AI Data Agents with Joe Reis
AgileData Podcast #53
Join Shane Gibson as he chats with Joe Reis on how the potential adoption of GenAI and LLM's in the way Data teams work.
Listen
Listen on all good podcast hosts or over at:
https://agiledata.podbean.com/e/agiledata-53-ai-data-agents-with-joe-reis/
Read
Read the podcast transcript at:
https://agiledata.io/podcast/agiledata-podcast/ai-data-agents-with-joe-reis/#read
Google NoteBookLLM Briefing
Briefing Document: AI Data Agents - AgileData Podcast with Joe Reis
Date: 29 June 2024 Source: AgileData Podcast - "AI Data Agents with Joe Reis" Presenters: Shane Gibson, Joe Reis
Overall Summary
This podcast episode delves into the current state of Generative AI (GenAI), particularly Large Language Models (LLMs), within the data domain. Joe Reis offers his typically opinionated and pragmatic take, cutting through the hype to discuss real use cases, pitfalls, and future directions. He argues that the market is currently experiencing two camps; one believes AI is a truly transformative technology and the other that it is a box ticking exercise driven by FOMO. The conversation explores different levels of AI integration, from "Ask AI" chatbots, to "Assisted AI" co-pilots, and the aspirational "Automated AI" systems, acknowledging the unique dangers, value and potential of each. A key theme throughout is that while LLMs hold immense promise, their non-deterministic nature and tendency to "hallucinate" poses significant risks, particularly when applied to data analysis and decision-making. The pair express a genuine excitement for the technology, alongside a level of caution.
Key Themes and Ideas
The Hype Cycle and Two Camps:
Observation: "If you didn’t know any better, you would think that every company and vendor and open source project is an AI company or project these days."
Two Camps:Transformative: Those who believe GenAI is a game-changing technology that requires immediate adaptation.
FOMO: Those who are implementing GenAI because it's expected, rather than based on a genuine need or understanding. Many in this group view GenAI as "just a toy" but feel compelled to participate.
Hype Cycle Comparison: Joe sees this following similar hype cycles to previous tech like OLAP cubes, Tableau, Big Data, and Data Mesh. Shane uses the analogy of the "Hadoop" hype to see if this tech has "stickiness".
Consumer vs Commercial: There is still a lot of "imagining" with consumers, whereas commercial businesses are now testing things with POC's before rubber hits the road next year,
Use Cases: Ask, Assisted, and Automated AI
Ask AI (Chatbots): This is the most common application, with many vendors offering text-to-SQL capabilities.
Table Stakes: "Everyone’s got a co pilot, for example. Everyone’s got, some sort of LLM search." These features are becoming table stakes, losing their uniqueness.
Hallucination Risks: Joe and Shane are both concerned about the use of LLMs making up information and then answering questions using this info.
Assisted AI (Co-pilots): These systems augment human work by providing coding help or suggesting data quality rules, and the human stays in the loop.
Value: Co-pilots are seen as productivity boosters, but Shane raises the issue of the quality of the prompt ( the instructions used) as being an area to focus on.
Prompt Reviews: There is a need to include a way of tracking prompts when the co-pilot generates code.
Automated AI (Autonomous): The ideal scenario where AI takes action without human intervention.
Examples: Anomaly detection, triggering actions based on insights (like customer retention campaigns, fixing a data quality problem autonomously). This was not often part of traditional Machine Learning.
Real-Time Requirements: For "live data" environments and real-time data integration, automated AI is necessary because "things are happening at such a speed that humans can’t, Respond".
LLM Limitations and Risks:
Non-Determinism: "You can ask it the same question five times, and you will get variability of answers." This unreliability is problematic for data analysis where consistent results are required.
Hallucination: LLMs generate incorrect information and are particularly inaccurate with text-to-SQL tasks. As Shane says, "the machine is making shit up about our data and then using it to answer our questions".
Trust: There's a question of whether we should trust an LLM to the same degree as we trust an analyst and there are many human checks and balances with analysts that are hard to do with machines.
Dependency on Vendors: Building on top of a platform like OpenAI risks the platform provider launching their own version that kills that market and competitive moat.
"Black Box" Functionality: LLM's are a black box that no one truly understands, not even their creators.
Emerging Trends and Opportunities:
Multimodal Capabilities: The ability of AI to process images, audio, and video as well as text, is going to open opportunities for new capabilities. Shane and Joe talk about using images of reports to diagnose issues with data and look at trends.
Agentic Workflows: Joe highlights interest in "agentic workflows," where multiple AI agents work together to complete tasks (e.g., code writing, testing, and deployment). This may impact software development, in that the cost of developing software will drop to near zero.
Micro Use Cases: Rather than broad applications of AI, there is potential for AI to solve very specific micro problems with clear constraints. This allows for much better prompts.
Data Governance: Data professionals may become more like BA's with good data skills, that are also able to bridge the gap with good communication and prompting.
AI Empowered Businesses: There is a strong desire from a lot of businesses to experiment with AI powered businesses to reduce costs and complexity, and streamline processes.
The Future of Data Teams:
Potential Displacement: There is a dystopian view of LLMs where data teams become redundant as businesses rely on LLMs to self-serve answers. Shane predicts that stakeholders may prefer to use self-serve options rather than rely on data teams with long turnaround times, even if the data is only "good enough".
Evolving Roles: Joe suggests that data teams may evolve into domain experts with good communication and prompting skills rather than being heavily focused on technical aspects.
Data Management at Scale: AI and ML may be the answer to solving the decades-old data management problems, because AI can operate at a scale that humans cannot.
Quotes and Key Takeaways
Joe: "I use [chat GPT] every day...do I use it more as a conversational agent to poke holes in my ideas? I don’t use it to write."
Shane: "I can see using LLMs for actions...But that reinforcement model where the LLM is making shit up about our data and then using it to answer our questions, I still haven’t got my head around how dangerous that is."
Joe: "Would you stick an LLM on top of your corporate data set today? And the second part of that question is, would you bet your job that it’s going to work awesomely? Nobody raises their hand."
Shane: "I’m just going to focus on small problems where it can take time... and let me do them in a minute or 10 minutes or teach me something I didn’t know."
Joe: "This is probably how a lot of the data management problems that have plagued the business world for ages get solved because it’s going to operate at such a scale that it would have to work, I would say."
Shane: "…if there’s something that takes a team an hour and the AI data agent makes it 10 minutes and doesn’t increase your risk, why wouldn’t you invest in that?"
Joe: "Not everyone’s an expert in this stuff. I would say not a lot of people are… Go with people who have, experience in this field who have done the work."
Shane "Make them do it because I guarantee you ask me to do that. I’m gonna go, I’ll do it, but I don’t trust anything I get back. "
Actionable Points
Be Wary of Hype: Approach GenAI with a critical eye, understanding that not every solution is genuinely transformative.
Focus on Use Cases: Experiment with specific applications of GenAI (assisted AI), prioritising quick wins and clear ROI, rather than broadly implementing AI solutions.
Address Risks: Acknowledge the dangers of non-determinism and hallucination in LLMs and implement checks and balances with human reviews.
Investigate Automated AI: Explore how automated systems can reduce manual effort and improve decision-making with real time feedback, rather than just generating reports.
Develop Domain Expertise: Data professionals should invest in understanding the business context and refining their communication and prompting skills.
Concluding Thoughts While Generative AI offers significant potential for data-related tasks, it should be approached with pragmatism and a clear understanding of its current limitations. It is another tool in the toolkit. The key will be to adopt small experiments and keep an open mind about the technology without being caught up in the hype. Joe and Shane plan to review the progress of AI in data in another year, which will be a good opportunity to see what's changed.