In this episode, you’ll explore the field of deep reinforcement learning and the ways it influences the real world with Anna Goldie, Senior Staff Research Scientist at Google DeepMind. 

Anna’s current role has her working on Large Language Model (LLM) research for Gemini and Bard. Prior to that, she worked on reinforcement learning for LLMs and retrieval-augmented LLMs at Anthropic, and was co-founder/lead of the ML for Systems team in Google Brain.

During this wide-ranging discussion, you’ll learn about her contributions to the field of reinforcement learning, and how we can leverage reinforcement learning effectively for real world applications going forward. 

Sponsored by: https://odsc.com/ 
Find more ODSC lightning interviews, webinars, live trainings, certifications, bootcamps here – https://aiplus.training/ 

Topics:
1. Professional journey and the key moments 
2. Core principles of deep reinforcement learning
3. Deep reinforcement learning for chip design vs  traditional approaches
4. Key complexities in modern chip design and how deep reinforcement learning can address these complexities
5. Discuss Google’s TPUs - Tensor Processing Units - built specifically for accelerating machine learning workloads. 
6. The potential of Deep Reinforcement learning in computer systems or other domains within Google Deepmind.
7. Deep reinforcement learning use in Large Language Models (LLMs)
8. Reinforcement Learning from Human Feedback (RLHF), designing effective rewards and providing feedback at scale
9. Scalable supervision techniques, for developing methods to efficiently gather feedback that aligns the LLM with human preferences
10. Implement the Constitutional AI framework where AI models are guided by a set of foundational principles or 'constitutional' directives
11. How Retrieval Augment Generation (RAG) systems improve the accuracy and relevance of responses compared to standard large language models even LLMs with large retrieval context windows
12. How “RAPTOR: RECURSIVE ABSTRACTIVE PROCESSING FOR TREE-ORGANIZED RETRIEVALcompares to traditional RAG approaches that retrieve short, contiguous chunks?
13. Hierarchical summaries with RAPTOR
14. LLM Finetuning With Low-Rank Adaptation (LoRA)
15. Google's Gemini 1.5 next-generation LLM model and mixture of experts architecture
16. CALM—Composition to Augment Language Models
17. How dual undergraduate degrees in both computer science and linguistics from MIT has contributed to your innovative work in machine learning,
18. Constitutional AI at Antropic https://www.anthropic.com/news/claudes-constitution
19. What is the best way to follow your work?
20. Keynote address at ODSC East in mid-April. 


SHOW NOTES
More about Anna Godie
https://www.linkedin.com/in/adgoldie/
https://www.annagoldie.com/


More about Constitutional AI at Antropic
https://www.anthropic.com/news/claudes-constitution
Constitutional AI: Harmlessness from AI Feedback
 https://arxiv.org/pdf/2212.08073.pdf


More about Anna’s Paper
RAPTOR: RECURSIVE ABSTRACTIVE PROCESSING FOR TREE-ORGANIZED RETRIEVAL
https://openreview.net/pdf?id=GN921JHCRw
The official Code implementation of RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval
https://github.com/parthsarthi03/raptor


More about Large Language Models
Direct Preference Optimization: Your Language Model is Secretly a Reward Model
https://arxiv.org/abs/2305.18290
LLM Finetuning With Low-Rank Adaptation (LoRA)
https://lightning.ai/pages/community/article/lora-llm


CALM—Composition to Augment Language Models
https://arxiv.org/pdf/2401.02412.pdf
https://www.anthropic.com/news/claudes-constitution

Deep Reinforcement Learning in the Real World with Anna Goldie

In this episode of the ODSC AiX Podcast, host Sheamus McGovern reconnects with Paige Bailey, Engineering Lead at Google DeepMind for the Developer Experience team. Paige shares how the Gemini ecosystem has evolved since her last appearance, including the launch of Gemini 2.5 DeepThink, multimodal video generation with Veo 3, real-time music creation with Lyria RT, and groundbreaking advances in agentic and on-device AI systems. The conversation explores the rapid rise of agent-based workflows, AI-powered robotics, and the growing divide between cutting-edge tools and real-world adoption.

Key Topics Covered:
Gemini 2.5 DeepThink & Reasoning Models
The model that won gold at the International Mathematical Olympiad (IMO)
Use cases for DeepThink, Pro, Flash, and FlashLite variants
Using Gemini Live API for real-world robotics and decision planning
Role of multimodal inputs (video, audio, text) in enabling embodied AI
On-Device AI & Ubiquity
Implications for edge deployment, cost reduction, and accessibility
Veo 3: Multimodal Video Generation
Lyria RT: Real-Time Music Generation
Gemini Live API & Voice Interfaces
Real-time bidirectional voice, screen understanding, and tool calling
Rise of voice as the dominant AI interface
Use of SynthID and digital watermarking to detect deepfakes
Future of AI-agent orchestration via MCP servers

Memorable Outtakes:
On the pace of model development: “A 4-billion parameter model on-device now outperforms our best cloud model from six months ago. That’s pretty magical.”  — Paige Bailey 
On the role of AI agents in robotics: “You can say, ‘Hey robot, go get me that apple,’ and Gemini will plan the task, route it, and call the right control models.” — Paige Bailey 
On the AI adoption gap: “In the Bay Area, we use AI hourly. But when I talk to developers in the Midwest, they often aren’t using it at all.” — Paige Bailey 

References & Resources:
Paige Bailey
Dynamic Web Paige: https://webpaige.dev/
LinkedIn: https://www.linkedin.com/in/dynamicwebpaige
GitHub: https://github.com/dynamicwebpaige
Medium: https://medium.com/@dynamicwebpaige
Previous podcast with Paige: https://podcasters.spotify.com/pod/show/ai-x-podcast/episodes/Googles-AI-Powered-Tools-for-Data-Scientists-Building-the-Automated-Future-of-Data-Science-with-Paige-Bailey-e2p3t6e

Resources mentioned
International Mathematical Olympiad (IMO): https://www.imo-official.org
Model Context Protocol (MCP): https://modelcontextprotocol.io/docs/getting-started/intro
Gemini 2.5 Deep Think: https://blog.google/products/gemini/gemini-2-5-deep-think/
Veo 3: https://deepmind.google/technologies/veo/
Lyria RT & Music AI Sandbox: https://deepmind.google/technologies/lyria/
SynthID & Deepfake Watermarking: https://deepmind.google/technologies/synthid/
Gemma Models: https://ai.google.dev/gemma
Gemini Live API Docs: https://ai.google.dev/gemini-api/docs/live
Google AI Studio: https://ai.google.dev

Sponsored by:
 🔥 ODSC AI West 2025 – The Leading AI Training Conference
Join us in San Francisco from October 28th–30th for expert-led sessions on generative AI, LLMOps, and AI-driven automation.
Use the code podcast for 10% off any ticket.
Learn more: https://odsc.ai




Inside Google’s New AI Stack with Paige Bailey

In this episode, Sheamus McGovern sits down with João “Joe” Moura, founder and CEO of CrewAI, to discuss the rapid evolution of multi-agent AI systems and their real-world impact in enterprise environments. From his early experiments with agent-based automations to supporting clients like PwC, PepsiCo, and the U.S. Department of Defense, Joe shares how CrewAI is enabling companies to move beyond hype and flashy demos toward scalable, secure agent workflows. This conversation covers everything from open-source momentum to agent orchestration, governance, deployment challenges, and what it takes to go from prototype to production.
Key Topics Covered:
* João’s background in AI and the early days of CrewAI
* The problem CrewAI was originally built to solve—and how it scaled from a personal tool to a global platform
* Definition of a true agent: “Agency is the key. If there’s no decision-making, it’s not an agent.”
* The evolution to robust multi-agent enterprise systems
* CrewAI’s dual product strategy: open source for developers, enterprise-grade platform for production use
* Use cases driving the most traction: go-to-market workflows, internal GPTs, code automation at scale
* The MIT report on AI adoption: why many enterprises are failing to see ROI—and how CrewAI is addressing the gap
* Empowering non-technical users through CrewAI Studio
* Security, governance, and the importance of deterministic control in agent deployment
* Why waiting for the next-gen LLM is a mistake: “The models we have today are good enough for 99% of enterprise use cases.”
* CrewAI’s take on interoperability, agent-to-agent protocols, and why standards matter
Memorable Outtakes:
* “If you want an agent, you’ve got to have agency. Otherwise, it’s just a script.”
* “LLMs today are good enough for 99% of enterprise use cases.”
*  “We’re now running over 475 million agent automations a month.”
References & Resources:
João “Joe” Moura – LinkedIn https://www.linkedin.com/in/joaomdmoura/CrewAI  https://www.crewai.com/
MIT Report on GenAI adoption (referenced in discussion) https://ide.mit.edu/research/95-of-generative-ai-projects-fail-how-to-make-yours-succeed
Humanity's Last Exam https://agi.safe.ai/
AgentBench⁠ https://github.com/THUDM/AgentBench⁠
HumanEval Benchmark⁠ https://github.com/openai/human-eval⁠
MCP – Model Coordination Protocolhttps://modelcontextprotocol.io/docs/getting-started/intro
Agent to Agent  Communication  https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/
Sponsored by: 🔥 ODSC AI West 2025 – The Leading AI Training Conference Join us in San Francisco from October 28th–30th for expert-led sessions on generative AI, LLMOps, and AI-driven automation. Use the code podcastfor 10% off any ticket. Learn more:⁠ https://odsc.ai⁠

CrewAI and the Rise of Autonomous Agents in Enterprise AI with João (Joe) Moura

In this episode of the ODSC AiX Podcast, host Sheamus McGovern sits down with Kwindla Hultman Kramer, co-founder of Daily and creator of the open-source project PipeCat. Kwindla is a pioneer in real-time audio, video, and human-computer interaction. He shares his journey from MIT Media Lab to building voice interfaces that are now shaping the future of software interaction.
Together, they explore why voice is poised to become the primary interface for AI, what makes voice agents so challenging to build, and how real-world use cases—from call centers to appointment scheduling—are proving the value of well-designed voice AI systems.
Key Topics Covered:
Why voice is at a “2007 iPhone moment” for user interfaces
What makes a great voice AI experience (and why Siri/Alexa missed the mark)
The rise of voice agents in production use: healthcare, banking, customer service, and more
Real-world engineering challenges in building voice-first AI agents
How latency, streaming, turn detection, and interruptibility define user experience
The architecture and motivation behind the PipeCat open-source toolkit
Why LLMs need context engineering in voice workflows
Differences between text-first vs. voice-first agent design
Hybrid models: routing between cloud and on-device inference
What’s missing in voice AI today: memory, context, and bi-directional streaming
Synthetic data and the future of training audio models
Advice for developers just starting with voice AI
Memorable Outtakes:
“Voice is at the tipping point—like mobile was in 2007.”
“All UI is going to become liquid, generated on the fly by natural voice speech.”
“You’ve probably already spoken to an AI agent—you just didn’t realize it.”
References & Resources:
PipeCat GitHub: https://github.com/pipecat-ai/pipecat
Kwindla Hultman Kramer: https://www.linkedin.com/in/kwkramer/
Daily (video/audio infrastructure company): https://www.daily.co
Voice AI and Voice Agents Guide: https://voiceaiandvoiceagents.com
Smart Turn Detection Model: https://github.com/pipecat-ai/smart-turn
Gemini 1.5 and Google Live API: https://blog.google/products/gemini/gemini-live-updates-august-2025/
NotebookLM by Google: https://notebooklm.google/
Suno AI (music generation): https://www.suno.ai/
Whisper (OpenAI ASR Model): https://github.com/openai/whisper
Minority_Report: https://en.wikipedia.org/wiki/Minority_Report_(film)
Sponsored by:
 🔥 ODSC AI West 2025 – The Leading AI Training Conference
Join us in San Francisco from October 28th–30th for expert-led sessions on generative AI, LLMOps, and AI-driven automation.
Use the code podcast for 10% off any ticket.
Learn more: https://odsc.ai



Voice AI is About to Get Loud with Kwindla Kramer

In this episode of the ODSC AI Podcast, we sit down with Paco Nathan, veteran AI practitioner, data science leader, and currently with Senzing. Paco shares how AI, entity resolution, and graph analytics are being deployed to uncover massive financial crime networks, from shell companies and money laundering to cybercrime and trafficking. Drawing from real-world cases like the Azerbaijani Laundromat and Danske Bank scandal, Paco explains how cutting-edge AI techniques, RAG, and open data are transforming the fight against hidden financial networks while also powering legitimate business intelligence.

Key Topics Covered:
Paco Nathan’s career journey: AI, neural networks, Databricks, and now Senzing
Senzing’s origins in casinos, card counting detection, and “non-obvious relationship awareness”
Industrial-scale money laundering: Azerbaijani Laundromat, Danske Bank Estonia branch, and $200B flows exceeding Estonia’s GDP
Entity resolution as the backbone of fraud detection and why accuracy is critical in high-stakes domains
How graph analytics expose hidden structures in money laundering networks
DSPy, BAML, and MLflow: AI tooling for programmatic prompt engineering and evaluation
The rise of Graph-RAG: combining retrieval-augmented generation with knowledge graphs
Forensic accounting, fraud tradecraft patterns, and synthetic data simulators
Open-source and open-data initiatives like OpenSanctions, OpenOwnership, and GLEIF
Opportunities for data professionals to engage in combating financial crime and supporting transparency

Memorable Outtakes:
Paco Nathan: “Defenders think in lists. Attackers think in graphs — and as long as that holds, attackers win.”
Paco Nathan: “Sometimes your best customers and your worst criminals look much the same.”
Paco Nathan: “One Estonian bank branch moved more money than the GDP of its host country.”

References & Resources:
Paco Nathan’s LinkedIn: https://www.linkedin.com/in/ceteri/
Senzing: https://senzing.com/
DSPy: https://github.com/stanfordnlp/dspy
BAML: https://github.com/BoundaryML/baml
MLflow: https://mlflow.org/
OpenSanctions: https://www.opensanctions.org/
OpenOwnership: https://www.openownership.org/
GLEIF (Global Legal Entity Identifier Foundation): https://www.gleif.org/en

Sponsored by:
🔥 ODSC AI West 2025 – The Leading AI Training Conference
Join us in San Francisco from October 28th–30th for expert-led sessions on generative AI, LLMOps, and AI-driven automation.
Use the code podcast for 10% off any ticket.
Learn more: https://odsc.ai


Follow the (Dark) Money: AI and the Hunt for Hidden Networks with Paco Nathan

In this episode of the ODSC Ai X Podcast, host Alex Landa speaks with Georgios N. Yannakakis, Professor at the Institute of Digital Games, University of Malta, co-founder of Modl.ai and Human Feedback AI, and Editor-in-Chief of IEEE Transactions on Games. A leading researcher in AI and games, Georgios shares insights from two decades of work at the intersection of artificial intelligence, player experience, and creative design. The discussion spans his seminal paper Choose Your Weapons: Survival Strategies for Depressed AI Academics, the value of human feedback in AI development, and the transformative role of AI in game research and design.

Key Topics Covered:
Georgios’s background in AI and game research, including co-founding Modl.ai and Human Feedback AI
The challenges facing AI academics today, including scaling maximalism and the need for smaller, human-centered models
Insights from the satirical yet practical paper Choose Your Weapons: Survival Strategies for Depressed AI Academics
The underrated importance of human data and human-in-the-loop approaches for AI alignment
Applications of AI in gaming, from procedural content generation to adaptive player modeling
How reinforcement learning and emotional modeling can create personalized, player-centric game experiences
The shift from building superhuman AI agents toward co-creative systems that assist game designers
Strategies like “try things that shouldn’t work” and how disruptive, risky research can lead to breakthroughs
Current projects in affective computing, user modeling, and generative AI for co-creation in games

Memorable Outtakes:
“As recently as 10 years ago, if you had a decent desktop computer and an internet connection, you had everything you needed to compete with the best researchers. Well, guess what? It’s not the case anymore.” – Georgios N. Yannakakis
“Humans are the next frontier of AI. We need human-centered models to build safer, explainable, and more trustworthy AI.” – Georgios N. Yannakakis
“The best use of generative AI in games is as a creative partner—an assistant in the design process—not as a fully autonomous game designer.” – Georgios N. Yannakakis

References & Resources:
Georgios Yannakakis University of Malta profile: https://www.um.edu.mt/profile/georgiosyannakakis
Georgios’ LinkedIn: https://www.linkedin.com/in/georgios-n-yannakakis-67a5b05
Modl.ai: https://modl.ai/
Human Feedback AI: https://humanfeedback.ai/
IEEE Transactions on Games: https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=7782673
Book Artificial Intelligence and Games: http://gameaibook.org/
Paper: Choose Your Weapons: Survival Strategies for Depressed AI Academics: https://ieeexplore.ieee.org/document/10458714
Paper: Pixels and Sound of Emotion: General Purpose Representations of Arousal in Games (IEEE Transactions on Affective Computing): https://arxiv.org/abs/2101.10706
Paper: BehAVE: Behaviour Alignment of Video Game Encodings: https://arxiv.org/html/2402.01335v3
Podcast with his colleague, Julian Togelius - From Turing’s Chess to Neural Game Engines: AI in Video Games Today: https://youtu.be/f_z2PpR3GaU

Sponsored by:
ODSC AI West 2025 – The Leading AI Training Conference
Join us in San Francisco from October 28th–30th for expert-led sessions on generative AI, LLMOps, and AI-driven automation.
Use the code podcast for 10% off any ticket.
Learn more: https://odsc.ai

Gaming Research in AI and the Need for Human-Centered Design with Georgios Yannakakis

In this episode of the ODSC Ai X Podcast, host Sheamus McGovern speaks with Sandi Besen, AI Engineer and Ecosystem Lead at IBM Research, about designing rule-based, production-ready AI agents using IBM’s open-source BeeAI framework. Sandi shares her journey from the performing arts to AI engineering, unpacks the challenges of agent reliability and trust, and explores how frameworks like BeeAI are solving real-world issues through enforceable rules, observability, and open standards. This episode dives deep into multi-agent systems, communication protocols, and what it really takes to deploy AI agents safely and effectively in enterprise environments.

Key Topics Covered:
Sandi’s background and current role at IBM Research
What inspired the creation of the BeeAI framework and platform
The concept of "agent chaos" and how BeeAI brings structure and reliability
Comparison of popular agent frameworks: LangChain, LangGraph, CrewAI
Introduction to emerging agent communication protocols (MCP and A2A)
Enforcing hardcoded rules (conditional requirements) in agents
Building agent behavior guardrails and preventing rogue actions
Importance of observability, memory, and tracing in agent frameworks
Using OpenTelemetry for tracing agent decision-making
Challenges in agent safety, validation, and trust for enterprise use
How BeeAI makes deploying an agent with a UI fast and easy (in under an hour)
When to build from scratch vs. using a framework like BeeAI 
Common mistakes startups make when adopting cutting-edge AI tech
Trends in context engineering and long-term memory
The future of consulting in the age of AI

Memorable Outtakes:
“Agent chaos is real. We needed a way to bring structure, enforce rules, and make agents safe for production.”
“Conditional requirements let you hardcode business logic into your agents — not just suggest behavior, but enforce it.”

References & Resources:
Sandi Besen – AI Engineer & Ecosystem Lead at IBM Research: https://research.ibm.com/people/sandi-besen
Website: https://sandibesen.com/
LinkedIn: https://www.linkedin.com/in/sandibesen/
BeeAI: https://www.ibm.com/think/news/beeai-open-source-multiagent
BeeAI GitHub Repository: https://github.com/i-am-bee/beeai-framework
Agent Communication Protocol (ACP) Course (DeepLearning.AI): https://www.deeplearning.ai/short-courses/acp-agent-communication-protocol/
OpenTelemetry:  https://opentelemetry.io/
Google’s P2 Payment Protocol (for AI agents): https://ap2-protocol.org/
AP2 GitHub: https://github.com/google-agentic-commerce/AP2
OpenAI Agent Commerce Protocol: https://developers.openai.com/commerce/

Sponsored by:
🔥 ODSC AI West 2025 – The Leading AI Training Conference
Join us in San Francisco from October 28th–30th for expert-led sessions on generative AI, LLMOps, and AI-driven automation.
Use the code podcast for 10% off any ticket.
Learn more: https://odsc.ai


AI Agents That Follow the Rules: Sandi Besen

In this episode of the ODSC Ai X Podcast, we speak with Noah Giansiracusa, a professor of mathematics at Bentley University and the author of Robin Hood Math (MIT Press, 2024). Noah makes the case that you don’t need to be a math genius to use numbers to challenge power, spot manipulation, and think more clearly in today’s data-saturated world. From social media algorithms to misleading rankings and risk perception, Noah explains how math is a tool for empowerment—not just abstraction.
Key Topics Covered:
Why “being pretty good at math” is enough for most real-world reasoning
The real-world implications of expected value and decision theory
How college and law school rankings are gamed—and what’s wrong with the metrics
How social media platforms like Facebook use engagement formulas to shape behavior
The problem with overestimating rare events 
Why being “numerically literate” is now a form of civic engagement
The role of anecdotes vs. data in shaping perception and risk
Memorable Outtakes:
“What I wanted to do with this book was show people that you don’t have to be 
“When we talk about risk, the things we fear the most are often not the things that are most likely to happen—and that’s a statistical insight.”
References & Resources:
Robin Hood Math by Noah Giansiracusa    https://www.noahgian.com/books
Noah Giansiracusa’s website  https://www.noahgian.com
Noah Giansiracusa’s academic page: https://faculty.bentley.edu/profile/ngiansiracusa
Mentioned research on COMPAS algorithm and algorithmic bias: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
College ranking methodology critique   https://en.wikipedia.org/wiki/Criticism_of_college_and_university_rankings_in_North_America
Sponsored by:
 🔥 ODSC AI West 2025 – The Leading AI Training Conference
 Join us in San Francisco from October 28th–30th for expert-led sessions on generative AI, LLMOps, and AI-driven automation.
 Use the code podcast for 10% off any ticket.
 Learn more: https://odsc.ai



Robin Hood Math: Taking the Algorithms Back with Noah Giansiracusa

In this episode, we speak with Kyle Corbitt, co-founder and CEO of OpenPip, recently acquired by CoreWeave, to explore the evolving role of reinforcement learning in building smarter, more reliable AI agents. Kyle shares the journey of OpenPipe from supervised fine-tuning to developing ART (Agent Reinforcement Trainer), their open-source RL toolkit designed to train AI agents that can think, adapt, and perform with greater autonomy. The discussion spans technical insights, practical applications, startup lessons from YC’s Startup School, and the future of agent-based AI systems.

Key Topics Covered:
Why reinforcement learning is gaining attention in modern Agent development
The transition from supervised fine-tuning (SFT) to reinforcement learning (RL)
Practical differences between RL and SFT, including weight movement and model reliability
OpenPipe’s approach with ART: supporting multi-turn agent training and tool use
How ART differs from OpenAI’s RFT implementation
The importance of consistent agent behavior in production and how RL helps
Avoiding reward hacking and the role of Ruler, OpenPipe’s LLM-based judging system
Cost-efficiency strategies in RL training using serverless infrastructure
OpenPipe’s long-term vision for self-improving agents
Advice for AI startup founders on building in a rapidly evolving ecosystem

Memorable Outtakes:
On why reinforcement learning matters now:
"What RL does… is it actually lets you solve the reliability problem. You can make your smaller model significantly more reliable—even more performant in your domain—through RL." 
On OpenAI's RFT vs. OpenPipe’s ART:
"What you're missing [with OpenAI's RFT API] is the ability to define custom tools and multi-turn agent behavior. With ART, you own the code, define the tools, and have full control of the reward signal."  

Guest information, Mentioned Tools and Projects:
Kyle Corbitt: https://www.linkedin.com/in/kcorbitt/
OpenPipe: https://openpipe.ai 
OpenPipe GitHub: https://github.com/OpenPipe/OpenPipe
ART (Agent Reinforcement Trainer): https://github.com/OpenPipe/ART
Ruler: LLM-based evaluation tool for training agents (part of the ART package): https://openpipe.ai/blog/ruler
DeepSeek R1: https://api-docs.deepseek.com/
OpenAI RFT: https://platform.openai.com/docs/guides/reinforcement-fine-tuning
SkyPilot: scale AI workloads  https://github.com/skypilot-org/skypilot

Notable Blog Posts:
Ruler (LLM-as-a-Judge): https://www.openpipe.ai/blog/ruler
Reward Hacking Post: https://www.openpipe.ai/blog/reward-hacking
Serverless RL Backend: https://www.openpipe.ai/blog/serverless-rl
ART Trainer: A New RL Trainer for Agents: https://openpipe.ai/blog/art-trainer-a-new-rl-trainer-for-agents

Sponsored by:
🔥 ODSC AI West 2025 – The Leading AI Training Conference
Join us in San Francisco from October 28th–30th for expert-led sessions on generative AI, LLMOps, and AI-driven automation.
Use the code podcast for 10% off any ticket.
Learn more: https://odsc.ai


Training Agents with Reinforcement Learning: Kyle Corbitt

In this special episode of the ODSC AI Podcast, host Sheamus McGovern, Founder of ODSC, shares a curated selection of short, focused Minisodes recorded live at ODSC AI West 2025. Across three days of hands-on sessions, keynotes, and workshops, these conversations capture the voices of eight experts driving the next wave of enterprise and applied AI, from responsible governance and trust engineering to coding with AI, open models, and the automation of discovery itself.
Guests include: Cal Al-Dhubaib, Ivan Lee, Sinan Ozdemir, Chip Huyen, Jepson Taylor, Helen O’Sullivan, Andrew Radin, and Steven Pousty — spanning startups, academia, and industry innovation.
Key Topics Covered:
▪️ Trust Engineering and AI Governance
▪️ Labeling Agents and Human-in-the-Loop Systems
▪️ Benchmarking and Real-World Performance Gaps
▪️ AI Engineering and the Future of Coding
▪️ Algorithmic Invention and Superhuman Code
▪️ Enterprise AI Infrastructure and Cost Optimization
▪️ AI Entrepreneurship and Measuring ROI
▪️ Computer Vision, Data Quality, and AI Sovereignty


Memorable Outtakes:
▪️ Cal Al-Dhubaib: “Trust is the currency of enterprise AI — you earn it over time, and one bad incident can spend it all.”
▪️ Ivan Lee: “Supervised learning is back. It never really left — we’ve just made it faster and smarter with labeling agents.”️
▪️Chip Huyen: “Saying prompt engineering is dead is like saying communication is dead.”


References & Resources:
▪️ Cal Al-Dhubaib – ⁠https://www.linkedin.com/in/dhubaib/⁠
▪️ Ivan Lee – Datasaur:⁠ ⁠https://www.linkedin.com/in/iylee/
▪️ Sinan Ozdemir – Crucible AI:⁠https://www.linkedin.com/in/sinan-ozdemir/⁠
▪️ Chip Huyen – The AI Engineer (O’Reilly Media):https://www.linkedin.com/in/chiphuyen/
▪️ Jepson Taylor – VEOX Inc.:⁠ ⁠⁠https://www.linkedin.com/in/jepsontaylor/⁠
▪️ Helen O’Sullivan – Dell Technologies:⁠ ⁠⁠https://www.linkedin.com/in/helen-o-sullivan-0596b1150/⁠
▪️ Andrew Radin – StartX:⁠ ⁠https://www.linkedin.com/in/andrewradin/
▪️ Steven Pousty – Red Hat:⁠ ⁠https://www.linkedin.com/in/thesteve0/
Sponsored by:
🔥 ODSC AI East 2026 – The Leading AI Training Conference
Join us in Boston from April 28th–30th for expert-led sessions on generative AI, LLMOps, and AI-driven automation.
Use the code podcast for 10% off any ticket. Learn more:⁠ https://odsc.ai/east⁠

ODSC West Minisodes: Trust, Agents, and the Next Wave of Applied AI

In this episode of the ODSC Ai X Podcast, host Sheamus McGovern speaks with Dr. Denis Garagić, CTO at Palladyne AI, about how intelligent systems are moving beyond the screen and into the physical world.
They explore the rise of embodied and physical AI, how it’s transforming robotics and manufacturing, and why autonomous drone swarms represent a powerful shift in real-time, decentralized decision-making. Denis also shares why hybrid models often outperform LLMs in production environments, and what’s needed to safely deploy these systems in industry and defense.
Key Topics Covered
How Denis Garagić’s work connects robotics, AI software, and autonomy.
What Embodied AI and Physical AI mean, and how they differ from traditional automation.
Why embedded, adaptive systems are replacing rigid, pre‑programmed robots in manufacturing.
The role of data, edge compute, and multi‐modal perception in enabling real‑world autonomy.
Internal limitations of large language models (LLMs) in industrial and physical systems, and why hybrid autonomy is often preferable.
How true drone swarms work: decentralized decision‑making, edge planning, and self‑organising UAV fleets.
Use cases in industrial manufacturing (automotive, aerospace, defense, construction) and in public‑safety/drone swarms.
The operational, organisational, and safety challenges of deploying embodied/physical AI in production: workforce readiness, verification & validation, regulatory considerations.
The future of embodied & physical AI: world‑models, runtime memory, sim‑to‑real transfer, and where the industry is headed next.
Memorable Outtakes
“Embodied AI is about learning through actions. You perceive, adapt, and reason while interacting with the physical world—just like we do as humans.”
“Drone swarms shouldn’t act like fleets. They should act like hives—decentralised, self‑organising, and communicating with only what’s necessary.”
“LLMs don’t scale well in industrial environments. Data is proprietary, compute is constrained, and the models need to be reliable — not just impressive.”

References & Resources
Dr. Denis Garagić’s bio (Palladyne AI): https://www.palladyneai.com/about-palladyne-ai/dr-denis-garagic/
Palladyne AI homepage: https://www.palladyneai.com
Blog: Autonomous Swarming Takes Flight: The Next Era of UAV Operations: https://www.palladyneai.com/blog/autonomous-swarming-takes-flight-the-next-era-of-uav-operations
Press release: Palladyne AI Achieves Key Milestone with First Small Drone Autonomous Tracking Flight:  https://www.tmcnet.com/usubmit/2024/12/23/10124356.htm TMCnet
News article: Palladyne AI and Draganfly partner for advanced UAV swarm autonomy: https://www.dronexl.co/2025/10/21/draganfly-ai-palladyne-autonomous-drone-swarms/ DroneXL.co
News article: Palladyne AI secures US Patent for advanced swarming and autonomy technology: https://www.asdnews.com/news/defense/2025/11/03/palladyne-ai-secures-us-patent-advanced-swarming-autonomy-technology ASD News


Sponsored by:
🔥 ODSC AI East 2026 – The Leading AI Training Conference
Join us in Boston from April 28th–30th for expert‑led sessions on Agentic AI, AI Engineering, Data Science, Machine Learning, LLMOps, and AI‑driven automation.
Learn more: https://odsc.ai/east


🧠 Agentic AI Summit – Learn to Build and Deploy Autonomous AI Systems
Happening online this January, the Agentic AI Summit is your chance to dive into hands‑on workshops and expert sessions on AI agents, workflows, and real‑world automation.
 Use the code podcast for 10% off any ticket for either event
Learn more: https://summit.ai



From Embodied and Physical AI to Autonomous Drone Swarms with Dr. Denis Garagic

With Ai X Podcast, Open Data Science Conference (ODSC) brings its vast experience in building community and its knowledge of the data science and AI fields to the podcast platform. The interests and challenges of the data science community are wide ranging. To reflect this Ai X Podcast will offer a similarly wide range of content, from one-on-one interviews with leading experts, to career talks, to educational interviews, to profiles of AI Startup Founders. Join us every two weeks to discover what’s going on in the data science community.

Find more ODSC lightning interviews, webinars, live trainings, certifications, bootcamps here - https://aiplus.training/  
Don't miss out on this exciting opportunity to expand your knowledge and stay ahead of the curve.

Deep Reinforcement Learning in the Real World with Anna Goldie

ODSC's Ai X Podcast

Related tracks