StarZero All shows Try the podcast clipper →

Sep 27, 2023 · 50m · mad

Guardrails AI: The Playbook for Safer, Hallucination-Free LLMs — Shreya Rajpal Explains

Shreya Rajpal · 39m spoken Matt Turck · 8m spoken
0:00 / 0:00
▶ Watch on YouTube →

gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions

Shreya Rajpal, Co-Founder & CEO of Guardrails AI, joins host Matt Turck on The MAD Podcast to discuss strategies for deploying generative AI safely in enterprise environments. She details how runtime verification, modular validation guards, and open-source tooling mitigate hallucinations, compliance risks, and operational failures in mission-critical applications.

How this conversation actually went

Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 17.1% of the talking time here. How this is scored →

Matt as informed peer 3.6 Guest teaching 4.3 Guest disagreement 0.6 Matt pushing back 0.9
05100:0015:0030:0045:000:14–5:48 · Matt as informed peer 2/10 Shreya Rajpal's Background & The Origin of Guardrails AI Matt opens with a supportive question connecting Shreya's autonomous vehicle background to Guardrails AI. Shreya takes the floor to explain how building ML applications is more akin to self-driving systems than traditional predictive ML models.5:48–9:21 · Matt as informed peer 3/10 The Spectrum of Generative AI Failure Modes Matt prompts Shreya to define the broader spectrum of GenAI risks beyond hallucinations. Shreya educates him on compliance failures, brand reputation risks, and privacy leakage, while Matt interjects with friendly banter.9:21–23:27 · Matt as informed peer 5/10 Evaluating Mitigation Techniques: RAG vs. Fine-Tuning Matt asks technical questions about RAG versus fine-tuning and pushes for concrete hallucination reduction metrics. Shreya reframes expectations by explaining that LLMs remain next-token predictors regardless of fine-tuning, offering a nuanced perspective on open-source versus proprietary benchmarks.23:27–35:23 · Matt as informed peer 5/10 Guardrails AI Framework: Guards, Validators, and Provenance Matt presses Shreya on the paradox of using probabilistic machine learning to correct other machine learning models in order to reach reliability. Shreya clarifies how ensembling diverse validators creates overlapping safeguards similar to a stack of sieves.35:23–40:22 · Matt as informed peer 4/10 Managing Failure Modes, Execution Policies, and Latency Matt asks how failure modes are handled practically and how latency is impacted at runtime. Shreya details self-healing re-asking loops, programmatic fallback policies, and parallel execution trade-offs.40:22–49:13 · Matt as informed peer 5/10 Open-Source Strategy and the AI Safety Ecosystem Matt demonstrates strong industry awareness by referencing key companies and ecosystem leaders such as LangChain, Weights & Biases, and LlamaIndex. Shreya elaborates on academic and open-source contributions from Stanford Hazy Research, Microsoft RAIL, and UW.49:13–50:49 · Matt as informed peer 1/10 Connecting with Guardrails AI and Career Opportunities A brief and friendly wrap-up covering open roles, social handles, and contact information, marked by joking about referral fees.0:14–5:48 · Guest teaching 4/10 Shreya Rajpal's Background & The Origin of Guardrails AI Matt opens with a supportive question connecting Shreya's autonomous vehicle background to Guardrails AI. Shreya takes the floor to explain how building ML applications is more akin to self-driving systems than traditional predictive ML models.5:48–9:21 · Guest teaching 5/10 The Spectrum of Generative AI Failure Modes Matt prompts Shreya to define the broader spectrum of GenAI risks beyond hallucinations. Shreya educates him on compliance failures, brand reputation risks, and privacy leakage, while Matt interjects with friendly banter.9:21–23:27 · Guest teaching 6/10 Evaluating Mitigation Techniques: RAG vs. Fine-Tuning Matt asks technical questions about RAG versus fine-tuning and pushes for concrete hallucination reduction metrics. Shreya reframes expectations by explaining that LLMs remain next-token predictors regardless of fine-tuning, offering a nuanced perspective on open-source versus proprietary benchmarks.23:27–35:23 · Guest teaching 6/10 Guardrails AI Framework: Guards, Validators, and Provenance Matt presses Shreya on the paradox of using probabilistic machine learning to correct other machine learning models in order to reach reliability. Shreya clarifies how ensembling diverse validators creates overlapping safeguards similar to a stack of sieves.35:23–40:22 · Guest teaching 5/10 Managing Failure Modes, Execution Policies, and Latency Matt asks how failure modes are handled practically and how latency is impacted at runtime. Shreya details self-healing re-asking loops, programmatic fallback policies, and parallel execution trade-offs.40:22–49:13 · Guest teaching 4/10 Open-Source Strategy and the AI Safety Ecosystem Matt demonstrates strong industry awareness by referencing key companies and ecosystem leaders such as LangChain, Weights & Biases, and LlamaIndex. Shreya elaborates on academic and open-source contributions from Stanford Hazy Research, Microsoft RAIL, and UW.49:13–50:49 · Guest teaching 0/10 Connecting with Guardrails AI and Career Opportunities A brief and friendly wrap-up covering open roles, social handles, and contact information, marked by joking about referral fees.0:14–5:48 · Guest disagreement 0/10 Shreya Rajpal's Background & The Origin of Guardrails AI Matt opens with a supportive question connecting Shreya's autonomous vehicle background to Guardrails AI. Shreya takes the floor to explain how building ML applications is more akin to self-driving systems than traditional predictive ML models.5:48–9:21 · Guest disagreement 1/10 The Spectrum of Generative AI Failure Modes Matt prompts Shreya to define the broader spectrum of GenAI risks beyond hallucinations. Shreya educates him on compliance failures, brand reputation risks, and privacy leakage, while Matt interjects with friendly banter.9:21–23:27 · Guest disagreement 2/10 Evaluating Mitigation Techniques: RAG vs. Fine-Tuning Matt asks technical questions about RAG versus fine-tuning and pushes for concrete hallucination reduction metrics. Shreya reframes expectations by explaining that LLMs remain next-token predictors regardless of fine-tuning, offering a nuanced perspective on open-source versus proprietary benchmarks.23:27–35:23 · Guest disagreement 1/10 Guardrails AI Framework: Guards, Validators, and Provenance Matt presses Shreya on the paradox of using probabilistic machine learning to correct other machine learning models in order to reach reliability. Shreya clarifies how ensembling diverse validators creates overlapping safeguards similar to a stack of sieves.35:23–40:22 · Guest disagreement 0/10 Managing Failure Modes, Execution Policies, and Latency Matt asks how failure modes are handled practically and how latency is impacted at runtime. Shreya details self-healing re-asking loops, programmatic fallback policies, and parallel execution trade-offs.40:22–49:13 · Guest disagreement 0/10 Open-Source Strategy and the AI Safety Ecosystem Matt demonstrates strong industry awareness by referencing key companies and ecosystem leaders such as LangChain, Weights & Biases, and LlamaIndex. Shreya elaborates on academic and open-source contributions from Stanford Hazy Research, Microsoft RAIL, and UW.49:13–50:49 · Guest disagreement 0/10 Connecting with Guardrails AI and Career Opportunities A brief and friendly wrap-up covering open roles, social handles, and contact information, marked by joking about referral fees.0:14–5:48 · Matt pushing back 0/10 Shreya Rajpal's Background & The Origin of Guardrails AI Matt opens with a supportive question connecting Shreya's autonomous vehicle background to Guardrails AI. Shreya takes the floor to explain how building ML applications is more akin to self-driving systems than traditional predictive ML models.5:48–9:21 · Matt pushing back 0/10 The Spectrum of Generative AI Failure Modes Matt prompts Shreya to define the broader spectrum of GenAI risks beyond hallucinations. Shreya educates him on compliance failures, brand reputation risks, and privacy leakage, while Matt interjects with friendly banter.9:21–23:27 · Matt pushing back 2/10 Evaluating Mitigation Techniques: RAG vs. Fine-Tuning Matt asks technical questions about RAG versus fine-tuning and pushes for concrete hallucination reduction metrics. Shreya reframes expectations by explaining that LLMs remain next-token predictors regardless of fine-tuning, offering a nuanced perspective on open-source versus proprietary benchmarks.23:27–35:23 · Matt pushing back 3/10 Guardrails AI Framework: Guards, Validators, and Provenance Matt presses Shreya on the paradox of using probabilistic machine learning to correct other machine learning models in order to reach reliability. Shreya clarifies how ensembling diverse validators creates overlapping safeguards similar to a stack of sieves.35:23–40:22 · Matt pushing back 1/10 Managing Failure Modes, Execution Policies, and Latency Matt asks how failure modes are handled practically and how latency is impacted at runtime. Shreya details self-healing re-asking loops, programmatic fallback policies, and parallel execution trade-offs.40:22–49:13 · Matt pushing back 0/10 Open-Source Strategy and the AI Safety Ecosystem Matt demonstrates strong industry awareness by referencing key companies and ecosystem leaders such as LangChain, Weights & Biases, and LlamaIndex. Shreya elaborates on academic and open-source contributions from Stanford Hazy Research, Microsoft RAIL, and UW.49:13–50:49 · Matt pushing back 0/10 Connecting with Guardrails AI and Career Opportunities A brief and friendly wrap-up covering open roles, social handles, and contact information, marked by joking about referral fees.

speaking balance: gold is Matt, purple is the guest (3 minute bins)

0:00 · Matt 38.2% · guest 61.8%0:00 · Matt 38.2% · guest 61.8%3:00 · Matt 7% · guest 93%3:00 · Matt 7% · guest 93%6:00 · Matt 8.6% · guest 91.4%6:00 · Matt 8.6% · guest 91.4%9:00 · Matt 27.1% · guest 72.9%9:00 · Matt 27.1% · guest 72.9%12:00 · Matt 16.9% · guest 83.1%12:00 · Matt 16.9% · guest 83.1%15:00 · Matt 1.7% · guest 98.3%15:00 · Matt 1.7% · guest 98.3%18:00 · Matt 22.9% · guest 77.1%18:00 · Matt 22.9% · guest 77.1%21:00 · Matt 14% · guest 86%21:00 · Matt 14% · guest 86%24:00 · Matt 1.1% · guest 98.9%24:00 · Matt 1.1% · guest 98.9%27:00 · Matt 8.2% · guest 91.8%27:00 · Matt 8.2% · guest 91.8%30:00 · Matt 20.3% · guest 79.7%30:00 · Matt 20.3% · guest 79.7%33:00 · Matt 16.2% · guest 83.8%33:00 · Matt 16.2% · guest 83.8%36:00 · Matt 5.7% · guest 94.3%36:00 · Matt 5.7% · guest 94.3%39:00 · Matt 28.3% · guest 71.7%39:00 · Matt 28.3% · guest 71.7%42:00 · Matt 15.3% · guest 84.7%42:00 · Matt 15.3% · guest 84.7%45:00 · Matt 38.1% · guest 61.9%45:00 · Matt 38.1% · guest 61.9%48:00 · Matt 23.5% · guest 76.5%48:00 · Matt 23.5% · guest 76.5%
Sharpest disagreement ▶ 17:40 Spicy take on fine-tuning and hallucinations

Shreya mildly counters general assumptions by calling it a spicy take that fine-tuning cannot fundamentally fix hallucinations because models are strictly next-token predictors.

 Hardest push from Matt ▶ 32:35 Matt challenges ML correcting ML

Matt directly challenges Shreya on how using imperfect machine learning techniques to validate other imperfect machine learning outputs can achieve 100 percent system accuracy.

 Biggest teaching moment ▶ 17:15 Explaining fine-tuning limits vs structured tasks

Shreya educates the host on how fine-tuning improves structured task formats like JSON output but fails to eliminate underlying factual hallucinations in unstructured text.

 Matt holds his own ▶ 45:13 Matt lists leading AI ecosystem figures

Matt displays deep ecosystem familiarity by citing recent guests, specific company announcements, and key founders across LangChain, Weights & Biases, and LlamaIndex.
the scores for every segment, with the reasoning behind each
ChapterTopicMatt as informed peerGuest teachingGuest disagreementMatt pushing backWhy
Shreya Rajpal's Background & The Origin of Guardrails AI 2400 Matt opens with a supportive question connecting Shreya's autonomous vehicle background to Guardrails AI. Shreya takes the floor to explain how building ML applications is more akin to self-driving systems than traditional predictive ML models.
The Spectrum of Generative AI Failure Modes 3510 Matt prompts Shreya to define the broader spectrum of GenAI risks beyond hallucinations. Shreya educates him on compliance failures, brand reputation risks, and privacy leakage, while Matt interjects with friendly banter.
Evaluating Mitigation Techniques: RAG vs. Fine-Tuning 5622 Matt asks technical questions about RAG versus fine-tuning and pushes for concrete hallucination reduction metrics. Shreya reframes expectations by explaining that LLMs remain next-token predictors regardless of fine-tuning, offering a nuanced perspective on open-source versus proprietary benchmarks.
Guardrails AI Framework: Guards, Validators, and Provenance 5613 Matt presses Shreya on the paradox of using probabilistic machine learning to correct other machine learning models in order to reach reliability. Shreya clarifies how ensembling diverse validators creates overlapping safeguards similar to a stack of sieves.
Managing Failure Modes, Execution Policies, and Latency 4501 Matt asks how failure modes are handled practically and how latency is impacted at runtime. Shreya details self-healing re-asking loops, programmatic fallback policies, and parallel execution trade-offs.
Open-Source Strategy and the AI Safety Ecosystem 5400 Matt demonstrates strong industry awareness by referencing key companies and ecosystem leaders such as LangChain, Weights & Biases, and LlamaIndex. Shreya elaborates on academic and open-source contributions from Stanford Hazy Research, Microsoft RAIL, and UW.
Connecting with Guardrails AI and Career Opportunities 1000 A brief and friendly wrap-up covering open roles, social handles, and contact information, marked by joking about referral fees.

Statements from this episode (7)

Insight
Production LLM deployment challenges mirror autonomous vehicle development
“The trajectory of issues and concerns that people are running into are similar to, you know, like the path is similar to what it was in self-driving, which is, you know, How do I get like this, you know, runtime safety? How do I get like runtime constraints, e…”
Shreya Rajpal Sep 27, 2023 ▶ 4:56
Insight
Existing model risk frameworks fail for third-party AI models
“Existing, for example, like model risk management frameworks don't really apply when you haven't built the model yourself. You know, you didn't like curate the data that the model was trained on. And so you can't make any claims to that.”
Shreya Rajpal Sep 27, 2023 ▶ 7:10
Insight
Fine-tuning cannot eliminate LLM hallucinations
“At the end of the day, these models are like next token predictors, which is, you know, like they kind of look at like what they've predicted until now, and then, you know, figure out like what the next token they're on is. And from that, like, even with fine-…”
Shreya Rajpal Sep 27, 2023 ▶ 17:46
Assertion Not checkable as stated
Fine-tuned Llama 2 achieves performance comparable to GPT-3.5 and GPT-4
“Straight out of the bat, if you just use Lama tool directly, I don't think you could get like comparable performance, you know, with GPD, 3.5 or four today. But like with fine tuning, if you make that investment in curating your data set in running that fine t…”
Shreya Rajpal Sep 27, 2023 ▶ 22:48
Insight
GenAI solves ML's first mile, but traditional tools handle the last
“LLMs and Generative AI really helps solve, like the first mile problem in ML, right? But the last mile problem, which is, like, how do you take this generic generalizable technology and make it work, like, specifically for your use cases and for your actual ap…”
Shreya Rajpal Sep 27, 2023 ▶ 34:43
Insight
Open source turns abstract AI safety into an actionable engineering problem
“What the open source really ends up doing as you know, a framework is taking down this very abstract problem of what it means to do safe AI development, right? Like it's a very abstract problem. It's almost an academic problem to some degree, and it takes that…”
Shreya Rajpal Sep 27, 2023 ▶ 41:47
Opinion
RAG is the definitive way to build generative AI today
“Rag is the way to build you know, these models today”
Shreya Rajpal Sep 27, 2023 ▶ 46:33
Made with StarZero

Turn any episode into a week of clips.

This entire site, over 400 conversations transcribed, diarized, checked and made playable, runs on the StarZero media pipeline. Drop in your own episode and the podcast clipper finds the moments worth sharing, cuts them, captions them, and reframes them for every feed.

Try the podcast clipper Learn more