May 1, 2023 · 27m · mad
Build and Deploy AI with Pytorch | Lightning AI Founder & CEO, William Falcon
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
William Falcon, Founder and CEO of Lightning AI, joins Matt Turck on The MAD Podcast to discuss the evolution of PyTorch Lightning, open-source AI strategies, enterprise deployment risks, and the future landscape of specialized language models.
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 14.9% of the talking time here. How this is scored →
Matt as informed peer 3.6 Guest teaching 5.3 Guest disagreement 2.1 Matt pushing back 0.9
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 18:23 Dismissing GenAI tooling stack as noise
Falcon forcefully rejects market consensus around tools like LangChain and vector databases, calling 99% of the current hype temporary research noise fueled by VC capital.
Hardest push from Matt ▶ 15:18 Host clarifies regulated industry constraintsTurck cuts in to sharpen Falcon's argument about sequential adoption, explicitly identifying regulatory requirements as the core impediment in banking and health.
Biggest teaching moment ▶ 21:30 Reframing LLM scale as scientific ignoranceFalcon reframes the multi-billion parameter model arms race, explaining that throwing massive compute at LLMs merely substitutes for a current lack of mathematical elegance in loss functions.
Matt holds his own ▶ 18:03 Host cites modern tooling primitivesTurck shows strong domain knowledge by specifically identifying vector databases and LangChain to press the guest on how the enterprise GenAI architecture fits together.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Welcome and Overview of Recent Milestones | 3 | 4 | 1 | 0 | Matt sets up the interview with accurate financial and product milestones before handing the floor to Will. Will provides an extensive history of deep learning frameworks and the engineering motivations behind PyTorch Lightning. | |
| Lit Llama and Open Source AI Models | 3 | 5 | 1 | 1 | Matt interrupts constructively to ask Will to explain Meta's Llama model for non-technical listeners and inquires about monetization. Will explains the open-source licensing differences between GPL and Apache 2.0 for enterprise use. | |
| Risks and Reality of Enterprise LLM Deployment | 4 | 5 | 1 | 1 | Matt references past interviews where Will cautioned enterprises against rushing into LLMs. Will explains why chat interfaces distort real production readiness, highlighting hallucinations and auditability risks. | |
| Timeline for Hallucination Solutions and Regulated Industries | 4 | 5 | 2 | 2 | Matt presses Will on adoption timelines and explicitly highlights regulatory burdens in healthcare and finance. Will enthusiastically argues that all AI advances ultimately stem from and return to open source. | |
| Evaluating the Emerging Generative AI Tooling Stack | 5 | 6 | 4 | 1 | Matt demonstrates knowledge of the modern AI stack by asking specifically about vector databases and LangChain. Will bluntly characterizes 99% of current GenAI tooling hype as ephemeral research noise. | |
| Small vs. Large Models and Future AI Architectures | 4 | 6 | 3 | 1 | Matt outlines two competing industry hypotheses regarding massive models versus specialized multi-model ecosystems. Will argues that giant models reflect a temporary lack of mathematical understanding rather than a permanent paradigm. | |
| Lightning AI's Three-Year Vision and Enterprise Roadmap | 2 | 6 | 3 | 0 | Will explains Lightning AI's open-core business model in response to an audience question about cloud provider monetization. He then rejects the premise of AGI, framing human and machine intelligence as specialized. |
Statements from this episode (15)
Falcon: Stability AI trained Stable Diffusion using PyTorch Lightning
“Who's heard of stability, AI, and stable diffusion? That was trained using Lightning, right?”
Falcon: Nvidia NeMo services are powered by PyTorch Lightning
“NVIDIA just announced all these Nemo services. Those are all powered by Lightning.”
Falcon: Lightning AI rebuilt Meta's LLaMA from scratch under Apache 2.0
“So we took the Lama, like, paper and implemented from scratch completely in Apache two, and we open sourced it. So it's fully, fully usable for enterprises, but we also gave you the training code, not just inference, and we also gave you fine tuning methods as…”
Falcon: Lightning AI will not offer hosted model APIs
“But no, we're not gonna be offering like services and APIs and all of that. And really that's so that we can have the same incentives as the open source community.”
Falcon: Avoid full enterprise LLM deployment without human-in-the-loop
“So I wouldn't go all in on this unless you can have like a human in the loop who's like helping.”
Falcon: Solving AI hallucination will take roughly five years
“So I think it's still going to be, I don't know, like five years, probably like, I don't think it's going to be like one day suddenly it's solved. I think it's going to unlock industries sequentially. So like, year one, maybe we can do this industry, and then …”
Falcon: Open-source AI will quickly catch up to proprietary models
“So I think all, all roads lead to open source at the end of the day. No matter what you do or how much money you have, you cannot compete with the world's resources put together to do something. So some companies will have an edge for a bit, yes, but open sour…”
Falcon: Modern generative AI would not exist without open-source research
“None of this wouldn't exist if Google hadn't put published transformers and put the paper out there, right? None of this would have existed if Ah, the attention stuff would have been out there if the Anno hadn't been created, if TensorFlow hadn't been created.”
Falcon: Probably only 1% of emerging generative AI tools will endure
“I think it's a lot of research. Like, what's gonna last from there? I'm not sure. Like, probably one percent of the things.”
Falcon: PyTorch Lightning's original tooling was insufficient for post-GPT-3 architectures
“The tool that we had was good for 2016 through 20 19 deep learning, but as of GPT three deep learning, it wasn't good enough, and so we had to upgrade to this kind of new paradigm that we introduced.”
Falcon: AI development continuously cycles between scaling up and compression
“There's always going to be a pattern of go big to get some result and then compress it back.”
Falcon: Scaling massive AI models compensates for gaps in scientific understanding
“I kind of think that the whole large scale model thing, It's really a gap in our scientific understanding of these models. Instead of looking at the math and coming up with a better loss function, we figured out that we could throw a compute at it, and get us …”
Falcon: Mathematical advances will eventually make future AI models small
“I think the math field will catch up, and we'll have some fancy regularizer that will make all this go away, and now we're back to, like, a small model. So I guess, I think that the future is small, large language models, I guess, that are specialized”