Summary
In this episode of the AI Engineering Podcast Kasimir Schulz, Director of Security Research at HiddenLayer, talks about the complexities and security challenges in AI and machine learning models. Kasimir explains the concept of shadow genes and shadow logic, which involve identifying common subgraphs within neural networks to understand model ancestry and potential vulnerabilities, and emphasizes the importance of understanding the attack surface in AI integrations, scanning models for security threats, and evolving awareness in AI security practices to mitigate risks in deploying AI systems.


Announcements

• Hello and welcome to the AI Engineering Podcast, your guide to the fast-moving world of building scalable and maintainable AI systems
• Your host is Tobias Macey and today I'm interviewing Kasimir Schulz about the relationships between the various models on the market and how that information helps with selecting and protecting models for your applications

Interview

• Introduction
• How did you get involved in machine learning?
• Can you start by outlining the current state of the threat landscape for ML and AI systems?
• What are the main areas of overlap in risk profiles between prediction/classification and generative models? (primarily from an attack surface/methodology perspective)
  • What are the significant points of divergence?
• What are some of the categories of potential damages that can be created through the deployment of compromised models?
• How does the landscape of foundation models introduce new challenges around supply chain security for organizations building with AI?
• You recently published your findings on the potential to inject subgraphs into model architectures that are invisible during normal operation of the model. Along with that you wrote about the subgraphs that are shared between different classes of models. What are the key learnings that you would like to highlight from that research?
  • What action items can organizations and engineering teams take in light of that information?
• Platforms like HuggingFace offer numerous variations of popular models with variations around quantization, various levels of finetuning, model distillation, etc. That is obviously a benefit to knowledge sharing and ease of access, but how does that exacerbate the potential threat in the face of backdoored models?
• Beyond explicit backdoors in model architectures, there are numerous attack vectors to generative models in the form of prompt injection, "jailbreaking" of system prompts, etc. How does the knowledge of model ancestry help with identifying and mitigating risks from that class of threat?
  • A common response to that threat is the introduction of model guardrails with pre- and post-filtering of prompts and responses. How can that approach help to address the potential threat of backdoored models as well?
• For a malicious actor that develops one of these attacks, what is the vector for introducing the compromised model into an organization?
  • Once that model is in use, what are the possible means by which the malicious actor can detect its presence for purposes of exploitation?
• What are the most interesting, innovative, or unexpected ways that you have seen the information about model ancestry used?
• What are the most interesting, unexpected, or challenging lessons that you have learned while working on ShadowLogic/ShadowGenes?
• What are some of the other means by which the operation of ML and AI systems introduce attack vectors to organizations running them?

Contact Info

• LinkedIn

Parting Question

• From your perspective, what are the biggest gaps in tooling, technology, or training for AI systems today?

Closing Announcements

• Thank you for listening! Don't forget to check out our other shows. The Data Engineering Podcast covers the latest on modern data management. Podcast.__init__ covers the Python language, its community, and the innovative ways it is being used.
• Visit the site to subscribe to the show, sign up for the mailing list, and read the show notes.
• If you've learned something or tried out a project from the show then tell us about it! Email [email protected] with your story.
• To help other people find the show please leave a review on iTunes and tell your friends and co-workers.

Links

• HiddenLayer
• Zero-Day Vulnerability
• MCP Blog Post
• Python Pickle Object Serialization
• SafeTensors
• Deepseek
• Huggingface Transformers
• KROP == Knowledge Return Oriented Prompting
• XKCD "Little Bobby Tables"
• OWASP Top 10 For LLMs
• CVE AI Systems Working Group
• Refusal Vector Ablation
• Foundation Model
• ShadowLogic
• ShadowGenes
• Bytecode
• ResNet == Resideual Neural Network
• YOLO == You Only Look Once
• Netron
• BERT
• RoBERTA
• Shodan
• CTF == Capture The Flag
• Titan Bedrock Image Generator

The intro and outro music is from Hitman's Lovesong feat. Paola Graziano by The Freak Fandango Orchestra/CC BY-SA 3.0

Summary
In this episode of the AI Engineering podcast Julian LaNeve, CTO of Astronomer, talks about transitioning from simple LLM applications to more complex agentic AI systems. Julian shares insights into the challenges and considerations of this evolution, emphasizing the importance of starting with simpler applications to build operational knowledge and intuition. He discusses the parallels between microservices and agentic AI, highlighting the need for careful orchestration and observability to manage complexity and ensure reliability, and explores the technical requirements for deploying AI systems, including data infrastructure, orchestration tools like Apache Airflow, and understanding the probabilistic nature of AI models.


Announcements

• Hello and welcome to the AI Engineering Podcast, your guide to the fast-moving world of building scalable and maintainable AI systems
• Seamless data integration into AI applications often falls short, leading many to adopt RAG methods, which come with high costs, complexity, and limited scalability. Cognee offers a better solution with its open-source semantic memory engine that automates data ingestion and storage, creating dynamic knowledge graphs from your data. Cognee enables AI agents to understand the meaning of your data, resulting in accurate responses at a lower cost. Take full control of your data in LLM apps without unnecessary overhead. Visit aiengineeringpodcast.com/cognee to learn more and elevate your AI apps and agents.
• Your host is Tobias Macey and today I'm interviewing Julian LaNeve about how to avoid putting the cart before the horse with AI applications. When do you move from "simple" LLM apps to agentic AI and what's the path to get there?

Interview

• Introduction
• How did you get involved in machine learning?
• How do you technically distinguish "agentic AI" (e.g., involving planning, tool use, memory) from "simpler LLM workflows" (e.g., stateless transformations, RAG)? What are the key differences in operational complexity and potential failure modes?
• What specific technical challenges (e.g., state management, observability, non-determinism, prompt fragility, cost explosion) are often underestimated when teams jump directly into building stateful, autonomous agents?
• What are the pre-requisites from a data and infrastructure perspective before going to production with agentic applications?
  • How does that differ from the chat-based systems that companies might be experimenting with?
• Technically, where do you most often see ambitious agent projects break down during development or early deployment?
• Beyond generic data quality, what specific data engineering practices become critical when building reliable LLM applications? (e.g., Designing data pipelines for efficient RAG chunking/embedding, versioning prompts alongside data, caching strategies for LLM calls, managing vector database ETL).
• From an implementation complexity standpoint, what characterizes tasks well-suited for initial LLM workflow adoption versus those genuinely requiring agentic capabilities?
  • Can you share examples (anonymized if necessary) highlighting how organizations successfully engineered these simpler LLM workflows? What specific technical designs, tooling choices, or MLOps practices were key to their reliability and scalability?
• What are some hard-won technical or operational lessons from deploying and scaling LLM workflows in production environments? Any surprising performance bottlenecks, cost issues, or monitoring challenges engineers should anticipate?
• What technical maturity signals (e.g., robust CI/CD for ML, established monitoring/alerting for pipelines, automated evaluation frameworks, cost tracking mechanisms) suggest an engineering team might be ready to tackle the challenges of building and operating agentic systems?
• How does the technical stack and engineering process need to evolve when moving from orchestrated LLM workflows towards more complex agents involving memory, planning, and dynamic tool use? What new components and failure modes must be engineered for?
• How do you foresee orchestration platforms evolving to better serve the needs of AI engineers building LLM apps? 
• What are the most interesting, innovative, or unexpected ways that you have seen organizations build toward advanced AI use cases?
• What are the most interesting, unexpected, or challenging lessons that you have learned while working on supporting AI services?
• When is AI the wrong choice?
• What is the single most critical piece of engineering advice you would give to fellow AI engineers who are tasked with integrating LLMs into production systems right now?

Contact Info

• LinkedIn
• GitHub

Parting Question

• From your perspective, what are the biggest gaps in tooling, technology, or training for AI systems today?

Links

• Astronomer
• Airflow
• Anthropic
• Building Effective Agents post from Anthropic
• Airflow 3.0
• Microservices
• Pydantic AI
• Langchain
• LlamaIndex
• LLM As A Judge
• SWE (SoftWare Engineer) Bench
• Cursor
• Windsurf
• OpenTelemetry
• DAG == Directed Acyclic Graph
• Halting Problem
• AI Long Term Memory

The intro and outro music is from Hitman's Lovesong feat. Paola Graziano by The Freak Fandango Orchestra/CC BY-SA 3.0

Summary
In this episode of the AI Engineering Podcast Emmanouil (Manos) Koukoumidis, CEO of Oumi, about his vision for an open platform for building, evaluating, and deploying AI foundation models. Manos shares his journey from working on natural language AI services at Google Cloud to founding Oumi with a mission to advance open-source AI, emphasizing the importance of community collaboration and accessibility. He discusses the need for open-source models that are not constrained by proprietary APIs, highlights the role of Oumi in facilitating open collaboration, and touches on the complexities of model development, open data, and community-driven advancements in AI. He also explains how Oumi can be used throughout the entire lifecycle of AI model development, post-training, and deployment.


Announcements

• Hello and welcome to the AI Engineering Podcast, your guide to the fast-moving world of building scalable and maintainable AI systems
• Your host is Tobias Macey and today I'm interviewing Manos Koukoumidis about Oumi, an all-in-one production-ready open platform to build, evaluate, and deploy AI models

Interview

• Introduction
• How did you get involved in machine learning?
• Can you describe what Oumi is and the story behind it?
• There are numerous projects, both full suites and point solutions, focused on every aspect of "AI" development. What is the unique value that Oumi provides in this ecosystem?
• You have stated the desire for Oumi to become the Linux of AI development. That is an ambitious goal and one that Linux itself didn't start with. What do you see as the biggest challenges that need addressing to reach a critical mass of adoption?
• In the vein of "open source" AI, the most notable project that I'm aware of that fits the proper definition is the OLMO models from AI2. What lessons have you learned from their efforts that influence the ways that you think about your work on Oumi?
• On the community building front, HuggingFace has been the main player. What do you see as the benefits and shortcomings of that platform in the context of your vision for open and collaborative AI?
• Can you describe the overall design and architecture of Oumi?
  • How did you approach the selection process for the different components that you are building on top of?
  • What are the extension points that you have incorporated to allow for customization/evolution?
• Some of the biggest barriers to entry for building foundation models are the cost and availability of hardware used for training, and the ability to collect and curate the data needed. How does Oumi help with addressing those challenges?
• For someone who wants to build or contribute to an open source model, what does that process look like?
  • How do you envision the community building/collaboration process?
• Your overall goal is to build a foundation for the growth and well-being of truly open AI. How are you thinking about the sustainability of the project and the funding needed to grow and support the community?
• What are the most interesting, innovative, or unexpected ways that you have seen Oumi used?
• What are the most interesting, unexpected, or challenging lessons that you have learned while working on Oumi?
• When is Oumi the wrong choice?
• What do you have planned for the future of Oumi?

Contact Info

• LinkedIn

Parting Question

• From your perspective, what are the biggest gaps in tooling, technology, or training for AI systems today?

Closing Announcements

• Thank you for listening! Don't forget to check out our other shows. The Data Engineering Podcast covers the latest on modern data management. Podcast.__init__ covers the Python language, its community, and the innovative ways it is being used.
• Visit the site to subscribe to the show, sign up for the mailing list, and read the show notes.
• If you've learned something or tried out a project from the show then tell us about it! Email [email protected] with your story.
• To help other people find the show please leave a review on iTunes and tell your friends and co-workers.

Links

• Oumi
• Cloud PaLM
• Google Gemini
• DeepMind
• LSTM == Long Short-Term Memory
• Transfomers)
• ChatGPT
• Partial Differential Equation
• OLMO
• OSI AI definition
• MLFlow
• Metaflow
• SkyPilot
• Llama
• RAG
  • Podcast Episode
• Synthetic Data
  • Podcast Episode
• LLM As Judge
• SGLang
• vLLM
• Function Calling Leaderboard
• Deepseek

The intro and outro music is from Hitman's Lovesong feat. Paola Graziano by The Freak Fandango Orchestra/CC BY-SA 3.0

Summary
In this episode of the AI Engineering Podcast Adil Hafiz talks about the Arch project, a gateway designed to simplify the integration of AI agents into business systems. He discusses how the gateway uses Rust and Envoy to provide a unified interface for handling prompts and integrating large language models (LLMs), allowing developers to focus on core business logic rather than AI complexities. The conversation also touches on the target audience, challenges, and future directions for the project, including plans to develop a leading planning LLM and enhance agent interoperability.


Announcements

• Hello and welcome to the AI Engineering Podcast, your guide to the fast-moving world of building scalable and maintainable AI systems
• Your host is Tobias Macey and today I'm interviewing Adil Hafeez about the Arch project, a gateway for your AI agents

Interview

• Introduction
• How did you get involved in machine learning?
• Can you describe what Arch is and the story behind it?
• How do you think about the target audience for Arch and the types of problems/projects that they are responsible for?
• The general category of LLM gateways is largely oriented toward abstracting the specific model provider being called. What are the areas of overlap and differentiation in Arch?
• Many of the features in Arch are also available in AI frameworks (e.g. LangChain, LlamaIndex, etc.), such as request routing, guardrails, and tool calling. How do you think about the architectural tradeoffs of having that functionality in a gateway service?
• What is the workflow for someone building an application with Arch?
• Can you describe the architecture and components of the Arch gateway?
• With the pace of change in the AI/LLM ecosystem, how have you designed the Arch project to allow for rapid evolution and extensibility?
• What are the most interesting, innovative, or unexpected ways that you have seen Arch used?
• What are the most interesting, unexpected, or challenging lessons that you have learned while working on Arch?
• When is Arch the wrong choice?
• What do you have planned for the future of Arch?

Contact Info

• LinkedIn
• GitHub

Parting Question

• From your perspective, what are the biggest gaps in tooling, technology, or training for AI systems today?

Closing Announcements

• Thank you for listening! Don't forget to check out our other shows. The Data Engineering Podcast covers the latest on modern data management. Podcast.__init__ covers the Python language, its community, and the innovative ways it is being used.
• Visit the site to subscribe to the show, sign up for the mailing list, and read the show notes.
• If you've learned something or tried out a project from the show then tell us about it! Email [email protected] with your story.
• To help other people find the show please leave a review on iTunes and tell your friends and co-workers.

Links

• Arch Gateway
• Gradient Boosting
• Envoy
• LLM Gateway
• Huggingface
• Katanemo Models
• Qwen2.5
• Rust Clippy

The intro and outro music is from Hitman's Lovesong feat. Paola Graziano by The Freak Fandango Orchestra/CC BY-SA 3.0

Summary
In this episode of the AI Engineering Podcast Ali Golshan, co-founder and CEO of Gretel.ai, talks about the transformative role of synthetic data in AI systems. Ali explains how synthetic data can be purpose-built for AI use cases, emphasizing privacy, quality, and structural stability. He highlights the shift from traditional methods to using language models, which offer enhanced capabilities in understanding data's deep structure and generating high-quality datasets. The conversation explores the challenges and techniques of integrating synthetic data into AI systems, particularly in production environments, and concludes with insights into the future of synthetic data, including its application in various industries, the importance of privacy regulations, and the ongoing evolution of AI systems.


Announcements

• Hello and welcome to the AI Engineering Podcast, your guide to the fast-moving world of building scalable and maintainable AI systems
• Seamless data integration into AI applications often falls short, leading many to adopt RAG methods, which come with high costs, complexity, and limited scalability. Cognee offers a better solution with its open-source semantic memory engine that automates data ingestion and storage, creating dynamic knowledge graphs from your data. Cognee enables AI agents to understand the meaning of your data, resulting in accurate responses at a lower cost. Take full control of your data in LLM apps without unnecessary overhead. Visit aiengineeringpodcast.com/cognee to learn more and elevate your AI apps and agents.
• Your host is Tobias Macey and today I'm interviewing Ali Golshan about the role of synthetic data in building, scaling, and improving AI systems

Interview

• Introduction
• How did you get involved in machine learning?
• Can you start by summarizing what you mean by synthetic data in the context of this conversation?
• How have the capabilities around the generation and integration of synthetic data changed across the pre- and post-LLM timelines?
• What are the motivating factors that would lead a team or organization to invest in synthetic data generation capacity?
• What are the main methods used for generation of synthetic data sets?
  • How does that differ across open-source and commercial offerings?
• From a surface level it seems like synthetic data generation is a straight-forward exercise that can be owned by an engineering team. What are the main "gotchas" that crop up as you move along the adoption curve?
  • What are the scaling characteristics of synthetic data generation as you go from prototype to production scale?
• domains/data types that are inappropriate for synthetic use cases (e.g. scientific or educational content)
• managing appropriate distribution of values in the generation process
• Beyond just producing large volumes of semi-random data (structured or otherwise), what are the other processes involved in the workflow of synthetic data and its integration into the different systems that consume it?
• What are the most interesting, innovative, or unexpected ways that you have seen synthetic data generation used?
• What are the most interesting, unexpected, or challenging lessons that you have learned while working on synthetic data generation?
• When is synthetic data the wrong choice?
• What do you have planned for the future of synthetic data capabilities at Gretel?

Contact Info

• LinkedIn

Parting Question

• From your perspective, what are the biggest gaps in tooling, technology, or training for AI systems today?

Closing Announcements

• Thank you for listening! Don't forget to check out our other shows. The Data Engineering Podcast covers the latest on modern data management. Podcast.__init__ covers the Python language, its community, and the innovative ways it is being used.
• Visit the site to subscribe to the show, sign up for the mailing list, and read the show notes.
• If you've learned something or tried out a project from the show then tell us about it! Email [email protected] with your story.
• To help other people find the show please leave a review on iTunes and tell your friends and co-workers.

Links

• Gretel
• Hadoop
• LSTM == Long Short-Term Memory
• GAN == Generative Adversarial Network
• Textbooks are all you need MSFT paper
• Illumina

The intro and outro music is from Hitman's Lovesong feat. Paola Graziano by The Freak Fandango Orchestra/CC BY-SA 3.0

Summary
In this episode of the AI Engineering podcast Viraj Mehta, CTO and co-founder of TensorZero, talks about the use of LLM gateways for managing interactions between client-side applications and various AI models. He highlights the benefits of using such a gateway, including standardized communication, credential management, and potential features like request-response caching and audit logging. The conversation also explores TensorZero's architecture and functionality in optimizing AI applications by managing structured data inputs and outputs, as well as the challenges and opportunities in automating prompt generation and maintaining interaction history for optimization purposes.


Announcements

• Hello and welcome to the AI Engineering Podcast, your guide to the fast-moving world of building scalable and maintainable AI systems
• Seamless data integration into AI applications often falls short, leading many to adopt RAG methods, which come with high costs, complexity, and limited scalability. Cognee offers a better solution with its open-source semantic memory engine that automates data ingestion and storage, creating dynamic knowledge graphs from your data. Cognee enables AI agents to understand the meaning of your data, resulting in accurate responses at a lower cost. Take full control of your data in LLM apps without unnecessary overhead. Visit aiengineeringpodcast.com/cognee to learn more and elevate your AI apps and agents. 
• Your host is Tobias Macey and today I'm interviewing Viraj Mehta about the purpose of an LLM gateway and his work on TensorZero

Interview

• Introduction
• How did you get involved in machine learning?
• What is an LLM gateway?
  • What purpose does it serve in an AI application architecture?
• What are some of the different features and capabilities that an LLM gateway might be expected to provide?
• Can you describe what TensorZero is and the story behind it?
  • What are the core problems that you are trying to address with Tensor0 and for whom?
• One of the core features that you are offering is management of interaction history. How does this compare to the "memory" functionality offered by e.g. LangChain, Cognee, Mem0, etc.?
• How does the presence of TensorZero in an application architecture change the ways that an AI engineer might approach the logic and control flows in a chat-based or agent-oriented project?
• Can you describe the workflow of building with Tensor0 and some specific examples of how it feeds back into the performance/behavior of an LLM?
• What are some of the ways in which the addition of Tensor0 or another LLM gateway might have a negative effect on the design or operation of an AI application?
• What are the most interesting, innovative, or unexpected ways that you have seen TensorZero used?
• What are the most interesting, unexpected, or challenging lessons that you have learned while working on TensorZero?
• When is TensorZero the wrong choice?
• What do you have planned for the future of TensorZero?

Contact Info

• LinkedIn

Parting Question

• From your perspective, what are the biggest gaps in tooling, technology, or training for AI systems today?

Closing Announcements

• Thank you for listening! Don't forget to check out our other shows. The Data Engineering Podcast covers the latest on modern data management. Podcast.__init__ covers the Python language, its community, and the innovative ways it is being used.
• Visit the site to subscribe to the show, sign up for the mailing list, and read the show notes.
• If you've learned something or tried out a project from the show then tell us about it! Email [email protected] with your story.
• To help other people find the show please leave a review on iTunes and tell your friends and co-workers.

Links

• TensorZero
• LLM Gateway
• LiteLLM
• OpenAI
• Google Vertex
• Anthropic
• Reinforcement Learning
• Tokamak Reactor
• Viraj RLHF Paper
• Contextual Dueling Bandits
• Direct Preference Optimization
• Partially Observable Markov Decision Process
• DSPy
• PyTorch
• Cognee
• Mem0
• LangGraph
• Douglas Hofstadter
• OpenAI Gym
• OpenAI o1
• OpenAI o3
• Chain Of Thought

The intro and outro music is from Hitman's Lovesong feat. Paola Graziano by The Freak Fandango Orchestra/CC BY-SA 3.0

Summary
In this episode of the AI Engineering Podcast Ron Green, co-founder and CTO of KungFu AI, talks about the evolving landscape of AI systems and the challenges of harnessing generative AI engines. Ron shares his insights on the limitations of large language models (LLMs) as standalone solutions and emphasizes the need for human oversight, multi-agent systems, and robust data management to support AI initiatives. He discusses the potential of domain-specific AI solutions, RAG approaches, and mixture of experts to enhance AI capabilities while addressing risks. The conversation also explores the evolving AI ecosystem, including tooling and frameworks, strategic planning, and the importance of interpretability and control in AI systems. Ron expresses optimism about the future of AI, predicting significant advancements in the next 20 years and the integration of AI capabilities into everyday software applications.


Announcements

• Hello and welcome to the AI Engineering Podcast, your guide to the fast-moving world of building scalable and maintainable AI systems
• Seamless data integration into AI applications often falls short, leading many to adopt RAG methods, which come with high costs, complexity, and limited scalability. Cognee offers a better solution with its open-source semantic memory engine that automates data ingestion and storage, creating dynamic knowledge graphs from your data. Cognee enables AI agents to understand the meaning of your data, resulting in accurate responses at a lower cost. Take full control of your data in LLM apps without unnecessary overhead. Visit aiengineeringpodcast.com/cognee to learn more and elevate your AI apps and agents. 
• Your host is Tobias Macey and today I'm interviewing Ron Green about the wheels that we need for harnessing the power of the generative AI engine

Interview

• Introduction
• How did you get involved in machine learning?
• Can you describe what you see as the main shortcomings of LLMs as a stand-alone solution (to anything)?
• The most established vehicle for harnessing LLM capabilities is the RAG pattern. What are the main limitations of that as a "product" solution?
• The idea of multi-agent or mixture-of-experts systems is a more sophisticated approach that is gaining some attention. What do you see as the pro/con conversation around that pattern?
• Beyond the system patterns that are being developed there is also a rapidly shifting ecosystem of frameworks, tools, and point solutions that plugin to various points of the AI lifecycle. How does that volatility hinder the adoption of generative AI in different contexts?
  • In addition to the tooling, the models themselves are rapidly changing. How much does that influence the ways that organizations are thinking about whether and when to test the waters of AI?
• Continuing on the metaphor of LLMs and engines and the need for vehicles, where are we on the timeline in relation to the model T Ford?
  • What are the vehicle categories that we still need to design and develop? (e.g. sedans, mini-vans, freight trucks, etc.)
• The current transformer architecture is starting to reach scaling limits that lead to diminishing returns. Given your perspective as an industry veteran, what are your thoughts on the future trajectory of AI model architectures?
  • What is the ongoing role of regression style ML in the landscape of generative AI?
• What are the most interesting, innovative, or unexpected ways that you have seen LLMs used to power a "vehicle"?
• What are the most interesting, unexpected, or challenging lessons that you have learned while working in this phase of AI?
• When is generative AI/LLMs the wrong choice?

Contact Info

• LinkedIn

Parting Question

• From your perspective, what are the biggest gaps in tooling, technology, or training for AI systems today?

Closing Announcements

• Thank you for listening! Don't forget to check out our other shows. The Data Engineering Podcast covers the latest on modern data management. Podcast.__init__ covers the Python language, its community, and the innovative ways it is being used.
• Visit the site to subscribe to the show, sign up for the mailing list, and read the show notes.
• If you've learned something or tried out a project from the show then tell us about it! Email [email protected] with your story.
• To help other people find the show please leave a review on iTunes and tell your friends and co-workers.

Links

• Kungfu.ai
• Llama open generative AI models
• ChatGPT
• Copilot
• Cursor
• RAG == Retrieval Augmented Generation
  • Podcast Episode
• Mixture of Experts
• Deep Learning
• Random Forest
• Supervised Learning
• Active Learning)
• Yann LeCunn
• RLHF == Reinforcement Learning from Human Feedback
• Model T Ford
• Mamba selective state space
• Liquid Network
• Chain of thought
• OpenAI o1
• Marvin Minsky
• Von Neumann Architecture
• Attention Is All You Need
• Multilayer Perceptron
• Dot Product
• Diffusion Model
• Gaussian Noise
• AlphaFold 3
• Anthropic
• Sparse Autoencoder

The intro and outro music is from Hitman's Lovesong feat. Paola Graziano by The Freak Fandango Orchestra/CC BY-SA 3.0

Summary
In this episode of the AI Engineering Podcast Jim Olsen, CTO of ModelOp, talks about the governance of generative AI models and applications. Jim shares his extensive experience in software engineering and machine learning, highlighting the importance of governance in high-risk applications like healthcare. He explains that governance is more about the use cases of AI models rather than the models themselves, emphasizing the need for proper inventory and monitoring to ensure compliance and mitigate risks. The conversation covers challenges organizations face in implementing AI governance policies, the importance of technical controls for data governance, and the need for ongoing monitoring and baselines to detect issues like PII disclosure and model drift. Jim also discusses the balance between innovation and regulation, particularly with evolving regulations like those in the EU, and provides valuable perspectives on the current state of AI governance and the need for robust model lifecycle management.


Announcements

• Hello and welcome to the AI Engineering Podcast, your guide to the fast-moving world of building scalable and maintainable AI systems
• Your host is Tobias Macey and today I'm interviewing Jim Olsen about governance of your generative AI models and applications

Interview

• Introduction
• How did you get involved in machine learning?
• Can you describe what governance means in the context of generative AI models? (e.g. governing the models, their applications, their outputs, etc.)
• Governance is typically a hybrid endeavor of technical and organizational policy creation and enforcement. From the organizational perspective, what are some of the difficulties that teams are facing in understanding what those policies need to encompass?
  • How much familiarity with the capabilities and limitations of the models is necessary to engage productively with policy debates?
• The regulatory landscape around AI is still very nascent. Can you give an overview of the current state of legal burden related to AI?
  • What are some of the regulations that you consider necessary but as-of-yet absent?
• Data governance as a practice typically relates to controls over who can access what information and how it can be used. The controls for those policies are generally available in the data warehouse, business intelligence, etc. What are the different dimensions of technical controls that are needed in the application of generative AI systems?
  • How much of the controls that are present for governance of analytical systems are applicable to the generative AI arena?
• What are the elements of risk that change when considering internal vs. consumer facing applications of generative AI?
  • How do the modalities of the AI models impact the types of risk that are involved? (e.g. language vs. vision vs. audio)
• What are some of the technical aspects of the AI tools ecosystem that are in greatest need of investment to ease the burden of risk and validation of model use?
• What are the most interesting, innovative, or unexpected ways that you have seen AI governance implemented?
• What are the most interesting, unexpected, or challenging lessons that you have learned while working on AI governance?
• What are the technical, social, and organizational trends of AI risk and governance that you are monitoring?

Contact Info

• LinkedIn

Parting Question

• From your perspective, what are the biggest gaps in tooling, technology, or training for AI systems today?

Closing Announcements

• Thank you for listening! Don't forget to check out our other shows. The Data Engineering Podcast covers the latest on modern data management. Podcast.__init__ covers the Python language, its community, and the innovative ways it is being used.
• Visit the site to subscribe to the show, sign up for the mailing list, and read the show notes.
• If you've learned something or tried out a project from the show then tell us about it! Email [email protected] with your story.
• To help other people find the show please leave a review on iTunes and tell your friends and co-workers.

Links

• ModelOp
• Foundation Models
• GDPR
• EU AI Regulation
• Llama 2
• AWS Bedrock
• Shadow IT
• RAG == Retrieval Augmented Generation
  • Podcast Episode
• Nvidia NEMO
• LangChain
• Shapley Values
• Gibberish Detection

The intro and outro music is from Hitman's Lovesong feat. Paola Graziano by The Freak Fandango Orchestra/CC BY-SA 3.0

Summary
In this episode of the AI Engineering Podcast, Vasilije Markovich talks about enhancing Large Language Models (LLMs) with memory to improve their accuracy. He discusses the concept of memory in LLMs, which involves managing context windows to enhance reasoning without the high costs of traditional training methods. He explains the challenges of forgetting in LLMs due to context window limitations and introduces the idea of hierarchical memory, where immediate retrieval and long-term information storage are balanced to improve application performance. Vasilije also shares his work on Cognee, a tool he's developing to manage semantic memory in AI systems, and discusses its potential applications beyond its core use case. He emphasizes the importance of combining cognitive science principles with data engineering to push the boundaries of AI capabilities and shares his vision for the future of AI systems, highlighting the role of personalization and the ongoing development of Cognee to support evolving AI architectures.


Announcements

• Hello and welcome to the AI Engineering Podcast, your guide to the fast-moving world of building scalable and maintainable AI systems
• Your host is Tobias Macey and today I'm interviewing Vasilije Markovic about adding memory to LLMs to improve their accuracy

Interview

• Introduction
• How did you get involved in machine learning?
• Can you describe what "memory" is in the context of LLM systems?
• What are the symptoms of "forgetting" that manifest when interacting with LLMs?
  • How do these issues manifest between single-turn vs. multi-turn interactions?
• How does the lack of hierarchical and evolving memory limit the capabilities of LLM systems?
• What are the technical/architectural requirements to add memory to an LLM system/application?
• How does Cognee help to address the shortcomings of current LLM/RAG architectures?
• Can you describe how Cognee is implemented?
  • Recognizing that it has only existed for a short time, how have the design and scope of Cognee evolved since you first started working on it?
• What are the data structures that are most useful for managing the memory structures?
• For someone who wants to incorporate Cognee into their LLM architecture, what is involved in integrating it into their applications?
  • How does it change the way that you think about the overall requirements for an LLM application?
• For systems that interact with multiple LLMs, how does Cognee manage context across those systems? (e.g. different agents for different use cases)
• There are other systems that are being built to manage user personalization in LLm applications, how do the goals of Cognee relate to those use cases? (e.g. Mem0 - https://github.com/mem0ai/mem0)
• What are the unknowns that you are still navigating with Cognee?
• What are the most interesting, innovative, or unexpected ways that you have seen Cognee used?
• What are the most interesting, unexpected, or challenging lessons that you have learned while working on Cognee?
• When is Cognee the wrong choice?
• What do you have planned for the future of Cognee?

Contact Info

• LinkedIn

Parting Question

• From your perspective, what are the biggest gaps in tooling, technology, or training for AI systems today?

Closing Announcements

• Thank you for listening! Don't forget to check out our other shows. The Data Engineering Podcast covers the latest on modern data management. Podcast.__init__ covers the Python language, its community, and the innovative ways it is being used.
• Visit the site to subscribe to the show, sign up for the mailing list, and read the show notes.
• If you've learned something or tried out a project from the show then tell us about it! Email [email protected] with your story.
• To help other people find the show please leave a review on iTunes and tell your friends and co-workers.

Links

• Cognee
• Montenegro
• Catastrophic Forgetting
• Multi-Turn Interaction
• RAG == Retrieval Augmented Generation
  • Podcast Episode
• GraphRAG
  • Podcast Episode
• Long-term memory
• Short-term memory
• Langchain
• LlamaIndex
• Haystack
• dlt
  • Data Engineering Podcast Episode
• Pinecone
  • Podcast Episode
• Agentic RAG
• Airflow
• DAG == Directed Acyclic Graph
• FalkorDB
• Neo4J
• Pydantic
• AWS ECS
• AWS SNS
• AWS SQS
• AWS Lambda
• LLM As Judge
• Mem0
• QDrant
• LanceDB
• DuckDB

The intro and outro music is from Hitman's Lovesong feat. Paola Graziano by The Freak Fandango Orchestra/CC BY-SA 3.0

Summary
In this episode of the AI Engineering Podcast, Tanner Burson, VP of Engineering at Prismatic, talks about the evolving impact of generative AI on software developers. Tanner shares his insights from engineering leadership and data engineering initiatives, discussing how AI is blurring the lines of developer roles and the strategic value of AI in software development. He explores the current landscape of AI tools, such as GitHub's Copilot, and their influence on productivity and workflow, while also touching on the challenges and opportunities presented by AI in code generation, review, and tooling. Tanner emphasizes the need for human oversight to maintain code quality and security, and offers his thoughts on the future of AI in development, the importance of balancing innovation with practicality, and the evolving role of engineers in an AI-driven landscape.


Announcements

• Hello and welcome to the AI Engineering Podcast, your guide to the fast-moving world of building scalable and maintainable AI systems
• Your host is Tobias Macey and today I'm interviewing Tanner Burson about the impact of generative AI on software developers

Interview

• Introduction
• How did you get involved in machine learning?
• Can you describe what types of roles and work you consider encompassed by the term "developers" for the purpose of this conversation?
• How does your work at Prismatic give you visibility and insight into the effects of AI on developers and their work?
• There have been many competing narratives about AI and how much of the software development process it is capable of encompassing. What is your top-level view on what the long-term impact on the job prospects of software developers will be as a result of generative AI?
• There are many obvious examples of utilities powered by generative AI that are focused on software development. What do you see as the categories or specific tools that are most impactful to the development cycle?
• In what ways do you find familiarity with/understanding of LLM internals useful when applying them to development processes?
• As an engineering leader, how are you evaluating and guiding your team on the use of AI powered tools?
  • What are some of the risks that you are guarding against as a result of AI in the development process?
• What are the most interesting, innovative, or unexpected ways that you have seen AI used in the development process?
• What are the most interesting, unexpected, or challenging lessons that you have learned while using AI for software development?
• When is AI the wrong choice for a developer?
• What are your projections for the near to medium term impact on the developer experience as a result of generative AI?

Contact Info

• LinkedIn

Parting Question

• From your perspective, what are the biggest gaps in tooling, technology, or training for AI systems today?

Closing Announcements

• Thank you for listening! Don't forget to check out our other shows. The Data Engineering Podcast covers the latest on modern data management. Podcast.__init__ covers the Python language, its community, and the innovative ways it is being used.
• Visit the site to subscribe to the show, sign up for the mailing list, and read the show notes.
• If you've learned something or tried out a project from the show then tell us about it! Email [email protected] with your story.
• To help other people find the show please leave a review on iTunes and tell your friends and co-workers.

Links

• Prismatic
• Google AI Development announcement
• Tabnine
  • Podcast Episode
• GitHub Copilot
• Plandex
• OpenAI API
• Amazon Q
• Ollama
• Huggingface Transformers
• Anthropic
• Langchain
• Llamaindex
• Haystack
• Llama 3.2
• Qwen2.5-Coder

The intro and outro music is from Hitman's Lovesong feat. Paola Graziano by The Freak Fandango Orchestra/CC BY-SA 3.0