ai
Making artificial intelligence practical, productive & accessible to everyone. Practical AI is a show in which technology professionals, business people, students, enthusiasts, and expert guests engage in lively discussions about Artificial Intelligence and related topics (Machine Learning, Deep Learning, Neural Networks, GANs, MLOps, AIOps, LLMs & more). The focus is on productive implementations and real-world scenarios that are accessible to everyone. If you want to keep up with the latest advances in AI, while keeping one foot in the real world, then this is the show for you!
Stanford HealthCare developed a Surgical Co-Management Navigator tool using the O3 Mini model to screen next-day elective surgeries and identify patients needing surgical co-management. According to Heidi Porter, the institution established its own strict rubric in the prompt so the model could only apply the criteria and output yes, maybe, or no without hallucinating.
Heidi Porter highlighted a peer-reviewed study of a rubber hose production line where a modified vision model detects defects without an LLM. An LLM is subsequently used only to draft candidate explanations and corrective actions based on retrieved maintenance history, accompanied by rule-based guardrails to ensure output stays within evidence.
Seed company Syngenta built an AI application where users ask questions in plain language regarding seed selection, and an LLM interprets requests before a Bedrock agent pulls weather, soil, and agronomic data. Heidi Porter explained that the LLM is restricted to routing requests and explaining results, while existing models handle actual agronomy predictions.
Kevin Solorio discussed his team's engagement at Focused building AI applications for blue-collar field technicians. He emphasized that adoption fails when tools are built through management proxies without direct input, advocating instead for building alongside workers and utilizing data flywheels and offline evaluation sets.
Chris Schenkel and Andrew Dunlap from SEP presented a sense-making framework for software development teams utilizing AI, categorized by agent autonomy from individual human-in-the-loop modes to full dark factory autonomy. They emphasized that token budgets, domain knowledge, and verifiability constrain which mode fits best.
Spencer McIntyre explains that while AI models like Mythos and Opus are proficient at identifying vulnerabilities in code, they often struggle to bridge the gap toward creating functional exploits that prove attacker utility.
McIntyre discusses VulnCheck's analysis of open-source proof-of-concept exploits, noting that 2026 marked a tipping point where the majority of circulating PoCs are broken or illegitimate due to an influx of AI-generated content.
McIntyre details that AI models are successful at exploiting open-source web application logic flaws, token logic issues, and file leaks, but struggle significantly with closed-source proprietary software and firmware requiring patch diffing.
McIntyre notes that offensive security researchers are experiencing increasingly strict guardrails on American proprietary AI models, leading many shops to downgrade or transition toward local open-weight models like GLM.
McIntyre highlights that AI tools are significantly accelerating vulnerability analysis and helping researchers react faster to new security flaws, thereby reducing the window of advantage that attackers historically held.
Mingyu Liu from NVIDIA explains the critical role of open models in advancing the AI field, enabling innovation, and providing developers with greater insight and customization options compared to API-based solutions. NVIDIA actively promotes open models as part of its strategy to empower the entire AI ecosystem.
Mingyu Liu highlights that open models offer significant commercial advantages by providing customers with the freedom to innovate beyond API limitations. This customization is crucial for businesses developing unique AI applications tailored to their specific needs.
NVIDIA sees 'Physical AI' as a key area encompassing automotive, factory automation, and robotics, where AI is deployed in devices that interact with and alter the physical world. The company envisions humanoid robots playing a significant role in assisting humans with various tasks.
Mingyu Liu explains that open models are crucial for physical AI due to the varied sensor setups and unique requirements of different robots and devices. Open models allow for customization and fine-tuning, enabling better adaptation to specific hardware and use cases, which is difficult with one-size-fits-all API solutions.
Mingyu Liu distinguishes 'World Models' from language models, explaining they are built from sensory observations to implicitly encode physics and enable simulation. NVIDIA's Cosmos lab develops these models for tasks like world understanding and simulation, creating a shared representation for physical AI applications.
Mingyu Liu differentiates between classical simulation, based on physics equations, and neural simulation, which is data-driven. Neural simulation, used in world models, approximates predictions by learning from vast amounts of observational data, offering a pattern recognition-based approach.
Mingyu Liu explains how world models significantly accelerate the development of AI for applications like self-driving cars. Instead of relying solely on real-world testing, policies can be rapidly iterated and verified within the simulated environment provided by a world model, allowing for faster development cycles.
Mingyu Liu envisions a future where complex tasks are handled by a multi-agent system of specialized AIs that collaborate. These agents, ranging from digital assistants to robots and drones, will have varying degrees of autonomy and will need to work together, emphasizing the importance of intelligence per dollar and specialized capabilities.
Mingyu Liu outlines NVIDIA's commitment to open models, which includes not just open weights but also training frameworks, open-sourced datasets, and developer resources like blogs and recipes. This comprehensive support aims to empower developers and foster innovation in the AI ecosystem.
The ML Ops community has been donated to the Linux Foundation, leading to the creation of the Agentech AI Foundation. This move aims to provide a neutral governing body for ML Ops and agents, attracting significant interest from top industry players.
AgentCon, the flagship event for the Agentech AI community, will feature a 36-hour livestream to integrate both in-person and virtual attendees. The event aims for an 'ESPN-style' live coverage of talks and expo hall exploration.
The Agentech AI Foundation is organizing several conferences, including an MLOP summit in Japan, AgentCon in Amsterdam on September 17th, and another AgentCon in San Jose on October 22nd. These events are part of the rebranding of the ML Ops community under the Agentech AI umbrella.
Demetrius Frankman, Head of Development Experience at the Agentech AI Foundation, discussed his new role following the ML Ops community's merger. He highlighted the foundation's aim to provide resources and foster a deep technical community around agents and ML Ops.
Chaiten Gupta, Chief AI Officer at Rackspace, shared his diverse career path, starting with a PhD in mathematics and progressing through research roles at Hewlett Packard Labs and Hitachi. He eventually led global AI research at Hitachi before moving to Rackspace.
Chaiten Gupta attributes his career advancement at Hitachi to the company's early and risky investment in industrial AI. This strategic focus on a nascent field, including establishing a research lab in the Bay Area, demonstrated foresight and confidence in internal talent to drive innovation.
Chaiten Gupta described the evolution of industrial AI, moving from early focus on problems like failure prediction and recommended repairs to incorporating computer vision for defect detection and quality control with the rise of deep learning. More recently, large language models (LLMs) have expanded possibilities in robotics and human-machine interaction within industrial settings.
Chaiten Gupta observed that Japan has a broader cultural acceptance of robots for daily interactions, including for elderly support, compared to North America. He noted that while the US may be behind in this area, China is currently leading in industrial robotics applications.
Daniel Weineck and Chris Benson discuss the definition of AI models, likening them to software functions that transform data. They explain how models like large language models (LLMs) and vision models process specific inputs to generate outputs, and note that not all models are neural network-based.
The discussion clarifies the distinction between open-weight/open-source AI models and closed models. Open models release both software and parameters (weights/biases), allowing local execution, while closed models restrict access to these components, typically accessed via APIs.
Chris Benson explains that AI models come in various sizes, with 'frontier models' being the largest and most sophisticated but requiring expensive hardware. Smaller models, like Gemma, can be downloaded and run locally on available hardware, with China emerging as a significant source for open models.
Daniel Weineck clarifies the concept of agentic harnesses, describing them as software systems that manage and orchestrate multiple AI agents. He likens these harnesses to an 'operating system for AI,' enabling agents to work together to accomplish tasks.
The podcast discusses a recent cybersecurity incident where OpenAI agents reportedly exploited Hugging Face, a popular repository for AI models and datasets. The incident has drawn comparisons to a murder mystery due to its developing nature and various facets involved, including the use of agentic AI, cybersecurity vulnerabilities, and potential geopolitical implications with US versus Chinese models.
CoreWeave is adopting an AI-centric approach to infrastructure, drawing parallels to the early days of cloud computing. Cory Sanders, Senior Vice President of Product at CoreWeave, explained that this involves building services with AI as a core component, rather than an add-on, to avoid hindering AI development.
Cory Sanders highlighted that CoreWeave focuses on optimizing AI training, a computationally intensive process requiring specialized infrastructure. The company identifies and addresses bottlenecks in this workflow, such as GPU failures or storage loading inefficiencies, offering unique services like an observability platform and a storage platform tailored for large-scale AI training workloads.
Cory Sanders discussed the challenge of adapting to new technological waves, referring to it as the 'innovator's dilemma.' He emphasized the importance of hiring people who can challenge existing assumptions and question the status quo. This mindset is crucial for evolving and building the next iteration of technology, especially in the rapidly changing AI landscape.
Hamza Tahir, co-founder of ZenML, discusses how core MLOps principles are being rediscovered and adapted for the emerging field of AI agents. He highlights that the fundamental concepts of writing reliable, safe, and retryable code are just as relevant for agents as they were for traditional machine learning pipelines.
Hamza Tahir explains that all software, including AI agents, can be viewed as a graph of execution steps. While traditional ML pipelines were more deterministic, agents involve dynamic, potentially cyclical workflows. ZenML's evolution to a dynamic mode addresses this by allowing graphs to be defined in real-time.
Hamza Tahir describes how AI "harnesses" or frameworks, like Claude Code, become deeply coupled with the specific large language models they are designed for. This coupling, driven by reinforcement learning and specialized tool use, means that models become 'self-aware' of the harness they are running within, leading to more accurate performance.
Hamza Tahir identifies a key industry tension between proprietary AI agent frameworks, like Claude Code, which are tightly coupled with specific models, and open frameworks like LangGraph or Pythontic AI. The latter aim for open standards, seeking to avoid vendor lock-in and reduce dependency on specific model providers.
Hamza Tahir points out that large language models are primarily designed for text processing, not direct real-world interaction. He notes that agents act as a bridge, but this requires significant scaffolding to handle issues like the LLM making incorrect tool calls or responding in unexpected formats, highlighting the need for robustness and error handling.