Andrew Bosworth argues the industry is overly focused on model benchmarks. He believes that as models become rentable commodities, the real, defensible value will be in the consumer-facing product experience. Users care about functionality, not which model version they are using.
According to Meta's CTO, the era of one monolithic model doing everything is over. The current frontier involves using a 'harness' that intelligently routes tasks to a collection of different, specialized models based on cost, latency, and capability.
Meta's internal tracking program is designed to create a unique dataset for a fundamental AI challenge: teaching models how to proficiently use computer interfaces. Bosworth notes AIs are currently 'weirdly bad' at this task, which is a key bottleneck for agentic capabilities.
Meta's CTO distinguishes between unproductive pain (e.g., manual math) and productive pain (the struggle of adapting to new technology like AI). He argues that embracing the discomfort of organizational change is crucial for progress, and avoiding it leads to failure.
To release the highly-regarded Llama 3 model, Meta's researchers pulled forward all future research bets. This cannibalization of the R&D pipeline left them without the necessary pathfinding work for Llama 4, causing them to fall behind on newer techniques like mixture of experts.
Andrew Bosworth frames AI not as a path to merging with machines, but as a tool to drastically increase the speed and fidelity of information transfer between human intent and computer execution. This follows the historical HCI trend of tools like the mouse and autocorrect.
While Meta uses third-party models from Google or Anthropic, CTO Andrew Bosworth states that having a competitive in-house model is crucial. It acts as a backstop, preventing providers from charging exorbitant rent and ensuring Meta can control its own destiny if needed.
Meta's CTO believes consumer AI hasn't taken off because current applications are not easy enough or valuable enough to change people's daily routines. The technology has passed the hype peak and is now in the hard-work phase of solving user experience and friction problems.
The future interface for wearables won't be a traditional app store. Instead, users will describe their needs in real-time ('find me a 5k race and sign me up'), and the onboard AI will generate and execute the necessary 'app' on the fly.