Behind the biggest consumer AI adoption trap with Meta’s CTO

This article summarizes an episode of Alex Kantrowitz’s video series featuring Andrew Bosworth, CTO of Meta.

Andrew Bosworth, CTO of Meta/ Photo credit: Andrew Bosworth

Andrew Bosworth, CTO of Meta, says that an AI race can be lost before the next one begins. He attributes Meta’s recent AI setbacks not to a lack of chips or talent, but to a depleted research pipeline.

Even a well-resourced company can exhaust its future options to win a current release, leaving it unprepared for the next wave of innovation.

Meta used up research meant for the next model

Meta’s Llama development highlights a critical risk inside top AI labs where meeting an immediate launch goal can cripple the next one. Bosworth says that while Llama 3 appeared successful, it consumed research intended for Llama 4.

By pulling future ideas into the current release, Meta left itself with too few ongoing experiments when the next development cycle began.

Bosworth explains,  “When it came time for Llama 4, we didn’t have any of the pathfinding the other labs still had going. So we were behind on reasoning and mixture of experts.”

This creates a distinct management challenge to protect research that lacks an immediate launch date. Cutting-edge AI development requires multiple parallel paths.

If a company forces every useful idea into the immediate release, it may ship today only to fall decisively behind tomorrow.

Using many models changes who has power

This planning failure exposes the secondary challenge of navigating a multi-model ecosystem without becoming dependent on external suppliers.

Bosworth expects future products to route tasks to different models depending on the task. To succeed, products must route requests efficiently while the company maintains its own competitive models to prevent vendor lock-in.

The industry is shifting away from monolithic models. “We’ve moved past this world where one model rules everything,” he notes. “What you want is a very expensive-to-run, intelligent model and to use its exquisite intelligence only when necessary.”

Companies no longer build their own AI just to have a famous brand name. Instead, they build in-house models to control costs, keep their products running smoothly, and avoid being at the mercy of outside vendors.

Because choosing a model affects the bottom line so heavily, AI strategy, budgeting, and product design are now completely tied together.

Consumer AI struggles when it asks people to change habits

Expert training work shows the cost of moving fast

Work data is becoming part of AI training