 2025
- The traditional approach to data annotation (Annotation 1.0) of just blindly sending all data out for labeling is shifting towards a more efficient, question-driven method (Annotation 2.0), prompted by the emergence of powerful, semantically-rich foundation models.
- Voxil 51 is designed to be a Rosetta Stone for visual data formats that facilitates analysis via interactive exploration of embeddings to clean data, find corner cases, and integrate with existing annotation workflows.
- Recent research indicates that "zeroshot" autolabeling with foundation models can rival human performance in object detection while being significantly cheaper and faster, even if the autolabels contain some noise.