A series of informal conversations with thought leaders, researchers, practitioners, and writers on a wide range of topics in technology, science, and of course big data, data science, artificial intelligence, and related applications. Anchored by Ben Lorica (@BigData), the Data Exchange also features a roundup of the most important stories from the worlds of data, machine learning and AI. Detailed show notes for each episode can be found on https://thedataexchange.media/ The Data Exchange podcast is a production of Gradient Flow [https://gradientflow.com/].
Thomson Reuters has developed a domain-specific frontier AI model, "Thomson," to enhance its legal, tax, and compliance software offerings. By utilizing open-weight checkpoints like Qwen 3.5, the company avoids the limitations of basic fine-tuning APIs, instead employing a rigorous pipeline of continuous pre-training a...
AI integration in gaming requires moving beyond cost-reduction gimmicks toward foundational mechanics that enable new player experiences. While the industry has historically adopted new technologies quickly, current AI applications often prioritize automated asset production, which risks cheapening the artistic quality...
AI development is shifting toward post-transformer architectures that prioritize latent reasoning and continual learning over massive, static parameter scaling. Zuzanna Stamirovska, CEO of Pathway, argues that current transformer-based models are inefficient for reasoning and prone to catastrophic forgetting. By utiliz...
OpenAI’s new "Jalapeno" chip aims to reduce inference costs and provide greater control over the AI infrastructure stack. By prioritizing high affinity between memory and compute, the architecture achieves both high throughput and low latency, allowing for flexible deployment across varying model requirements. The deve...
The rapid advancement of AI technology is fundamentally reshaping knowledge work, shifting the primary challenge from content generation to validation and verification. While AI tools enable high-speed output, human expertise remains critical for peer-reviewed validation and effective communication. This shift mirrors ...
Spatial data, specifically stereo imagery, serves as a critical foundation for training advanced AI world models by providing accurate geometric representations that monocular video lacks. While current foundation models rely heavily on flat 2D data, integrating stereo information allows AI to better understand depth, ...
LLMOps represents a fundamental shift from traditional MLOps, moving beyond simple containerization to managing complex, distributed inference systems that require behavioral performance monitoring and rigorous cost control. As generative models scale, the focus has transitioned toward building reliable agentic workflo...
Visual reasoning remains a significant hurdle for current frontier models, which often fail at basic tasks like counting objects or navigating spatial constraints despite their massive scale. Andrew Dai, CEO of Elorian AI, argues that reasoning originates from the visual world rather than language, necessitating specia...
Topic modeling remains a robust tool for structuring large datasets, even as large language models (LLMs) dominate the AI landscape. Developers must master foundational concepts—specifically tokens, embeddings, and attention mechanisms—to build reliable, high-performance applications rather than relying solely on promp...
The AI compute market remains opaque and supply-constrained, characterized by rising rental costs for both legacy A100 and modern H100/B200 GPUs. Silicon Data brings transparency to this sector by tracking GPU rental indices, token expenditure, and forward curves, revealing that demand consistently outpaces supply. Whi...
AI agents causing security incidents highlight critical failures in governance, identity management, and network access control rather than inherent model danger. Enterprises must prioritize rigorous security guardrails over mere scaling, especially as autonomous agents move toward complex, non-human workflows. Meanwhi...
Video search and understanding require moving beyond simple metadata or transcript-based approaches to natively process the rich, spatial-temporal information inherent in video. Twelve Labs addresses this by utilizing specialized foundation models: Marengo, which creates vector embeddings for efficient search, and Pega...
Effective AI risk evaluation requires a holistic, high-dimensional approach rather than relying on simplistic, low-dimensional tests. Current industry practices often fail because they utilize single-model "LLM as a judge" frameworks, which produce inaccurate results and hinder deployment speed. Organizations must impl...
Time series foundation models are transforming predictive analytics by moving beyond classical statistical approaches to leverage massive, self-supervised datasets. Ameet Talwalkar, Professor at Carnegie Mellon and Chief Scientist at Datadog, explains that these models have reached a "BERT moment," where scaling parame...
Reliability in AI-driven systems requires a holistic approach that integrates monitoring across infrastructure, models, and data. As AI agents move from demo to production, they face significant risks from data drift, hallucinations, and security vulnerabilities, particularly in regulated industries. InsightFinder AI a...
Follow this podcast in Podwise
Sign in to get AI summaries, transcripts and mind maps for any episode, including new ones.
