Episode summary
A Masterclass on AI: How do LLMs Work and how does is Impact Investing?
The Intrinsic Value Podcast
· · 1 hr 29 min · Guests: Satish Terala
Summary
Satish Terala walks through how modern AI evolved from classical machine learning to deep learning, transformers, reasoning models, and agents. He argues that older ML systems were limited by manual feature engineering, poor handling of unstructured data, and scaling constraints, while deep learning improved performance by stacking layers that learn increasingly abstract representations. In his framing, the breakthrough behind large language models is not magic but large-scale matrix math over embeddings, tokens, and attention weights.
He explains that transformers changed the field because they can parallelize training, work across text, images, code, and audio, and use attention to decide which words matter most in context. Satish says generative AI differs from discriminative AI because it begins with a prompt and produces new output rather than classifying existing data. He also highlights why GPUs and Nvidia matter: training requires massive parallel computation, while inference is cheaper and can run on different hardware.
He closes by describing reasoning models and AI agents as the next layer. Reasoning models break problems into steps and self-check, while agents combine an LLM with tools, files, APIs, and enterprise data structures like ontologies so the model can act on real-world context rather than only answer questions from its training data.
Key takeaways
- Satish Terala says classical ML was constrained by manual feature engineering, limited scaling, and weak performance on unstructured text, image, and video data.
- Satish Terala says deep learning and transfer learning let models reuse learned layers across domains, which accelerated the pace of AI progress.
- Satish Terala says transformers matter because attention, embeddings, and tokenization let models predict the next word from context in parallel.
- Satish Terala says Nvidia’s moat comes from CUDA plus the software ecosystem around PyTorch, not just from building GPUs.
- Satish Terala says reasoning models improve accuracy by making the model self-check step by step before answering.
- Satish Terala says AI agents need an LLM plus tools and structured enterprise data, and ontology makes that data easier for the model to use.
Notable quotes
“the biggest innovation that has happened between the pure LMS and the reasoning models and where this agent stuff comes into play”
“nobody actually assigns weights to any of these models. Weights are the outputs that come out of the training exercise”
“The beauty of this thing is that what it enables is what is what we call as transfer learning”
“what this does is that it takes a word, converts that to a number, puts it up into some kind of high dimensional space”
How the episode opens
You know you'd get a requirement you'd basically say I'll do a database search I'll run a database query get this data you know based on this data I'm going to do x something else right programmers were the ones that were writing code to orchestrate all this stuff now the agent is able to figure out those things and then orchestrate all these tool calls based on this context window right that's the biggest innovation that has happened between the pure LMS and the reasoning models and where this agent stuff comes into play and why does programmers jobs feel threat at this point the beauty of code is that it follows set of rules like you know the syntax is extremely fixed once you learn a pattern you can apply the pattern repeatedly again and again there's nothing it's not a poem it's not poetry it's not pros there's no new words coming there's no there all right and there we are so um today Satish will give us or teach us some lessons on how he uses and leverages AI um I think we'll talk about open claw perhaps also cloud especially like the top tier tools that most of us are either currently using or we hope to use in the future. Um I think you'll do a quick demo right of how you use it share your screen. So yeah I'm excited for that and with that I would say how about you just start and teach us some AI lessons.
Yeah.
From the episode’s captions: roughly the first two minutes, lightly cleaned. Keep watching on YouTube →
Topics
- ai llms
- llm architecture
- transformers and attention
- tokens and embeddings
- ai agents
Companies and tickers mentioned
- NVDA
- Nvidia
- DeepSeek
- OpenClaw
Watch the episode
Watch “A Masterclass on AI: How do LLMs Work and how does is Impact Investing?” on YouTube
Summary, takeaways and quote selection by Octagon. Quotes are short excerpts; the full conversation is in the episode.
More from The Intrinsic Value Podcast
-
Canadian Pacific Kansas City Limited (CP.TO): Irreplaceable Moat, Unproven Compounding -
Reddit (RDDT): Record Cash Flow, Collapsing Share Price -
McDonald's ($MCD): Generational Compounder in Historical Drawdown -
Mercado Libre ($MELI): We Doubled Down on Meli! -
Airbnb (ABNB) The Business Outgrew Its Peak, the Stock Hasn't -
GE Aerospace: The Ultimate Compounder?
The Intrinsic Value Podcast, this episode and its artwork belong to The Investor's Podcast Network. Octagon is not affiliated with the show; this page is an independent summary.
