What if the secret to successful AI governance lies in understanding the evolution of model documentation? In this episode, our hosts challenge the common belief that model cards marked the start of documentation in AI. We explore model documentation practices, from their crucial beginnings in fields like finance to their adaptation in Silicon Valley. Our discussion also highlights the important role of early modelers and statisticians in advocating for a complete approach that includes the entire model development lifecycle.

Show Notes

Model documentation origins and best practices (1:03)

• Documenting a model is a comprehensive process that requires giving users and auditors clear understanding: 
  • Why was the model built? 
  • What data goes into a model? 
  • How is the model implemented? 
  • What does the model output? 

Model cards - pros and cons (7:33)

• Model cards for model reporting, Association for Computing Machinery
• Evolution from this research to Google's definition to today
• How the market perceives them vs. what they are
• Why the analogy “nutrition labels for models” needs a closer look

System cards - pros and cons (12:03)

• To their credit, OpenAI system cards somewhat bridge the gap between proper model documentation and a model card.
• Contains complex descriptions of evaluation methodologies along with results; extra points for reporting red-teaming results
• Represents 3rd-party opinions of the social and ethical implications of the release of the model

Automating model documentation with generative AI (17:17)

• Finding the balance in automation in a great governance strategy
• Generative AI can provide an assist in editing and personal workflow

Improving documentation for AI governance (23:11)

• As model expert, engage from the beginning with writing the bulk of model documentation by hand.
• The exercise of documenting your models solidifies your understanding of the model's goals, values, and methods for the business

What did you think? Let us know.

Do you have a question or a discussion topic for the AI Fundamentalists? Connect with them to comment on your favorite topics:

• LinkedIn - Episode summaries, shares of cited articles, and more.
• YouTube - Was it something that we said? Good. Share your favorite quotes.
• Visit our page - see past episodes and submit your feedback! It continues to inspire future episodes.

Are businesses ready for large language models as a path to AI? In this episode, the hosts reflect on the past year of what has changed and what hasn’t changed in the world of LLMs. Join us as we debunk the latest myths and emphasize the importance of robust risk management in AI integration. The good news is that many decisions about adoption have forced businesses to discuss their future and impact in the face of emerging technology. You won't want to miss this discussion.

• Intro and news: The veto of California's AI Safety Bill (00:00:03)
  • Can state-specific AI regulations really protect consumers, or do they risk stifling innovation? (Gov. Newsome's response)
  • Veto highlights the critical need for risk-based regulations that don't rely solely on the size and cost of language models 
  • Arguments to be made for a cohesive national framework that ensures consistent AI regulation across the United States
• Are businesses ready to embrace large language models, or are they underestimating the challenges? (00:08:35) 
  • The myth that acquiring a foundational model is a quick fix for productivity woes 
  • The essential role of robust risk management strategies, especially in sensitive sectors handling personal data
  • Review of model cards, Open AI's system cards, and the importance of thorough testing, validation, and stricter regulations to prevent a false sense of security
  • Transparency alone is not enough; objective assessments are crucial for genuine progress in AI integration
• From hallucinations in language models to ethical energy use, we tackle some of the most pressing problems in AI today (00:16:29)
  • Reinforcement learning with annotators and the controversial use of other models for review
  • Jan LeCun's energy systems and retrieval-augmented generation (RAG) offer intriguing alternatives that could reshape modeling approaches
• The ethics of advancing AI technologies, consider the parallels with past monumental achievements and the responsible allocation of resources (00:26:49)
  • There is good news about developments and lessons learned from LLMs; but there is also a long way to go.
  • Our original predictions in episode 2 for LLMs still reigns true: “Reasonable expectations of LLMs: Where truth matters and risk tolerance is low, LLMs will not be a good fit”
  • With increased hype and awareness from LLMs came varying levels of interest in how all model types and their impacts are governed in a business.

What did you think? Let us know.

Do you have a question or a discussion topic for the AI Fundamentalists? Connect with them to comment on your favorite topics:

• LinkedIn - Episode summaries, shares of cited articles, and more.
• YouTube - Was it something that we said? Good. Share your favorite quotes.
• Visit our page - see past episodes and submit your feedback! It continues to inspire future episodes.

Our special guest, astrophysicist Rachel Losacco, explains the intricacies of galaxies, modeling, and the computational methods that unveil their mysteries. She shares stories about how advanced computational resources enable scientists to decode galaxy interactions over millions of years with true-to-life accuracy. Sid and Andrew discuss transferable practices for building resilient modeling systems. 

• Prologue: Why it's important to bring stats back [00:00:03]
  • Announcement from the American Statistical Association (ASA): Data Science Statement Updated to Include “ and AI”  
• Today's guest: Rachel Losacco [00:02:10]
  • Rachel is an astrophysicist who’s worked with major galaxy formation simulations for many years. She hails from Leiden (Lie-den) University and the University of Florida. As a Senior Data Scientist, she works on modeling road safety.  
• Defining complex systems through astrophysics [00:02:52]
  • Discussion about origins and adoption of complex systems
  • Difficulties with complex systems: Nonlinearity, chaos and randomness, collective dynamics and hierarchy, and emergence.
• Complexities of nonlinear systems [00:08:20]
  • Linear models (Least Squares, GLMs, SVMs) can be incredibly powerful but they cannot model all possible functions (e.g. a decision boundary of concentric circles)
  • Non-linearity and how it exists in the natural world
• Chaos and randomness [00:11:30]
  • Enter references to Jurassic Park and The Butterfly Effect
  • “In universe simulations, a change to a single parameter can govern if entire galaxy clusters will ever form” - Rachel
• Collective dynamics and hierarchy [00:15:45]
  • Interactions between agents don’t occur globally and often is mediated through effects that only happen on specific sub-scales
  • Adaptation: components of systems breaking out of linear relationships between inputs and outputs to better serve the function of the greater system   
• Emergence and complexity [00:23:36]
  • New properties arise from the system that cannot be explained by the base rules governing the system
• Examples in astrophysics [00:24:34]
  • These difficulties are parts of solving previously impossible problems
  • Consider this lecture from IIT Delhi on Complex Systems to get a sense of what is required to study and formalize a complex system and its collective dynamics (https://www.youtube.com/watch?v=yJ39ppgJlf0)
• Consciousness and reasoning from a new point of view [00:31:45]
  • Non-linearity, hierarchy, feedback loops, and emergence may be ways to study consciousness. The brain is a complex system that a simple set of rules cannot fully define.
  • See: Brain modeling from scratch of C. Elgans 

What did you think? Let us know.

Do you have a question or a discussion topic for the AI Fundamentalists? Connect with them to comment on your favorite topics:

• LinkedIn - Episode summaries, shares of cited articles, and more.
• YouTube - Was it something that we said? Good. Share your favorite quotes.
• Visit our page - see past episodes and submit your feedback! It continues to inspire future episodes.

Can your AI models survive a big disaster? While a recent major IT incident with CrowdStrike wasn't AI related, the magnitude and reaction reminded us that no system no matter how proven is immune to failure. AI modeling systems are no different. Neglecting the best practices of building models can lead to unrecoverable failures. Discover how the three-tiered framework of robustness, resiliency, and anti-fragility can guide your approach to creating AI infrastructures that not only perform reliably under stress but also fail gracefully when the unexpected happens.

Show Notes

• Technology, incidents, and why basics matter (00:00:03)
  • While the recent Crowdstrike incident wasn't caused by AI, it's impact was a wakeup call for people and processes that support critical systems
  • As AI is increasingly being used at both experimental and production levels, we can expect AI incidents are a matter of if, not when. What can you do to prepare?

• The "7P's": Are you capable of handling the unexpected? (00:09:05)
  • The 7Ps is an adage, dating back to WWII, that aligns with our "do things the hard way" approach to AI governance and modeling systems.
  • Let’s consider the levels of building a performant system: Robustness, Resiliency, and Antifragility

• Model robustness (00:10:03)
  • Robustness is a very important but often overlooked component of building modeling systems. We suspect that part of the problem is due to: 
    • The Kaggle-driven upbringing of data scientists
    • Assumed generalizability of modeling systems, when models are optimized to perform well on their training data but do not generalize enough to perform well on unseen data.

• Model resilience (00:16:10)
  • Resiliency is the ability to absorb adverse stimuli without destruction and return to its pre-event state.
  • In practice, robustness and resiliency, testing, and planning are often easy components to leave out. This is where risks and threats are exposed.
  • See also, Episode 8. Model validation: Robustness and resilience

• Models and antifragility (00:25:04)
  • Unlike resiliency, which is the ability to absorb damaging inputs without breaking, antifragility is the ability of a system to improve from challenging stimuli. (i.e. the human body)
  • A key question we need to ask ourselves if we are not actively building our AI systems to be antifragile, why are we using AI systems at all?

What did you think? Let us know.

Do you have a question or a discussion topic for the AI Fundamentalists? Connect with them to comment on your favorite topics:

• LinkedIn - Episode summaries, shares of cited articles, and more.
• YouTube - Was it something that we said? Good. Share your favorite quotes.
• Visit our page - see past episodes and submit your feedback! It continues to inspire future episodes.

Join us as we chat with Patrick Hall, Principal Scientist at Hallresearch.ai and Assistant Professor at George Washington University. He shares his insights on the current state of AI, its limitations, and the potential risks associated with it. The conversation also touched on the importance of responsible AI, the role of the National Institute of Standards and Technology (NIST) AI Risk Management Framework (RMF) in adoption, and the implications of using generative AI in decision-making.

Show notes

Governance, model explainability, and high-risk applications 00:00:03 

• Intro to Patrick
• His latest book: Machine Learning for High-Risk Applications: Approaches to Responsible AI (2023)

The benefits of NIST AI Risk Management Framework 00:04:01 

• Does not have a profit motive, which avoids the potential for conflicts of interest when providing guidance on responsible AI. 
• Solicits, adjudicates, and incorporates feedback from the public and other stakeholders.
• NIST is not law, however it's recommendations set companies up for outcome-based reviews by regulators.

Accountability challenges in "blame-free" cultures 00:10:24 

• Cites these cultures have the hardest time with the framework's recommendations
• Practices like documentation and fair model reviews need accountability and objectivity
• If everyone's responsible, no one's responsible.

The value of explainable models vs black-box models 00:15:00 

• Concerns about replacing explainable models with LLMs for LLM's sake 
• Why generative AI is bad for decision-making 

AI and its impact on students 00:21:49 

• Students are more indicative of where the hype and market is today
• Teaching them how to work through the best model for the best job despite the hype

AI incidents and contextual failures 00:26:17 

• AI Incident Database 
• AI, as it currently stands, is a memorizing and calculating technology. It lacks the ability to incorporate real-world context.
• McDonald's AI Drive-Thru debacle is a warning to us all

Generative AI and homogenization problems 00:34:30

Recommended resources from Patrick:

• Ed Zitron “Better Offline” 
• NIST ARIA 
• AI Safety Is a Narrative Problem

What did you think? Let us know.

Do you have a question or a discussion topic for the AI Fundamentalists? Connect with them to comment on your favorite topics:

• LinkedIn - Episode summaries, shares of cited articles, and more.
• YouTube - Was it something that we said? Good. Share your favorite quotes.
• Visit our page - see past episodes and submit your feedback! It continues to inspire future episodes.

Ready to uncover the secrets of modern systems engineering and the future of AI? Join us for an enlightening conversation with Matt Barlin, the Chief Science Officer of Valence. Matt's extensive background in systems engineering and data lineage sets the stage for a fascinating discussion. He sheds light on the historical evolution of the field, the critical role of documentation, and the early detection of defects in complex systems. This episode promises to expand your understanding of model-based systems and data issues, offering valuable insights that only an expert of Matt's caliber can provide.

In the heart of our episode, we dive into the fundamentals and transformative benefits of data lineage in AI. Matt draws intriguing parallels between data lineage and the engineering life cycle, stressing the importance of tracking data origins, access rights, and verification processes. Discover how decentralized identifiers are paving the way for individuals to control and monetize their own data. With the phasing out of third-party cookies and the challenges of human-generated training data shortages, we explore how systems like retrieval-augmented generation (RAG) and compliance regulations like the EU AI Act are shaping the landscape of AI data quality and compliance. 

Don’t miss this thought-provoking episode that promises to keep you at the forefront of responsible AI.

What did you think? Let us know.

Do you have a question or a discussion topic for the AI Fundamentalists? Connect with them to comment on your favorite topics:

• LinkedIn - Episode summaries, shares of cited articles, and more.
• YouTube - Was it something that we said? Good. Share your favorite quotes.
• Visit our page - see past episodes and submit your feedback! It continues to inspire future episodes.

Explore the basics of differential privacy and its critical role in protecting individual anonymity. The hosts explain the latest guidelines and best practices in applying differential privacy to data for models such as AI. Learn how this method also makes sure that personal data remains confidential, even when datasets are analyzed or hacked.

Show Notes

• Intro and AI news (00:00) 
  • Google AI search tells users to glue pizza and eat rocks 
  • Gary Marcus on break? (Maybe and X only break)
• What is differential privacy? (06:34)
  • Differential privacy is a process for sensitive data anonymization that offers each individual in a dataset the same privacy they would experience if they were removed from the dataset entirely.
  • NIST’s recent paper SP 800-226 IPD: “Any privacy harms that result form a differentially private analysis could have happened if you had not contributed your data”.
  • There are two main types of differential privacy: global (NIST calls it Central) and local
• Why should people care about differential privacy? (11:30)
  • Interest has been increasing for organizations to intentionally and systematically prioritize the privacy and safety of user data
  • Speed up deployments of AI systems for enterprise customers since connections to raw data do not need to be established
  • Increase data security for customers that utilize sensitive data in their modeling systems
  • Minimize the risk of sensitive data exposure for your data privileges - i.e. Don’t be THAT organization
• Guidelines and resources for applied differential privacy
  • Guidelines for Evaluating Differential Privacy Guarantees: 
  • NIST De-Identification
• Practical examples of applied differential privacy (15:58)
  • Continuous Features - cite: Dwork, McSherry, Nissim, and Smith’s 2006 seminal paper "Calibrating Noise to Sensitivity in Private Data Analysis”[2], introduces a concept called ε-differential privacy
  • Categorical Features - cite: Warner (1965) created a randomized response technique in his paper titled: “Randomized Response: A Survey Technique for Eliminating Evasive Answer Bias” 
• Summary and key takeaways (23:59)
  • Differential privacy is going to be a part of how many of us need to manage data privacy
  • Data providers can’t provide us with anonymized data for analysis or when anonymization isn’t enough for our privacy needs
  • Hopeful that cohort targeting takes over for individual targeting
  • Remember: Differential privacy does not prevent bias!

What did you think? Let us know.

Do you have a question or a discussion topic for the AI Fundamentalists? Connect with them to comment on your favorite topics:

• LinkedIn - Episode summaries, shares of cited articles, and more.
• YouTube - Was it something that we said? Good. Share your favorite quotes.
• Visit our page - see past episodes and submit your feedback! It continues to inspire future episodes.

Artificial Intelligence (AI) stands at a unique intersection of technology, ethics, and regulation. The complexities of responsible AI are brought into sharp focus in this episode featuring Anthony Habayeb, CEO and co-founder of Monitaur,  As responsible AI is scrutinized for its role in profitability and innovation, Anthony and our hosts discuss the imperatives of safe and unbiased modeling systems, the role of regulations, and the importance of ethics in shaping AI.

Show notes

Prologue: Why responsible AI? Why now? (00:00:00)

• Deviating from our normal topics about modeling best practices
• Context about where regulation plays a role in industries besides big tech
• Can we learn from other industries about the role of "responsibility" in products?

 Special guest, Anthony Habayeb (00:02:59)

• Introductions and start of the discussion
• Of all the companies you could build around AI, why governance?

Is responsible AI the right phrase? (00:11:20)

• Should we even call good modeling and business practices "responsible AI"?
• Is having responsible AI a “want to have?” or a “need to have?”

Importance of AI regulation and responsibility (00:14:49)

• People in the AI and regulation worlds have started pushing back on Responsible AI.
• Do regulations impede freedom?
• Discussing the big picture of responsibility and governance: Explainability, repeatability, records, and audit

What about bias and fairness? (00:22:40)

• You can have fair models that operate with bias
• Bias in practice identifies inequities that models have learned
• Fairness is correcting for societal biases to level the playing field for safer business and modeling practices to prevail.

Responsible deployment and business management (00:35:10)

• Discussion about what organizations get right about responsible AI
• And what organizations can get completely wrong if they aren't careful.

Embracing responsible AI practices (00:41:15)

• Getting your teams, companies, and individuals involved in the movement towards building AI responsibly

What did you think? Let us know.

Do you have a question or a discussion topic for the AI Fundamentalists? Connect with them to comment on your favorite topics:

• LinkedIn - Episode summaries, shares of cited articles, and more.
• YouTube - Was it something that we said? Good. Share your favorite quotes.
• Visit our page - see past episodes and submit your feedback! It continues to inspire future episodes.

Baseline modeling is a necessary part of model validation. In our expert opinion, it should be required before model deployment. There are many baseline modeling types and in this episode, we're discussing their use cases, strengths, and weaknesses. We're sure you'll appreciate a fresh take on how to improve your modeling practices.

Show notes

Introductions and news: why reporting and visibility is a good thing for AI 0:03

• Spoiler alert: Providing visibility to AI bias audits does NOT mean exposing trade secrets. Some reports claim otherwise.
• Discussion about AI regulation in the context of current events and how regulation is playing out between Boeing and the FAA (tbc)

Understanding baseline modeling for machine learning 7:41

• Establishing baselines allows us to understand how models perform relative to simple rules-based models, aka heuristics.
• Reporting results without baselines to compare against is like giving a movie a rating of 5 without telling the listener that you were using a 10-point scale.
• Baseline modeling comparisons are part of rigorous model validations and should always be conducted during early model development and final production deployment.
• Pairs with analyses of theoretical upper bounds for modeling performance to show how your technique scores between acceptable worst and best case performance.
• We often find complex models being deployed in the real world that haven’t proven their value over simpler and explainable baseline models

Classification baselines and model performance comparison 19:40

• Uniform Random Selection - simulate how your model does against a baseline model that guesses classes randomly like a dice.
• Most Frequent Class (MFC) - the most telling test and often the most telling test in the case of highly skewed data with inappropriate metrics.
• Single-feature modeling - Validates how much the complex signal from your data and model improves over a bare minimum explainable model.
• And more…

Exploring regression and more advanced baselines for modeling 24:11

• Regression baselines: mean, median mode, Single-variable linear regression, Lag 1, and Least 5% re-interpretation
• Advanced baselines in language and vision

Conclusions 35:39

• Baseline modeling is a necessary part of model validation
• There are differing flavors of baselines that are appropriate for all types of modeling
• Baselines are needed to establish fair and realistic lower bounds for performance
• If your model can’t perform significantly better than a baseline consider scrapping the model and trying a new approach

What did you think? Let us know.

Do you have a question or a discussion topic for the AI Fundamentalists? Connect with them to comment on your favorite topics:

• LinkedIn - Episode summaries, shares of cited articles, and more.
• YouTube - Was it something that we said? Good. Share your favorite quotes.
• Visit our page - see past episodes and submit your feedback! It continues to inspire future episodes.

In this episode, we explore information theory and the not-so-obvious shortcomings of its popular metrics for model monitoring; and where non-parametric statistical methods can serve as the better option.

 Introduction and latest news 0:03

• Gary Marcus has written an article questioning the hype around generative AI, suggesting it may not be as transformative as previously thought.
• This in contrast to announcements out of the NVIDIA conference during the same week.

Information theory and its applications in AI. 3:45

• The importance of information theory in computer science, citing its applications in cryptography and communication.
• The basics of information theory, including the concept of entropy, which measures the uncertainty of a random variable.
• Information theory as a fundamental discipline in computer science, and how it has been applied in recent years, particularly in the field of machine learning.
• The speakers clarify the difference between a metric and a divergence, which is crucial to understanding how information theory is being misapplied in some cases

Information theory metrics and their limitations. 7:05

• Divergences are a type of measurement that don't follow simple rules like distance, and they have some nice properties but can be troublesome in certain use cases.
• KL Divergence is a popular test for monitoring changes in data distributions, but it's not symmetric and can lead to incorrect comparisons.
• Sid explains that KL divergence measures the slight surprisal or entropy difference between moving from one data distribution to another, and is not the same as KS test.

Metrics for monitoring AI model changes. 10:41

• The limitations of KL divergence and its alternatives, including Jenson Shannon divergence and population stability index.
• They highlight the issues with KL divergence, such as asymmetry and handling of zeros, and the advantages of Jenson Shannon divergence, which can handle both issues, and population stability index, which provides a quantitative measure of changes in model distributions.
• The popularity of information theory metrics in AI and ML is largely due to legacy and a lack of understanding of the underlying concepts.
•  Information theory metrics may not be the best choice for quantifying change in risk in the AI and ML space, but they are the ones that are commonly used due to familiarity and ease of use.

Using nonparametric statistics in modeling systems. 15:09

• Information theory divergences are not useful for monitoring production model performance, according to the speakers.
• Andrew Clark highlights the advantages of using nonparametric statistics in machine learning, including distribution agnosticism and the ability to test for significance without knowing the underlying distribution.
• Sid Mangali

What did you think? Let us know.

Do you have a question or a discussion topic for the AI Fundamentalists? Connect with them to comment on your favorite topics:

• LinkedIn - Episode summaries, shares of cited articles, and more.
• YouTube - Was it something that we said? Good. Share your favorite quotes.
• Visit our page - see past episodes and submit your feedback! It continues to inspire future episodes.