How to make AI models smarter for you without making them bigger?Iโm taking Berkeleyโs CS294 ๐๐๐ ๐ผ๐๐๐ฃ๐ฉ๐จ course, and I particularly enjoyed a recent lecture from Burak Gokturk where he explained several approaches that will help enterprises get significantly better results from foundation models.
๐ง๐๐ป๐ถ๐ป๐ด/๐๐ถ๐๐๐ถ๐น๐น๐ฎ๐๐ถ๐ผ๐ป
โข Tuning is customizing a model based on specific data and/or use case
This ranges from techniques as basic as prompt design (providing few shot examples as part of your prompt), to actual retraining of the model (either full [very costly] retraining, or more efficiently with a concept like โLoRAโ).
โข Distillation is creating a smaller model for improved cost/latency, using a teacher/student pattern, where your foundation model will โteachโ (generate training data and reward/penalise) a smaller (student) model.
Not every use case needs a model that can โanswer all the questions in the worldโ. Often, a smaller, specialized model is more effective.
๐๐ฟ๐ผ๐๐ป๐ฑ๐ถ๐ป๐ด
โข Combine with search to make it factual (also known as โRAGโ, Retrieval Augmented Generation)
โข Can be applied to web data or enterprise data
Think of this as making LLMs more reliable by connecting them to verified information sources (for example post training cutoff date), or sources that matter to your context (your enterprise data).
๐๐
๐๐ฒ๐ป๐๐ถ๐ผ๐ป๐/๐๐๐ป๐ฐ๐๐ถ๐ผ๐ป ๐๐ฎ๐น๐น๐ถ๐ป๐ด
Like the โPhone a Friendโ option in โWho Wants to Be a Millionaire,โ this involves 2 main complexities:
โข Recognizing when the model needs external help
โข Selecting the right tool for the task
For example, a LLM canโt book flights directly, but through function calling, it could use an Expedia extension to complete this task.
As a fractional CDO, I love exploring how these concepts can address real business challenges in real world AI implementations.
If youโre interested in the course, you can check it out here.
