EP 594: Data Dreams & Digital Delusions: The role of AI in health tech
🎯 Summary
Podcast Episode Summary: EP 594: Data Dreams & Digital Delusions: The role of AI in health tech
This episode of the Everyday AI Show, hosted by Jordan Wilson, features guest Smurthy Kurubandandhan, a health tech executive with expertise in robotics, public health, and data science. The core discussion revolves around the massive, multi-billion dollar investments in AI data centers and infrastructure, questioning whether this scale of investment truly translates into exponentially better, safer outcomes, particularly in the high-stakes environment of health technology. The conversation balances the excitement around AI’s potential with the critical need for data integrity, transparency, and robust grounding mechanisms to prevent “digital delusions” like hallucinations.
1. Focus Area
The primary focus is the application, challenges, and strategic investment surrounding Artificial Intelligence (specifically Generative AI) within the Health Tech sector. Key themes include the necessity of data quality over sheer quantity, the risk of AI hallucinations in clinical settings, and the strategic implementation of grounding techniques like RAG.
2. Key Technical Insights
- The Criticality of RAG (Retrieval Augmented Generation): RAG is emphasized as a vital technical checkpoint to ground AI outputs in live, accurate data sources, preventing hallucinations. Without this grounding, systems risk becoming “bodies of misinformation,” which is unacceptable in healthcare.
- Transparency in the AI Pipeline: The need for a “glass model” of Gen AI was proposed, where organizations must be transparent about the data sets used for training, the frequency of retraining, and the reasoning chain of the model to build trust and accountability.
- Data Volume vs. Data Quality: The discussion challenges the assumption that simply accumulating massive amounts of data (as evidenced by huge data center investments) will automatically solve AI problems. The quality, cleanliness, and relevance of the data are paramount, especially when dealing with high-stakes decisions.
3. Business/Investment Angle
- Investment Drivers are Competitive: Large-scale data center investments are driven by a business imperative to maintain a competitive edge in the Gen AI market, often coupled with geopolitical incentives (e.g., US-based job creation).
- Focus Beyond Infrastructure: While infrastructure spending is high, there is a call for responsible investment that includes funding for data education, implementing robust guardrails (like RAG), and security frameworks, rather than just raw compute power.
- Health Tech Adoption Pace: Healthcare is lagging in widespread Gen AI implementation compared to other sectors due to the extreme sensitivity surrounding patient lives (PHI) and the high cost of errors, necessitating a “slow to go fast” approach.
4. Notable Companies/People
- Smurthy Kurubandandhan: The expert guest, whose background spans robotics, public health, and data science, provided the critical perspective on balancing technological advancement with public health responsibility.
- Large Tech Companies: Mentioned generally as the entities making “hundreds of billions of dollars” in data center investments (e.g., referencing the scale of projects like Stargate).
- OpenAI/Gemma: Mentioned in the context of the growing availability of powerful, downloadable, open-source models, which could accelerate adoption in privacy-conscious sectors like health tech if run on-premise.
5. Future Implications
The industry is heading toward a necessary reckoning where data integrity and verifiable reasoning will supersede raw model size. There is an expectation that AI will significantly reduce physician burnout (e.g., through clinical note scribing) and optimize enterprise functions (HR, supply chain, F&A). However, the future success in health tech hinges on establishing unwavering transparency and accountability to mitigate the catastrophic impact of unchecked hallucinations on patient care, finances, and mental well-being.
6. Target Audience
This episode is highly valuable for Health Tech Executives, Digital Health Strategists, AI Product Leaders, and Investors focused on the intersection of AI infrastructure and regulated industries. It provides actionable insights on implementation strategy (RAG) and risk management in high-stakes data environments.
🏢 Companies Mentioned
đź’¬ Key Insights
"Say someone puts in an MRI and X-ray saying, 'Does this person have, say, cancer?' And say the AI system is hallucinating and says, 'Yes, the potential of this person having cancer is 80 percent,' and it's wrong. Think about the mental, the physical, the financial burden on the patient, on the provider."
"Where is the data being extracted from? That is the first step, right? The second one is understanding: is the data that's being extracted from the prompt, is it live? Is it accurate? Is it from a web source? Like, are those sources credible? That's why RAG comes into play, the grounding of that is key."
"Because healthcare, unlike other industries, such as people's lives, right? This is not retail. I mean, obviously, they touch people's lives, but not, you know, the actual lives. I think going slow to go fast is important."
"The grounding prevents hallucinations. Without the grounding, we're all just taking information and becoming these bodies of misinformation because perception is reality. What we read is true."
"I personally, Jordan, believe that everything in life needs a checkpoint, right? Every process needs a checkpoint, and it's healthy for especially a data-driven system to have a checkpoint, which is what RAG is in a very simple way, right?"
"how much of these models and answers that we are retrieving is accurate, right? So some of the things that I do advise clients and work on is how do they implement a RAG, which is Retrieval Augmented Generation, to make sure that the data that they're extracting is from the right data sets and is being cross-checked live and is not just, you know, from the abundance of stored data that is not real-time."