Generative AI in the Real World

In this episode, Ben Lorica and Hamel Husain talk about how to take the next steps with artificial intelligence. Developers don’t need to build their own models—but they do need basic data skills. It’s important to look at your data, to discover your model’s weaknesses, and to use that information to develop test suites and evals that show whether your model is behaving well.Links to ResourcesHamel's upcoming course on evaluating LLMs.Hamel's O'Reilly publications: “AI Essentials for Tech Executives” and “What We Learned from a Year of Building with LLMs”Hamel's website.Points of Interest0:39: What inspired you and your coauthors to create a series on practical uses of foundation models? What gaps in existing resources did you aim to address?0:56: We’re publishing “AI Essentials for Tech Executives”¹ now; last year, we published “What We Learned from a Year of Building with LLMs.”² Coming from the perspective of a machine learning engineer or data scientist—you don’t need to build or train models. You can use an API. But there are skills and practices from data science that are crucial.2:16: There are core skills around data analysis and error analysis and basic data literacy that you need to get beyond a demo.2:43: What are some crucial shifts in mindset that you’ve written about on your blog?3:24: The phrase we keep repeating is “look at your data." What does “look at your data" mean?3:51: There’s a process that you should use. Machine learning systems have a lot in common with modern AI. How do you test those? Debug them? Improve them? Look at your data; people fail on this. They do vibe checks, but they don’t really know what to do next.4:56: Looking at your data helps ground everything. Look at actual logs of user interactions. If you don’t have users, generate interactions synthetically. See how your AI is behaving and write detailed notes about failure modes. Do some analysis on those notes: Categorize them. You’ll start to see patterns and your biggest failure modes. This will give you a sense of what to prioritize.6:08: A lot of people are missing that. People aren’t familiar with the rich ecosystem of data tools, so they get stuck. We know that it’s crucial to sample some data and look at it.7:08: It’s also important that you have the domain expert do it with the engineers. On a lot of teams, the domain expert isn’t an engineer.7:44: Another thing is focusing on processes, not tools. Tools aren’t the problem—the problem is that your AI isn’t working. The tools won’t take care of it for you. There’s a process: how to debug, look at, and measure AI. Those are the main mind shifts.9:32: Most people aren’t building models (pretraining); they might be doing posttraining on a base model. But there are a lot of experiments that you still have to run. There’[re] knobs you have to turn, and without the ability to do it systematically and measure, you’re just mindless[ly] turning knobs without learning much.10:29: I’ve held open office hours for people to ask questions about evals. What people ask most is what to eval. There are many components. You can’t and shouldn’t test everything. You should be grounded in your actual failure modes. Prioritize your tests on that.11:30: Another topic is what I call the prototype purgatory. A lot of people have great demos. The demos work, and might even be deployable. But people struggle with pulling the trigger.12:15: A lot of people don’t know how to evaluate their AI systems if they don’t have any users. One way to help yourself is to generate synthetic data. Have an LLM generate realistic user inputs and brainstorm different personas and scenarios. That bootstraps you significantly towards production.13:57: There’s a new open source tool that does something like this for agents. It’s called IntelAgent. It generates synthetic data that you might not come up with yourself.

Join Shelby Heinecke, senior research manager at Salesforce, and Ben Lorica as they talk about agents, AI models that can take action on behalf of their users. Are they the future—or at least the hot topic for the coming year? Where are we with smaller models? And what do we need to improve the agent stack? How do you evaluate the performance of models and agents?About the Generative AI in the Real World podcast: In 2023, ChatGPT put AI on everyone’s agenda. In 2025, the challenge will be turning those agendas into reality. In Generative AI in the Real World, Ben Lorica interviews leaders who are building with AI. Learn from their experience to help put AI to work in your enterprise.Points of Interest0:29: Introduction—Our guest is Shelby Heinecke, senior research manager at Salesforce.0:43: The hot topic of the year is agents. Agents are increasingly capable of GUI-based interactions. Is this my imagination?1:20: The research community has made tremendous progress to make this happen. We’ve made progress on function calling. We’ve trained LLMs to call the correct functions to perform tasks like sending emails. My team has built large action models that, given a task, write a plan and the API calls to execute that. This is one piece. A second piece is when you don’t know the functions a priori, giving the agent the ability to reason about images and video.3:07: We released multimodal action models. They take an image and text and produce API calls. That makes navigating GUIs a reality.3:34: A lot of knowledge work relies on GUI interactions. Is this just robotic process automation rebranded?4:05: We’ve been automating forever. What’s special is that automation is driven by LLMs, and that combination is particularly powerful.4:32: The earlier generation of RPA was very tightly scripted. With multimodal models that can see the screen, they can really understand what’s happening. Now we’re beginning to see reasoning enhanced models. Inference scaling will be important.5:52: Multimodality and reasoning-enhanced models will make agents even more powerful.6:00: I’m very interested in how much reasoning we can pack into a smaller model. Just this week DeepSeek also released smaller distilled versions.7:08: Every month the capability of smaller models has been pushed. Smaller models right now may not compare to large models. But this year, we can push the boundaries.7:38: What’s missing from the agent stack? You have the model—some notion of memory. You have tools that the agent can call. There are agent frameworks. You need monitoring, observability. Everything depends on the model’s capabilities: There’s a lot of fragmentation, and the vocabulary is still unclear. Where do agents usually fall short?9:00: There’s a lot of room for improvement with function calling and multistep function calling. Earlier in the year, it was just single step. Now there’s multistep. That expands our horizons.9:59: We need to think about deploying agents that solve complex tasks that take multiple steps. We will need to think more about efficiency and latency. With increased reasoning abilities, latency increases.10:45: This year, we’ll see small language models and agents come together.10:58: At the end of the day, this is an empirical discipline and you need to come up with your own benchmarks and eval tools. What are you doing in terms of benchmarks and eval?11:36: This is the most critical piece of applied research. You’re deploying models for a purpose. You still need an evaluation set for that use case. As we work with a variety of products, we cocreate evaluation sets with our partners.12:38: We’ve released the CRM benchmark. It’s open. We’ve created CRM-style datasets with CRM-type tasks. You can see the open source models and small models on these leaderboards and how they perform.13:16: How big do these datasets have to be?

How do we measure skills in an age of AI? That question has an effect on everything from hiring to productive teamwork. Join Kian Katanforoosh, founder and CEO of Workera, and Ben Lorica for a discussion of how we can use AI to assess skills more effectively. How do we get beyond pass/fail exams to true measures of a person’s ability?Points of Interest0:28: Can you give a sense of how big the market for skills verification is?0:42: It’s extremely large. Anything that touches skills data is on the rise. When you extrapolate university admissions to someone’s career, you realize that there are many times when they need to validate their skills.1:59: Roughly what’s the breakdown between B2B and B2C?2:04: Workera is exclusively B2B and federal. However, there are also assessments focused on B2C. Workera has free assessments for consumers.3:00: Five years ago, there were tech companies working on skill assessment. What were prior solutions before the rise of generative AI?3:27: Historically, assessments have been used for summative purposes. Pass/fail, high stakes, the goal is to admit or reject you. We provided the use of assessments for people to know where they stand, compare themselves to the market, and decide what to study next. That takes different technology.4:50: Generative AI became much more prominent with the rise of ChatGPT. What changed?5:09: Skills change faster than ever. You need to update skills much more frequently. The half-life of skills used to be over 10 years. Today, it’s estimated to be around 2.5 years in the digital area. Writing a quiz is easy. Writing a good assessment is extremely hard. Validity is a concept showing that what you intend to measure is what you are measuring. AI can help.6:39: AI can help with modeling the competencies you want to measure.6:57: AI can help streamline the creation of an assessment.7:22: AI can help test the assessment with synthetic users.7:42: AI can help with monitoring postassessment. There are a lot of things that can go wrong.8:25: Five years ago in program, people used tests to filter people out. That has changed; people will use coding assistants on the job. Why shouldn’t I be able to use a coding assistant when I’m doing an assessment?9:16: You should be able to use it. The assessment has to change. The previous generation of assessments focused on syntax. Do you care if you forgot a semicolon? Assessments should focus on other cognitive levels, such as analyzing and synthesizing information.10:06: Because of generative models, it’s become easier to build an impressive prototype. Evaluation is the hard point. Assessment is all about evaluation, so the bar is much higher for you.10:48: Absolutely. We have a study that calculates the number of skills needed to prototype versus deploy AI. You need about 1,000 skills to prototype AI. You need about 10,000 skills for production AI.12:39: If I want to do skills assessment on an unfamiliar workflow, say full stack web development, what’s your process for onboarding?13:17: We have one agent that’s responsible for competency modeling. You can have a subject-matter expert (SME) share a job description or task analysis or job architecture. We take that information and granularize the tasks worth measuring. At that point, there’s a human in the loop.14:27: Where does AI help? What does the AI need? What would you like to see from people using your tool?15:04: Language models have been trained on pretty much everything online. You can get a pretty good answer from AI. The SME takes that from 80% to 100%. Now, there are issues with that process. We separate the core catalog of skills from the custom catalog, where customers create custom assessments. A standardized assessment lets you benchmark against other people or companies.16:32: If you take a custom assessment, it’s highly relevant to your needs, even though comparisons aren’t possible.16:41: It’s obviously anonymized, right?

Chloé Messdaghi and Ben Lorica discuss AI security—a subject of increasing importance as AI-driven applications roll out into the real world. There’s a knowledge gap: Security workers don’t understand AI, and AI developers don’t understand security. It’s important to be aware of all the resources that are available. Make sure to bring everyone together to develop AI security policies and playbooks, including AI developers and experts. Be aware of all the resources that are available; we expect to see AI security certifications and training becoming available in the coming year.Points of Interest0:24: How does AI security differ from traditional cybersecurity?0:44: AI is a black box: We don’t have transparency to show how AI works or explainability to show how it makes decisions. Black boxes are hard to secure.2:12: There’s a huge knowledge gap. Companies aren’t doing what is needed.2:24: When you talk to executives, do you distinguish between traditional AI and ML and the new generative AI models?2:43: We talk about older models as well. But security is as much about, What am I supposed to do? We’ve had AI for a while, but for some time, security has not been part of that conversation.3:26: Where do security folks go to learn how to secure AI? There are no certifications. We’re playing a massive catchup game.3:53: What’s the state of awareness about incident response strategies for AI?4:15: Even in traditional cybersecurity, we’ve always had an issue of making sure incident response plans aren’t ad hoc or expired. A lot of it is being aware of all the technologies and products that the company has been using. It’s hard to protect if you don’t know everything in your environment.5:19: The AI Threat Landscape report found that 77% of the companies reported breaches in their AI systems.5:40: Last year, a statistic came out about the adoption of AI-related cybersecurity measures. For North America, 70% of the organizations said they did one or two out of five security measures. 24% adopted two to four measures.6:35: What are some of the first things I should be thinking about to update my incident response playbook?6:51: Make sure you have all the right people in the room. We still have issues with department silos. CISOs can be dismissed or not even in the room when it comes to decisions. There are concerns about restricting innovation or product launch dates. You have to have CTOs, data scientists, ML developers, and all the right people to ensure that there is safety and that everyone has taken precautions.7:48: For companies with a mature cybersecurity incident playbook that they want to update for AI, what AI brings is that you have to include more people.8:17: You have to realize that there’s an AI knowledge gap, and that there’s insufficient security training for data scientists. Security folks don’t know where to turn for education. There aren’t a lot of courses or programs out there. We’ll see a lot of that develop this year.10:13: You’d think we’d have addressed communications silos by now, but AI has ripped the bandaids off. There are resources out there. I recommend Databricks’ AI Security Framework (DASF); it’s mapped to the MITRE ATLAS. Also be familiar with the NIST Risk Framework and the OWASP AI Exchange.11:40: This knowledge gap is on both sides. What are some of the best practices for addressing this two-sided knowledge gap?12:20: Be honest about where your company stands. Where are we right now? Are we doing a good job of governance? Am I doing a good enough job as a leader? Is there something I don’t know about the environment? Be the leader who’s a bridge, breaks down silos, knows who owns what, and who’s responsible for what.13:24: One issue is the notion of shadow AI. Knowledge workers go home and use things that aren’t sanctioned by companies. Are there specific things that companies should be doing about shadow AI?

Join Ben Lorica and Tom Smoker for a discussion of GraphRAG, one of the hottest topics of the last few months. GraphRAG goes a step beyond RAG to make the output of language models more consistent, accurate, and explainable. But what is a graph? A graph is a way of structuring data. In the end, it’s the structure that’s important, along with the work you do to create that structure.Points of Interest0:15: GraphRAG is RAG with a knowledge graph. Do you have a more strict definition?1:00: A lot of what I do is the R in RAG: retrieve. Retrieval is better if you have structured data. I’ve yet to find a definition for GraphRAG. You want to bring in structured data.2:03: At the end of the day, the lesson is structure. Sometimes structure is a SQL database. Don’t lose hope if you don’t have a knowledge graph.2:49: A knowledge graph is a knowledge base and a list of axioms (rules). The knowledge base is just a word connected to another word through a third word. Fundamentally, the benefit comes from the list of triples. The value is in having extracted and defined those triples.4:01: Knowledge graphs are cool again. What are your two favorite examples of GraphRag in production?4:57: My examples are people who are structuring their data so that it’s consistent. Then you can bring it into a context window and do something with it.5:18: LinkedIn and Pinterest are the best examples of existing graph structures that work.5:35: A new application is a veterinary radiology example. Without GraphRAG, the LLM kept recommending conditions specific to Labradors not bulldogs. GraphRAG controlled the problem.6:37: The underlying data was almost exclusively text. It’s difficult to build up a consistent dataset for veterinary radiology because animals move.7:12: My favorite examples: Google uses their data commons to build a Q&A application. Metaphor Data: The starting point is structured data, then they create a second graph from the first graph that maps technical terms to business terms. Then they construct a social graph based on who is using the data.9:41: Structured data can be the basis for a graph.10:06: Unstructured data is valuable, but you need a way to navigate and categorize unstructured data.11:04: Where are we on GraphRAG? Do you still have to explain what GraphRAG is?11:28: More people know about it, but I have to explain it more than I did previously. Exactly what are we referring to? Most people want accuracy in the beginning; the value is often that it is more explainable. People may have seen a fantastic example, but what they haven’t seen is the iterative process in schema design. The upfront cost of these systems is nontrivial.13:13: What are the key bottlenecks? How do I get a knowledge graph?13:23: The biggest question is: Do you need a graph in the first place? There’s a whole spectrum. It’s in most people's interest to stop before they get to the end.14:01: For people who come to us brand-new, we say, “You should try vector RAG first. If that doesn’t work, there’s a lot of good that structuring data can provide.”15:01: If the chunks are structured, and a lot of the work is done up front, then it’s possible to navigate through structured information. At that point, you get value out of vector RAG. Academic papers have to follow a certain structure. If you spend time making sure you know what the chunks are, where they’re split and why, and they’re labeled, you can get a lot of value.16:43: What are some of your pointers about how to get started?16:47: The knowledge base is often a compressed representation. That means less tokens. That means better rate limits and less cost. So some people want a graph to help scale. That’s one start. Another is the desire for a system to be explainable. Getting that information into a structured representation and tracing back that structured representation can be very useful.

Robert Nishihara is one of the creators of Ray and cofounder of Anyscale, a platform for high-performance distributed data analysis and artificial intelligence. Ben Lorica and Robert discuss the need for data for the next generation of AI, which will be multimodal. What kinds of data will we need to develop models for video and multimodal data? And what kinds of tools will we use to prepare that data?Points of Interest1:06: Are we running out of data?1:35: There is a paradigm shift in how ML is thinking about AI. The innovation is on the data side: finding data, evaluating sources of data, curating data, creating synthetic data, filtering low-quality data. People are curating and processing data using AI. Filtering out low-quality data or unimportant image data is an AI task.5:02: A lot of the tools were aimed at warehouses and lakehouses. Now we increasingly have more unstructured multimodal data. What's the challenge for tooling?5:44: Lots of companies have lots of data. They get value out of data by running SQL queries on structured data, but structured data is limited. The real insight is in unstructured data, which will be analyzed using AI. Data will shift from SQL-centric to AI-centric. And tooling for multimodal data processing is almost nonexistent.8:23: In part of the pipeline, you might be able to use CPUs instead of GPUs.8:44: Data processing is not just running inference with an LLM. You might want to decompress video, re-encode video, find scene changes, transcribe, or classify. Some stages will be GPU bound, some will be memory bound, some will be CPU bound. You will want to be able to aggregate these different resources.10:03: Most likely, with this kind of data, it's assumed you will have to go distributed and scale out. There is no choice but to scale the computation.10:46: In the past, we were only using structured data. Now we have multimodal data. We are only scratching the surface of what we can do with video—so people weren't collecting it as much. We will now collect more data.11:41: We need to enable training on 100 times more data.12:43: ML infrastructure teams are now on the critical path.13:52: Companies at the cutting edge have been doing this, but nearly every company has its own data about its specific business that they can use to improve their platform. The value is there. The challenge is the tooling and the infrastructure.15:15: There's another interesting angle around data and scale: experimentation. You will have to run experiments. Data processing and experimentation is part of experimentation.16:18: Customization isn't just at the level of the model. There are decisions to be made at every stage of the pipeline. What to collect, how to chunk, how to embed, how to do retrieval, what model to use, what data to use to fine tune—there are so many decisions to make. To iterate quickly, you need to try different choices and evaluate how they work. Companies should overinvest in evals early.17:29: If you don't have the right foundation, these experiments will be impossible.18:23: What's the next data type to get popular?18:42: Image data will be ubiquitous. People will do a lot with PDFs. Video will be the most challenging. Video combines images and audio; text can be in video too. But the data size is enormous. There are modeling challenges around video understanding. There's so much information in video that isn't being mined.22:50: Companies aren't saying that scaling laws are over, but scaling is slowing down. What's happening?

In this episode, Ben Lorica talks with Claire Vo, chief product officer at Launch Darkly and founder of ChatPRD. AI gives us a new set of tools that make everyone more productive and efficient. Those tools will allow more experimentation; they will allow more people to participate in product development; and they will create new opportunities for startups. As Claire says, this new tooling lets everyone get more ambitious—and if you start now, you’re on the leading edge. Lean in to the opportunities.Points of Interest0:25: ChatPRD is an AI copilot for product managers and people who build products. The goal is to make more efficient people who need to generate ideas, build our requirements.1:15: It improves the quality of product work: it’s an on-demand coach or colleague.2:05: In a hybrid world, there needs to be some kind of artifact describing what we want to build. No matter the culture, you should try to make high-quality documents to improve the thinking.3:44: We ingest your product documents for two reasons: to have context of what you’ve built, what matter, and to inform style and quality.5:13: To become a 100x PM you need to embrace tools and accelerate your work. It’s learning how to scale and do your best in a highly efficient way. Getting 2–3 days back in your week.7:17: Will the programming language of the future be natural language? You will still have to think and describe things as a software engineer or a product manager.7:54: My favorite users are engineers who don’t have product managers, sales people who get customer requests, and even founders who can’t afford a product manager.8:41: In frontier models, I’d like to see up-to-date training data. The killer feature is performance. The models need to support a workflow that requires speed. Models need more control over output mechanisms than they have now, so users don’t have to massage output.10:38: There isn’t capability parity between the models, so you have to make trade-offs between performance, features, API support, latency, user experience, and streaming.11:05: Always design your application to be model agnostic. LaunchDarkly allows engineers to decouple the configuration and release of their code from deploying in production.12:14: With AI, prompts become feature flags. You can measure things like latency and token count, and make informed decisions about what works best.13:21: It’s important to have the ability to experiment in classic software development. That matters even more with nondeterministic software, because the ability to predict output goes down. You need to think about instrumentation from the beginning.14:37: I have been through a couple of technology waves, but this one has stopped me in my tracks. The difference between what is possible and what is not possible is unbelievable. I could have built the product from my startup 10 years ago before lunchtime.16:01: People need to prepare to be expected to do more because the ability to do more is powered by these tools and automations. People should educate themselves on how to automate tasks in their current job, and they should add additional skills like the ability to code.16:42: The shape of organizations will change. The triad of the product manager, engineering lead, and design lead will collapse into an individual. Individual contributors will become more efficient.17:35: Everyone can get more ambitious. There won’t be less to do. More people will be empowered to do more things and have bigger impact.18:44: Everything requires a radical cultural shift inside companies. It can feel scary. You need to set the aspiration and why it matters; you need to organize among motivated individuals and reward the behavior you want to see; new organizations will fall out of the centers of gravity around people who are operating in an AI-native way.

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.