Podcast Summary: "GenAI Hot Takes and Bad Use Cases"
Practical AI – Feb 24, 2025
Hosts: Chris Benson (Principal AI Research Engineer, Lockheed Martin), Daniel Whitenack (CEO, Prediction Guard)
Overview
In this episode of Practical AI, Chris Benson and Daniel Whitenack dive into the "cold side" of generative AI, focusing on situations where GenAI is overhyped, misapplied, or simply not ready for real-world productivity. They share a series of "hot takes" on misuse cases and discuss why, despite the buzz around generative AI, there are critical limitations – both technical and practical – that make certain applications unreliable or risky. With a candid and insightful approach, they turn their years of practical AI experience into a must-listen cautionary guide for anyone considering deploying GenAI.
Key Discussion Points & Insights
1. The Trouble with Fully Autonomous Agents
[02:00–09:10]
- Hot Take #1: Fully autonomous GenAI agents (with no human in the loop) are, at present, “a source of sadness... when people try to create them.” – Chris Benson [02:36]
- Examples: Sales automation, intern/admin processes, manufacturing plant automation.
- Why They Fail:
- Lack of necessary guardrails.
- High sensitivity, errors, and system fragility.
- Real-world judgments and nuances are beyond current GenAI capabilities.
- Daniel: “[AI is better as a] tool that can help your sales professionals prospect or... create these dossiers... but not this end-to-end, completely autonomous automation.” [07:00]
2. GenAI and Time Series Forecasting – Don’t Do It!
[09:11–13:43]
- Hot Take #2: Using GenAI alone for time series forecasting (e.g., financial trading, predicting future trends) is generally a bad idea.
- Reasoning:
- GenAI lacks real-world grounding and numerical reasoning.
- Makes unreliable or nonsensical predictions.
- Traditional statistical tools (e.g., Facebook/Meta’s Prophet) remain superior.
- Chris: “If you paste in a bunch of time series data and try to create a forecast... with the GenAI model and nothing else... that’s going to end again in sadness for you.” [09:38]
- Alternative Advice: Use GenAI to generate code that leverages specialized forecasting tools, but not for prediction itself.
3. Complete Code Generation via GenAI – Not Production Ready
[15:04–18:05]
- Hot Take #3: Don't rely on GenAI to rewrite entire codebases or develop complete, production-grade applications.
- GenAI can be a good coding assistant but:
- Results are inconsistent, especially for complex or non-mainstream scenarios.
- Lacks understanding required for intricate project details.
- Daniel: “I haven’t gotten anything that I would say is a production grade program fully functional through nothing but generative AI. Just toy programs.” [15:21]
- GenAI is best seen as an assistant or “junior developer,” not as a replacement for skilled software engineers.
4. High Throughput, Real-Time, and Critical Applications
[19:40–23:16]
- Hot Take #4: Avoid GenAI for extremely high-throughput, low-latency, or real-time settings—especially when outcomes are critical (e.g., manufacturing quality checks, autonomous vehicles).
- Limitations:
- Insufficient speed (GenAI models are too slow for split-second decisions).
- Lack reliability in mission-critical or safety scenarios.
- Daniel: “Real time applications with critical outcomes... You may have generative AI as a component... but you’re going to have to have some guardrails around it.” [20:50]
- Traditional computer vision or statistical models are preferred for these tasks.
- Chris: “These models... don’t operate fast enough and they don’t operate in the types of environments you need for these... edge use cases.” [21:20]
5. Multilingual and Culturally Diverse Applications
[24:17–27:43]
- Hot Take #5: GenAI is not ready for use in most of the world’s languages or broad cultural contexts.
- Modern models perform decently in major languages (top 5–10) but poorly across the world’s 7,000+ spoken languages.
- Tooling, scripts, and contextual understanding are severely lacking for non-Western/non-mainstream cases.
- Chris: “It would be great if you could, you know, land anywhere in the world and change your ChatGPT... but I would say generally that’s not the case as of now.” [25:45]
- Ongoing significant bias due to how and where data is collected.
Notable Quotes & Memorable Moments
"Currently and for some time, [autonomous agents] are generally a source of sadness for people when they try to create them."
— Chris Benson [02:36]
"In the imagination, it would be great to think of just letting that run in the background and you getting sales all the time, but it just doesn’t really work very well. There’s a lot of fragility in that type of system.”
— Daniel Whitenack [07:00]
"If you paste in a bunch of time series data and try to create a forecast just with the GenAI model and nothing else, then I think that’s going to end again in sadness for you."
— Chris Benson [09:38]
"I haven’t gotten anything that I would say is a production grade program fully functional through nothing but generative AI. Just toy programs."
— Daniel Whitenack [15:21]
"You don’t want to put a general generative AI model in charge of doing things for which there are no guardrails."
— Daniel Whitenack [27:43]
Practical Alternatives Suggested
- Workflow Automation: Use orchestrators like Prefect for monitored, retriable processes with structured inputs/outputs.
- Time Series Forecasting: Employ statistical packages (e.g., Meta's Prophet) for accurate predictions; GenAI can help write code wrappers.
- Code Generation: Treat GenAI as an assistant, not a full developer (try tools like Devin, Cursor, Windsurf as code pairers).
- Critical Real-Time Applications: Stick with specialized, fast, and reliable models running locally/on edge devices.
- Multilingual AI: Recognize current limitations; consider custom training or alternative NLP pipelines for non-mainstream languages.
Hosts' Final Thoughts
- GenAI is improving rapidly, but users should avoid putting it in charge of high-risk, real-time, or critical tasks without guardrails.
- Use GenAI in collaboration with humans—think “assistant” not “autonomous operator.”
- The gap between AI hype and practical, reliable AI remains wide for non-mainstream, high-stakes, and high-speed applications.
- For tasks outside the “big tech” circles—especially across languages and cultures—expect to do extra work, or consider non-GenAI alternatives.
- Always prioritize guardrails and human oversight, especially where significant outcomes (financial, legal, medical, or safety) are at stake.
Episode takeaway:
Generative AI is a powerful tool, but it’s not a panacea. Know its limits, find productive integrations, and always supplement it with robust processes and human judgment—especially when outcomes matter most.