Large Language Models in the Enterprise: Going Beyond Chatbots
- Philippe Beaudoin
- November 20, 2024
- Turkyie AI Summit, Istanbul
A few years ago, building AI meant building it from scratch: collecting, cleaning and labelling a mountain of data to train a custom model. It was slow, costly, and it felt less like software than like building a new factory every time.
Large language models changed the game. In this talk I introduced what I called AI not from scratch: products built on top of pre-trained models, where the key question becomes simple. Can you describe your problem in words, even though it is hard to write rules for it? If so, an LLM may be the right starting point.
Key ideas
- Data is no longer the new oil. At Element AI we used to tell customers that data was the new oil. With pre-trained models, a whole new category of AI needs very little data. The advantage shifts from those who own data to those who learn quickly about their customers.
- A three-step recipe: start from a large model, describe the task in natural language (in other words, a prompt), then collect small data once the task is described.
- From prompt hacking to prompt engineering. Prompts tend to grow, break, and resist fixing. The fundamental question of software engineering — how do we split a task efficiently? — becomes how do we split a prompt efficiently? Prompt-based functions take natural language in and out, and can be composed with classical control structures: map-reduce, if-then-else, filters, loops.
- Transparent personalization. User profiles written in natural language, readable and editable by the users themselves, bring personalization to products that don’t have big-tech-sized datasets.
- Why Canada. Deep learning has deep Canadian roots — Yoshua Bengio, Geoffrey Hinton, Richard Sutton — and the attention mechanisms behind today’s LLMs were pioneered in Montreal.
Looking back
“Beyond chatbots,” for me, meant more control: more structure, more harness around the model. A few months later, agents were everywhere. I like keeping this talk in the library as a reminder of how quickly the ground moves in this field, and of why it’s worth revising one’s views out loud.