Generative AI in the Real World

Join us for a conversation between Ben Lorica and Matt Welsh, cofounder of Fixie.ai, former engineer at Apple and Google, and one of Mark Zuckerberg’s professors at Harvard. Learn how AI is changing computing. Whether it’s in C or a human language, programming is telling a computer what you want it to do—but AI opens up new classes of things that we can ask it to do.It’s not just simplifying (or replacing) coding; it’s creating new opportunities and new kinds of applications that we couldn’t imagine two or three years ago.Points of Interest0:00: Introduction.2:38: The changing nature of programming. What will replace programming?3:07: Ultimately, the idea of writing a program will be replaced by telling a language model what you want to do. The language model will do what you want directly.5:03: I can do things I couldn’t imagine doing—for example, summarize a transcript or find bios of speakers and relevant papers.7:01: There’s a whole new field of kinds of computation we couldn’t do before.7:48: People in fields like medicine used to have to ask computer scientists to do things for them. Now, you don’t have to get a computer scientist to translate an idea into reality.11:30: What is missing from the current tooling?11:40: It’s way too hard for people without programming ability to integrate language models into their workflows. Ultimately, AI needs to be deeply integrated into products and the OS.13:45: Are people in the UX community inventing new ways to interact?14:40: We are very embedded in a web/mobile-based way of thinking about interacting. AI changes the ways we interact with computers—for example, voice.16:07: There’s a lot of information encoded into voice that you miss when you encode it into text.18:15: What about programming itself?18:30: Programming is changing radically. At Fixie, we mandated that employees have access to ChatGPT and similar tools.20:34: What is the role of testing and QA?21:28: People will struggle to find the right trade-offs. We’re not throwing out all of the processes we’ve developed, like testing and code reviews.25:25: Every company can train AI to scale their best engineers.25:55: We’re being sloppy as an industry. Curation of good code and good documents will be important. We don’t just need more data, we need better data.28:23: What is Aryn doing?29:17: When people wanted to use AI models to ask questions about their data, they started with simple processes: break text into chunks, store in vector database, and at question time, feed them back in to the prompt.30:10: We need the ability to extract data from unstructured documents. The structure is there, but it’s hidden. The first part of Aryn: How do you extract the structure inherent in documents?32:46: The second part of Aryn: A Python framework, Sycamore, lets you build ETL pipelines from these documents. ETL does things like normalize location information.35:45: Another part of the Aryn stack is LLM-powered unstructured analytics (LUNA) that allows you to make queries based on the unstructured data in the documents.37:34: The future of programming is using language models as computers to perform computation that would be difficult to express in a programming language.38:22: People are talking about GraphRAG, which is RAG with knowledge graphs, but how do you get a knowledge graph? Can Aryn help that?39:15: Yes, we’re effectively doing knowledge graph construction. But once you have the right underlying structure, you may not need knowledge graphs at all.40:50: Are tools for evaluating AI lagging behind development tools?41:16: The meaning of “evaluation” is often not well-defined.43:03: Evaluation will come down to establishing trust.43:32: We need tools that will allow people to collaborate early on evaluations. You need to give people that help them understand what’s happening.

What can AI do to improve healthcare? Kingsley Ndoh, founder of Hurone AI, talks with Ben Lorica about how Hurone is making cancer care more effective for people who are underserved by the medical system. He discusses how AI can streamline the medical process, both helping doctors to treat patients more effectively and making clinical trials more diverse.Points of Interest0:36: What motivated you to apply AI to cancer care? What problems are you trying to solve?1:39: We need environments for training AI models that are effective for all populations.2:31: Current oncology solutions serve advanced healthcare systems, leaving community oncology centers and international markets underserved.3:31: Lack of diversity in clinical trials means we don’t have full evidence on the efficacy of drugs.5:00: What is an oncologist?6:10: Cancer is a very complex disease; every cancer is different and has its own solutions.6:43: What advantages do you bring as a domain expert?7:11: I’ve been a physician taking care of patients. I understand clinical workflows in Nigeria and the US. I’ve also been an entrepreneur since I was in high school. I’ve also worked in the global oncology space with governments and pharma companies. That network is very important.9:15: What was the situation before Gukiza [Hurone’s app]? What does Gukiza enable today?9:44: Gukiza makes care more accessible to patients and optimizes workflows for oncologists. They may have to travel long distances to see an oncologist; they may have side effects or even emergencies that are avoidable; data about events may be lost.12:53: Gukiza streamlines the process; it’s a two-way system that can be used standalone. There is a HIPPA-compliant API that can be integrated into major electronic medical records systems. Patients aren’t limited to an app; there is an API for WhatsApp, Telegram, and text messaging.14:13: Patients can describe their problems. Clinicians can click a button and generate a response that they can review and send to the patient. Clinicians can also call patients, do clinical summaries, and see how patients are progressing.17:08: One should think about this as a copilot. The app makes suggestions; the physician makes the decision.17:35: There are definitely risks. We are building our model and fine-tuning it to ensure that hallucination is limited. But there is still a final human review.18:40: What if I want to use the system in a completely new country? What does it take to get the system into a viable, usable state?19:41: We conform to the country’s guidelines for the management of patients. Cancer care is usually based on established guidelines. In the US, we have NCCN guidelines. To make sure guidelines are responsive to different regions, the NCCN looked at evidence for research done in different countries to harmonize guidelines. That gave birth to the resource stratified guidelines for regions like Sub-Saharan Africa. We don’t need to customize a lot.21:38: We are also building agreements for access to de-identified cancer data. As we scale, it will get better.24:02: Health data is the most sensitive data in the world, but also the most abundant. Compared to other industries, healthcare is lagging behind. But many regions are looking for disruption and innovation and are willing to be flexible to work with us.25:20: Our solution isn’t a magic bullet, but it will shift the needle.26:12: We are excited about LLMs with text and images. But before LLMs, people were excited about computer vision. What models are you using?27:10: We’re relying on LLMs and NLPs. There are established startups with computer vision for radiology and pathology; we are partnering with those companies. The major data we collect is genomic data. We are also incorporating wearable device data with things like geolocation, sleep patterns, heart rates, etc.28:28: Social determinants of health data are also important: ZIP code, employment status, activities, food.

Rikin Gandhi, CTO of Digital Green, talks with Ben Lorica about using generative AI to help farmers in developing countries become more productive. Farmer.Chat integrates information from training videos, sources of weather and crop information, and other data sources in a multimodal app that farmers can use in real-time.Points of Interest0:45: Digital Green helps farmers become more productive. Two years ago, Digital Green developed Farmer.Chat, an app that uses generative AI to put local language training videos together with weather data, market information, and other data.2:09: Our primary data source is our library of 10,000 videos in 40 languages that have been produced by farmers. We integrate additional sources for weather and market information. More recently, we’ve added information support tools.3:38: We have a smartphone app. Users who only have feature phones can call into a number and interact with a bot.5:00: Prior to Farmer.Chat, our work was primarily offline: videos shown on mobile projectors to an in-person audience. Sending content to phones flips the paradigm: rather than attending a video, farmers can ask questions relevant to their situation.6:40: When did you realize that generative AI opened up new possibilities? It was a gradual transition from offline videos on projectors. COVID didn’t allow us to get groups of farmers together. And more farmers came online in the same period.8:17: We had a deterministic bot before Farmer.Chat. But users had to traverse a tree to get the information they wanted. That tree was challenging to create and difficult to use.9:33: With GPT-3, we saw that we could move away from complexity and cost of using a deterministic bot.11:15: Did ChatGPT alert you to more possibilities? ChatGPT has scoured open internet knowledge. Farmers are looking for location and time-specific information. Even in the earliest version of ChatGPT, we saw that it had a lot of this information. Putting this world together with our video was powerful.13:07: Accuracy, precision, and recall are all important. Are you fine-tuning and using RAG to make sure you are accurate? We had problems with hallucinations even within our knowledge base. We implemented reranking and filtering, which reduced hallucinations to <1%. We’ve created a golden Q&A set.16:01: People are now talking about GraphRAG, the use of knowledge graphs for RAG. Can you create a knowledge graph because you know your data so well? A lot of concepts in agriculture are related—for example, crop calendars for how crops develop. We’re trying to build those relations into the system.17:05: We are leveraging agentic orchestration for the overall pipeline. Based on the user’s query, we may be able to answer questions directly rather than go through the RAG pipeline.18:44: Your situation is inherently multimodal: video, speech-to-text, voice; is this a challenge? We’re now using tools like GPT Vision to get descriptive metadata about what’s in videos. It becomes part of the database. We began with text queries; we added voice support. And now people can take a photo of a crop or an animal.21:04: Foundation models are becoming multimodal. What’s your user interface today? What are you moving towards? We started with messaging apps that the users already use. We’re plugging the bot into that ecosystem. We’re migrating towards a reality that isn’t text first: putting video first so farmers can speak and take a video. For many farmers, this is the first time they’ve interacted with a bot. Autoprompts are important so they know that it has weather and locale-specific information.23:57: What are specific challenges around AI—privacy, security, and ethics? Agriculture is often a sensitive subject. There’s a lot of personally identifiable information. We try to mask that information so it’s not used to train models. Farmers need to be able to trust that their information won’t be taken away from them.

Timothy Persons of PricewaterhouseCoopers (PwC) talks with Ben Lorica about adoption of AI in the enterprise. They discuss the challenges enterprises experience, including the need to change corporate culture. To succeed, it’s important to focus on solving well-defined problems rather than just doing something cool with AI. Good data strategies and data governance are essential. Persons also highlights the importance of training and education for everyone in the organization and the need to create safe environments where people can experiment.Points of Interest0:00: Introduction.1:00: We are seeing an uptick in adoption of AI in the enterprise. CEOs are planning to adopt AI and pursue business reinvention. Many companies are still kicking the tires. There is more adoption in the backend where risks are lower.3:36: AI budgets are on an upward trend. It is not a small spend and there’s a tendency to underestimate cost.4:54: What are some of the key challenges that enterprises face when they go to deployment?5:10: It’s all about trust and culture: getting employees and executives comfortable with the technology. That implies upskilling and internal conversations.7:09: What is a data strategy for generative AI?7:37: Companies need data governance, which must be more than a well-written policy document.Governance means operationalizing the policy. Once you focus on quality data and abide by governance, you have the foundation for a good future.9:26: How do you measure that you’re delivering ROI? How do you evaluate so that you know your LLM-backed application is ready to go?10:50: ROI—We need to separate R&D. For R, ROI doesn’t work well. But when you cross from R to D and investments scale, you need to think about ROI.12:15: Evaluation—We can measure LLMs today. But what does that mean in the context of the problem you’re solving? AI in autonomous vehicles is different from AI in medical systems.13:58: Companies need to invest in educating the workforce. Upskilling is not just for expertise; it is also for interdisciplinarity. Changing organizational culture means changing the way organizations communicate and partner.15:38: People underestimate the importance of creating a good user experience. Design thinking is needed. Focus on end-user experience and work back from that.16:59: What are some of the most common use cases for AI?17:17: In the back office, you often have a corpus of information customized to your situation. You can build fit-for-purpose chatbots for key support functions. The best lawyers can’t read everything possible in the corpus or keep up with all the regulatory changes coming in.21:11: AI will increase the value of labor investments. It will expedite the L&D curve for new employees. It will improve users’ lives. And AI is getting much better. We’ve only seen the floor, not the ceiling.24:38: Do you have a checklist or a playbook to help companies prioritize use cases?24:57: Companies need to think “What problems do I need to solve?” Think from a problem-centric approach.27:32 Are there best practices for sharing learning across different groups?28:17: We’ve seen centers of excellences rise. Sharing what didn’t work is important. GenAI is very democratizing—not everyone needs a PhD. When companies reward sharing, including what didn’t work, it really engenders collective learning and great ideas.30:15: What have leading companies done to prepare their workforces?30:31: PwC made a major investment in MyAI, which was focused on the ability to get AI into the hands of users, down to entry-level interns. It was an intentional L&D process that was focused on AI. We gave people the tools and a safe space to use them.32:43: It’s learning by doing, and it’s fun. And it can be customized to a company or a firm.33:03: If we didn’t provide a controlled environment, our people would go out into an uncontrolled environment.

Alfred Spector has been a leader in AI and machine learning at Google, IBM, and Two Sigma. He is now a visiting scholar at MIT, an advisor at Blackstone, and coauthor of the text book Data Science in Context. Alfred talks with Ben Lorica about what people developing with AI need to be successful. Succeeding with AI is about more than just a model. We need to think about the application and its context. We need humanities and social sciences in addition to technology. Alfred also discusses the AI skills gap, resistance to adopting AI, “hybrid intelligence,” and the calls to regulate AI.Points of Interest0:00: Intro0:54: What do we need to do to apply generative AI effectively?2:10: Why did you end up writing the book Data Science in Context?3:14: Data science is about more than the model. More than "just get some data and hope."8:22: Ethics alone isn't enough.11:08: Students need a good basis in economics, political science, history, and literature. We have to think more broadly than "which ad gets the most clicks."14:20: There's an AI literacy and skills gap, particularly outside Silicon Valley.15:43: Companies be probing opportunities.16:20: Is there resistance to adopting AI? Fear of displacement or distrust?18:18: Most people think there is more to do than people to do the work.19:21: To what extent are companies trying to come up with an overarching vision for AI?19:51: For some companies, GenAI will be formative. Others need to kick the tires and put together a road map.21:35: Internal applications can be more fault tolerant. Keep employees in the loop; don't be lazy.23:12: Prior to ChatGPT, barrier to entry was higher. AI is now very developer friendly.24:13: What level of data science or ML knowledge should companies have?25:01: There are two categories of expertise; broad perspective on products and services.28:25: It may take a long time to evaluate whether an application can be deployed.29:07: With agents, the stakes are higher.30:07: Hybrid intelligence will be a coalition that includes AI.32:38: Even task-specific agents can break. Agents are fragile. Humans aren't fast but are good at dealing with things we haven't encountered before.33:43: Regulate uses of technology, not technologies.

Andrew Ng is one of the pioneers of modern AI. He was Google Brain’s founding technical lead, Coursera’s founder, Baidu’s Chief Scientist, DeepLearning.ai’s founder, a Professor at Stanford—and much more. Andrew talks with Ben Lorica about scaling AI, agents, the future of open source AI, and openness among AI researchers. Have you experienced an “agentic moment” when you’re surprised and thrilled by AI’s ability to generate a plan and then to enact that plan? You will.Points of interest0:00: Introduction1:00: Advancing AI required scaling up. Better algorithms weren’t the issue.2:57: Just as we needed GPUs and other new hardware for training, we may need new hardware for inference.3:18: People are pushing Data-centric AI forward. Engineering the data is important—maybe even more important than engineering the model.4:41: The idea of agents has been around for a while. What’s new here?6:00: Agentic workflows let AI work iteratively, which yields a huge improvement in performance.8:01: Agent can be used for Robotic Process Automation (RPA), but it’s much bigger than that. We will experience “agentic moments” when we see AI that plans and executes a task without human intervention.10:42: Do you anticipate new Agentic applications that weren’t possible before?12:21: What are the risks of training on copyright-free datasets? Will using copyright-free datasets degrade performance?15:05: AI is a tool; I dispatch it to do things for me. I don’t see it as a different “species.”16:17: How do we know when an application is ready to release? What are best practices for enterprise use?17:18: It’s still very early. We need more work on evaluation. It’s easy to build applications—but when you build an app in a week, it’s hard to spend 10 weeks evaluating it.19:14: A lot of people build an application on one LLM, but won’t switch because evaluation is hard.20:12: Are you concerned that Meta is the only consistent supplier of open source language models?22:10: The cost of training is falling. The decrease in the cost of training means that the ability to train large models will become open to more players.26:15: The AI community seems less open than it was, and more dominated by commercial interests. Is it possible that the next big innovation won’t get published?26:50: We’re starting to see papers about alternatives to transformers. It’s very difficult to keep technical ideas secret for a long time.

Gwendolyn Stripling, author of Low-Code AI, talks about the democratization of AI, the primacy of data, the future of data science, and the coming of agents. It’s easy to think that AI is all about algorithms and models but it’s not; it’s really about understanding the business use case and the data that can be applied to that use case. We’re only beginning to have tools for the rest of the job: collecting, preparing, and exploring the data to find out what’s relevant to your business. Looking ahead, Gwendolyn sees generative AI automating even more of the workload. But focusing on the data, and collecting, understanding, and interpreting it, will always be the human part of the job.Points of interest0:57: What’s the boundary between no-code and low-code?3:10: Using the minimum amount of code necessary to achieve your goal.4:09: Low-code reduces the heavy lifting. But what if you want to learn about AI and ML?6:35: Learning ML isn’t about the tools; it’s about the business case and the data.7:55: What made you think about exposing more people to low-code AI?11:21: The key to all of this is the use case and then the data.14:32: What if I primarily use SQL?15:30: Is there an equivalent of AutoML for data collection and preparation?16:50: Generative AI looks like it will be able to help prepare data.19:22: How did the release of ChatGPT and other LLMs affect your book?24:00: Is there a low-code or no-code approach to RAG?26:30: The GenAI pipeline is becoming completely automated.26:49: The word of 2024 is agents. A lot of what can be automated will be automated.28:00: A lot of people are sharing lessons and best practices. That makes this an exciting time.29:17: Looking ahead five years, what will data scientists and ML Engineers do?

Justin Norman, author of Product Management for AI and co-founder of Vera, a startup focused on security for generative AI, talks with Ben Lorica about how product management has changed since Generative AI came on the scene. He discusses the issues retrieval-augmented generation (RAG) raises for product management; how reliability has become part of a product’s value; how companies that have lagged in their adoption of AI can use generative AI as a way to catch up; and the ability of open source AI in helping smaller companies compete with more established companies.Points of Interest0:00: You wrote Product Management for AI back in 2020 and 2021. How have things changed for product managers since then?3:04: Do companies that lead with operations and infrastructure for traditional AI maintain an advantage with Generative AI? Or does Generative AI allow companies that are just starting to catch up?5:09: Can new companies use open source to compete with established companies? Can open source help capture value as well as larger proprietary models?6:08: What do product managers struggle with when implementing RAG? What's the relationship between fine-tuning and RAG?10:58: RAG gives you value out of the box, but the key to success is how the data is organized.13:57: Are VCs underinvesting in certain parts of the pipeline? There is lots of investment in AI, but not as much investment in startups working on necessary technologies like ETL and data engineering.16:31: Why is reliability important for generative AI? How is generative AI different from other applications that we’re familiar with, and what implications does this have for product management?21:03: Are enterprises realizing that efficiency is important for succeeding with generative AI?23:44: We’re familiar with dashboards for monitoring and managing traditional software products. What would you imagine a dashboard for generative AI models to be? What do you need to be monitoring?28:49: Very few developers working in machine learning have also done frontend development or worked on user experience (UX). However, understanding user interaction can help you to improve your model.30:44: You're working with the father of digital forensics, Hany Farid. Should we be worried about DeepFakes?

Pete Warden, founder of Useful Sensors and co-author of TinyML, discusses use cases for artificial intelligence that we rarely think about: how can you run AI on very small systems? How can you put AI on consumer devices in ways that are actually useful and not just buzzword-compliant? AI doesn’t have to rely on massive GPU farms. Pete talks about what happens when you exchange one set of requirements (extreme power, heat, and expense) for another (minimal size, cost, and heat).Points of Interest00:00: Introductions, including Pete’s introduction to his company.2:22: What are some of the challenges and use cases for sensor-driven AI?4:11: Is sensor-driven AI relevant to industries other than hardware?6:22: Now we’re in the age of foundation models and large language models. Is “large” incompatible with “tiny”? Can you run language models on smaller devices?8:00: Will there be developments in tinyML that will benefit the broader LLM community?9:30: What’s deployable today in computer vision, speech, and language? What can be done with hardware that’s constrained by cost, size, and power consumption?11:15: How will product designers work with sensor-driven AI? Will they simply select from a palette of optional modules?12:37: Pete walks us through the development of AI-in-a-Box, from its conception to its reception.15:31: Your devices don’t have network connections. Without a network connection, how do you update models? Is it necessary?19:00: Do you do Retrieval Augmented Generation (RAG) on your devices?20:35: Our devices have user interfaces that combine voice and presence. A voice interface is central, but visual (and other channels) help to create an awareness of the speaker.21:35: What are some of your specific challenges, like power consumption and latency? How do you make tradeoffs?22:45: What is the future of large language models for sensor-driven AI?26:50: What are some of the security concerns for sensor-driven AI and what are you doing about them?28:22: What is Dark Compute and why is it important?30:48: What are the biggest opportunities for pushing AI into consumer devices? We need to start with problems that users actually care about.32:30: How can listeners connect to the broader movement around TinyML?

O’Reilly’s Generative AI in the Enterprise survey reported that people have trouble coming up with appropriate enterprise use cases for AI. Why is it hard to come up with appropriate use cases?Chip Huyen, cofounder of Claypot AI and author of Designing Machine Learning Systems, talks about why many companies have trouble coming up with appropriate use cases for AI, how to evaluate possible use cases, and the skills your company will need to put them into practice.Points of Interest0:00: Introduction0:49: O’Reilly’s Generative AI in the Enterprise survey report results.3:02: Now that generative AI is more accessible, will it be easier to come up with use cases?4:29: AI is easy to demo but hard to productize. Consistence, risk, and compliance.6:44: Is there a framework or checklist for thinking about applications?8:15: What are some of your favorite use cases?13:30: RAG is the “hello, world” of AI applications.17:24: How do you navigate between the desires and requirements of different stakeholders?19:00: When talking to stakeholders, you have to answer questions at the right level.21:10: How to think about staffing teams for generative AI.22:45: There’s less model development with generative AI, more application development.23:12: Frontend engineers and full-stack developers are very successful.26:27: What are companies’ concerns about risk?27:27: Understanding data gives a lot of clues about what it is good at and should be used for.29:00: The importance of documentation.30:25: Are there specific things you can do to ease the integration of AI into an organization?32:49: What companies that have deployed AI into products stand out?