Becoming an AI engineering company
By Steven Van ·
Vercel's CTO sets out how teams can build with AI SDK and v0, from prompting foundation models to cutting costs with evals and fine-tuning.
Vercel's CTO has published a post on how companies can become AI engineering companies, drawing on 30 years as an engineer. It is an opinion piece rather than a product release, and it uses Vercel's own AI SDK and v0 as examples.
The post says AI development no longer needs advanced degrees or specialised infrastructure. It contrasts the old approach (Python, small purpose-built models, months of training) with a newer one: TypeScript, large foundation models used as a service, and prompting instead of training.
For companies worried about competing with OpenAI, Google and Microsoft, it names three advantages:
- Exclusive data: customer data, internal documentation and domain-specific information that large models cannot access.
- Feedback cycle: the ability to test and adjust faster than the big vendors can.
- Complicated domain: business complexity that general models do not understand.
It also describes an optimisation cycle. Get something working even if it is slow or expensive, ship early and collect feedback, move to cheaper middle-tier models, use evals to catch regressions, then cut costs further with techniques like fine-tuning.
The AI SDK is presented as a way to switch between models with one or two lines of code, using functions such as generateText. The post argues AI should sit in the core of an application, with access to its components, data layer and business logic, rather than being a chat layer on top. v0, which turns natural language prompts into interactive web experiences, is the example of that approach.
