853: Generative AI for Business, with Kirill Eremenko and Hadelin de Ponteves

Episode Summary

Executive Summary: Kirill Aramenko and Adlain de Pontev explained how foundation models power modern generative AI, how they differ from and include large language models, and how businesses can customize them through fine-tuning, RAG, agents, and prompt templates. They outlined an eight-step foundation-model lifecycle, 12 selection criteria, and AWS tools (Q, Bedrock, SageMaker) for practical deployment and enterprise adoption.

Main Topics: What foundation models are and how they differ from LLMs: The guests defined foundation models as pre-trained base models for building custom AI applications, with large language models as one subset among broader model types including vision and multimodal models. Eight-step foundation model lifecycle: They walked through the end-to-end lifecycle: data prep, architecture choice, pre-training, fine-tuning, evaluation, deployment, monitoring, and maintenance/iteration. How businesses customize models during training: They described modification approaches during training, including from-scratch training, continued pre-training, domain-specific fine-tuning, instruction-based fine-tuning, and RLHF. How businesses customize models during inference: They explained inference-time customization via temperature and other parameters, RAG, agents/agentic AI, and prompt templates, emphasizing these can materially change outputs without retraining. Choosing the right foundation model: They presented 12 criteria for model selection, including cost, modality, customization, latency, benchmarks, language support, scalability, licensing, compliance, and environmental impact. AWS generative AI stack: They compared Amazon Q, Bedrock, and SageMaker as three levels of AWS tooling for GenAI, from plug-and-play to highly customizable and technical. Bravo Tech launch and enterprise implementation services: The episode announced Kirill and Adlain’s new consulting venture, Bravo Tech, offering implementation, consulting, and training for GenAI projects, including a free first three hours for podcast listeners.

Key Arguments: Foundation models are the practical base layer for enterprise GenAI: companies do not need to train the foundation model from scratch, only adapt and deploy it for their use case. The most important business customization path is usually domain-specific fine-tuning, because it aligns a general model with proprietary organizational data and workflows. Instruction-based fine-tuning can force a model to answer in a desired style, length, or format, not just learn domain knowledge. RAG is a highly effective way to combine a model’s internal knowledge with an organization’s documents and databases at inference time. Agentic AI can outperform a single large model on complex workflows by breaking tasks into smaller steps and orchestrating them. Model selection is multi-factor, not just about benchmark performance; cost, latency, compliance, scale, and environmental impact matter as much as raw capability. For many enterprises, success depends as much on change management and user adoption as on the technical model itself. AWS offers a gradient of options: Amazon Q for simple use cases, Bedrock for model choice and customization, and SageMaker for deep technical control.

Data Points: Podcast episode number: 853 - Current episode of the Super Data Science Podcast CloudWolf student count: over 5 million students between them - Kirill and Adlain’s combined reach as Udemy instructors AWS services discussed for GenAI: 3 - Amazon Q, Amazon Bedrock, and SageMaker Foundation model lifecycle steps: 8 - Data prep, model selection, pre-training, fine-tuning, evaluation, deployment, monitoring, maintenance Selection criteria for foundation models: 12 - Criteria listed for choosing a model for a business use case Fine-tuning example duration: a couple of dozens of minutes - Adlain’s medical fine-tuning lab on a Llama model Student lab app build time: 5 minutes - Adlain’s Yoda-style chatbot built in Bedrock Temperature range example: closer to one / at zero - Used to explain creativity vs determinism in inference parameters Amazon Q / Bedrock / SageMaker levels: 3 levels - High-level, mid-level, and low-level AWS GenAI tooling

Pivotal Quotes: "The first three steps are done for you, don't even need to worry about them." — Kirill Aramenko: Explaining why foundation models make GenAI accessible to businesses "All you have to do is take that bottom layer of cake and create your own cake and use it to your heart's desire in your organization." — Kirill Aramenko: Summarizing the foundation-model-as-cake analogy for enterprise customization "If you spend time prototyping with the most expensive model, then you'll have to kind of like redo the work when you just realize, oh, this is too cost-intensive." — Adlain de Pontev: Discussing why cost may matter earlier than expected for enterprise prototypes

Implications: The episode frames GenAI as commercially accessible but operationally nuanced: organizations should choose models strategically, customize thoughtfully, and invest in deployment, governance, and change management. AWS now offers a full stack for both quick wins and deep enterprise builds.