639 - How to Stop ChatGPT from Lying to You (AI PhD Reveals the Fix) with Garima Agrawal
🎯 Summary
Podcast Episode Summary: 639 - How to Stop ChatGPT from Lying to You (AI PhD Reveals the Fix) with Garima Agrawal
This episode of the Paul Higgins podcast features Dr. Garima Agrawal, an AI PhD and Head of AI at Minerva, who addresses the critical challenges consultants and businesses face when implementing Large Language Models (LLMs), primarily focusing on eliminating hallucinations and moving toward effective, goal-driven AI integration.
1. Focus Area
The discussion centers on Practical LLM Implementation and Mitigation of Hallucinations. Specific topics include advanced prompting techniques (layered prompting), strategic LLM selection for different tasks (e.g., analysis vs. coding), the distinction between interactive and agentic AI, and structuring context delivery (document uploads vs. verbal input).
2. Key Technical Insights
- Layered Prompting for Hallucination Reduction: Instead of single, blanket prompts, users should employ a multi-step, iterative approach. Start by testing the LLM’s baseline knowledge of a domain, then gradually introduce specific context about the user’s problem, and finally, ask the model to confirm its understanding before proceeding to solutions.
- LLM Specialization by Task: Different LLMs excel in different areas based on their training data. ChatGPT is recommended for broad analysis and natural language tasks (summarization, rephrasing). Claude is preferred for coding and technical tasks due to its strong intrinsic models (specifically Claude 3.5 Sonnet). Gemini is useful for technical research and gathering information about tools and libraries because of Google’s extensive training data.
- Agentic vs. Interactive AI: Interactive AI (like standard ChatGPT use) is static and reactive. Agentic AI is goal-driven, capable of taking actions, sensing environmental changes, and making decisions without constant human oversight (e.g., automatically sending reminders or preparing interview profiles based on real-time context).
3. Business/Investment Angle
- Consulting Bottlenecks: Many AI/tech consultants are stuck in the “LLM hamster wheel,” relying on basic prompting and getting unreliable results, hindering scalable growth for themselves and their clients.
- Investor Guidance: Dr. Agrawal advises investors on which AI ventures are sound, leveraging her technical depth to bridge the gap between complex AI development and commercial viability.
- Strategic Implementation: The focus for startups should be on implementing AI safely and effectively, reducing hallucinations to build trust in the solutions being deployed.
4. Notable Companies/People
- Garima Agrawal (Guest): AI PhD, Head of AI at Minerva, with 15 years of industry experience prior to her PhD, specializing in LLM research, RAG, and knowledge graphs.
- Paul Higgins (Host): Consultant specializing in helping other consultants build scalable businesses that don’t rely solely on the founder.
- Minerva: The company where Dr. Agrawal is Head of AI, currently building an agentic AI platform, particularly for handling customer service contacts.
- LLM Providers Mentioned: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Perplexity (for research).
5. Future Implications
The industry is moving rapidly from simple interactive AI (asking questions) toward sophisticated agentic AI systems capable of executing complex, goal-driven workflows autonomously. Success in AI implementation will depend on understanding the underlying mechanics and selecting the right model for the specific job, rather than relying on a single general-purpose tool.
6. Target Audience
This episode is highly valuable for AI/Tech Consultants, CTOs, Startup Founders, and Investors who are past the initial hype phase and need actionable, technically grounded strategies to move from basic LLM usage to robust, reliable, and scalable AI integration.
Comprehensive Summary Narrative
The podcast episode opens by framing the common frustration among consultants: using ChatGPT yields inconsistent, often hallucinated results, preventing true productivity gains. Dr. Garima Agrawal, bringing a unique perspective from her recent PhD in AI combined with 15 years of industry experience, steps in to provide the necessary technical grounding.
Dr. Agrawal emphasizes that LLMs are powerful, general-purpose tools, but relying on them blindly leads to failure. Her core solution for mitigating hallucinations is layered prompting—an iterative process where the user first probes the model’s existing knowledge, then provides context in small “bytes,” and continuously verifies the model’s understanding step-by-step. She warns that once an LLM begins to hallucinate, it creates a “snowball effect” that is difficult to correct, necessitating a restart.
A significant portion of the discussion focuses on model selection. Dr. Agrawal advises against using one model for everything. She recommends ChatGPT for general analysis and natural language manipulation, Claude for coding and technical tasks (citing its superior intrinsic models), and Gemini for gathering broad technical research, noting that ChatGPT’s tendency to fabricate citations makes it unreliable for serious research. For deep, cited research, Perplexity is suggested.
The conversation also touches on technical nuances like prompting structure. While JSON prompting is popular, Dr. Agrawal clarifies it is most beneficial when interacting via API for structured data consumption, not necessarily when using the standard UI. For UI interactions, providing context via attachments (like transcripts) is helpful, but feeding too much data at once can still trigger errors. She also introduces the concept of few-shot prompting (providing examples) as a powerful way to guide the model toward desired output formats, such as creating specific LinkedIn posts versus general content.
Finally, the episode tackles the next frontier: **agentic
🏢 Companies Mentioned
đź’¬ Key Insights
"Brilliant execution, terrible strategy. You can solve any kind of problem, but you'll be stuck in the same revenue patterns, same deal sizes, same time-for-money trap."
"I have a paper on this, Mindful RAG, where I illustrate on what could be the failure points of AI, and how I've figured that by analyzing the error logs through the AI."
"totally relying on AI—these are the two bottlenecks I see mostly. And exploring AI, but have your own thinking going on. You have to be very, very mindful."
"from the conversation in real-time, we are able to take the context, and the AI takes the decision to create a question and understand that the customer has asked some question. Let me go back, do the RAG, get the answer for the agent, keep it ready before he has to answer, right?"
"there is an AI which is goal-driven. It knows these are the tasks. It can act at a certain point, able to take a decision, able to get the context from the environment that something has changed, I need an action, right?"
"So, he knows that the customer is calling, and from the conversation in real-time, we are able to take the context, and the AI takes the decision to create a question and understand that the customer has asked some question. Let me go back, do the RAG, get the answer for the agent, keep it ready before he has to answer, right?"