CTO Arquitectura y Plataformas Core at Falabella

After more than 20 years at the helm of technology organizations, I felt the urge to return to my roots last year. There was a looming risk of losing touch with the latest advancements in technology. Trading my role as Walmart's CIO, I embarked on a new journey as the CTO of Architecture and Core Platforms at Falabella, a leading multi-format retailer and bank in Latin America. Alongside this transition, I found myself diving back into coding. And then, like a storm on the horizon, ChatGPT took my attention. The moment I engaged with it, I envisioned a future for my children, unlike anything I had ever known. I was compelled to be part of this transformation. Since then, I've dedicated countless hours to studying, designing, and coding applications, balancing my personal endeavors with the responsibilities at Falabella. Here, I present some insights from my dual perspective as an executive and developer.

The Impact of LLMs on Companies and IT Organizations

At its core, the engine driving ChatGPT offers the potential to craft personalized solutions. Available as an API, it can process a singular request or an entire conversation, mirroring the ChatGPT interaction. When referring to ChatGPT or OpenAI’s API, it's essential to understand that I'm highlighting the broader category of large language models (LLMs). Though GPT-4 reigns supreme for now, its dominance might be challenged soon.

Here are some applications of these APIs that I’ve already successfully applied:

Q&A: Despite the inability to ‘train’ the API about your company's specifics, there are straightforward methods to utilize a knowledge base for answering queries. The simplicity I'm emphasizing means you need no more than 20 lines of code for a preliminary test.

Planning and Executing: The open-source community brims with innovation. A project named Langchain showcases a unique approach using the model to comprehend problems and devise solutions. You can introduce custom functions (e.g., checking vacation balances) and guide these ‘Agents’ to incorporate them as necessary. With minimal coding, you can initialize an Agent equipped to plan, invoke your functions, process their outputs, and respond to users.

 ​Although LLMs falter with numerical data, their prowess in coding, planning, and action remains unmatched 

Coding: As a seasoned coder adapting to contemporary techniques (i.e., an old dog learning new tricks), GPT's assistance in coding feels revolutionary. You can not only ask it to code but also brief it on an error, ask it to explain code, teach you a concept, refactor code, and more! This partnership enhances productivity, especially once you master the dialogue with it.

Data Analytics: Although LLMs falter with numerical data, their prowess in coding, planning, and action remains unmatched. Consequently, an Agent can decode a request to scrutinize a structured source (like a database table), draft a code for analysis (such as linear regression), present the results, and add its interpretations.

The fusion of these capabilities equips us with tools only bounded by our creativity. While there's room for refining these models, we can harness numerous applications currently.

Some Tips for Your Journey with LLMs:

Not Everything is a Chatbot:Our star project right now is reading an Excel file with eCommerce products and returning a new file with beautiful descriptions for those products. Not the sexiest solution, I know, but the value vs effort is just nuts.

Prototyping vs. Productionizing: Prototyping is swift, but transforming it into a full-fledged application demands time. Aspects like MLOps, quality validation (especially challenging with LLMs), guardrails, and more mandate a learning curve. Nevertheless, beginwithprototypingtoinitiateyourlearningjourney.

Persistence: If initial results falter, persist! Quality outputs require multiple iterations and inventive approaches. For example, if the model fumbles in meeting your entire criteria, feed it requirements progressively, building on previous interactions.

Leverage Existing Solutions: Existing frameworks, like Langchain, offer valuable constructs (like Agents). While crafting your own may seem tempting, it might not be worthwhile. Instead, streamline infrastructure issues for your developers, such as granting them API keys or providing a vector database for semantic searches—integral for most LLM architectures.

Nurture Communities: Empower your developers with these cutting-edge tools. By providing support and fostering collaboration, you'll be amazed by their innovations.

Transformation is afoot, with inertia being the prime adversary. This technological wave isn't a fleeting trend—it's poised to grow exponentially. In a few years, ensure you're among those who foresaw this revolution and seized the opportunity.