Devansh Tandon argues that recommendation quality improves as data, model capacity and compute scale, drawing on Meta research and product examples. He frames a feedback loop in which improved recommendations increase engagement and revenue, which can fund another round of model training. These are speaker-presented examples and forecasts, not independent measurements in this draft.
Devansh Tandon describes four overlapping approaches: traditional feature and embedding systems, language-model-inspired recommenders, LLM-native models adapted to recommendation tasks, and a prospective agentic loop that plans, retrieves, ranks and critiques results. The agentic stage is presented as an emerging research direction rather than a deployed norm.
A proposed training recipe tokenizes content into compact semantic IDs, teaches a model to connect those IDs with natural language and interaction sequences, then tunes it for ranking and engagement tasks. A shared foundation can be adapted across several product surfaces. The Instagram examples illustrate how a user might inspect or steer a feed with language instead of relying only on clicks and likes.
The consumer-business argument compares a feed model that outputs pointers to existing creator content with a chat model that generates many output tokens itself. Devansh Tandon contends that this makes recommendation feeds comparatively efficient at serving long sessions and could support significant consumer AI applications; the cost gap and future adoption remain his projections.
Watch on YouTube
