Why LLM Recommenders Will Be AI's Biggest Consumer App — Devansh Tandon, Meta

Devansh Tandon, who leads Meta Recommendations Research, opens with two claims: recommendation systems follow scaling curves much like LLMs, and LLM recommenders will become one of AI’s biggest consumer applications. Recommendation, in his view, is still early in its climb. Follow the opening at 0:28.

The starting point is the power-law relationship familiar from language-model scaling: increasing model size, training data, and compute produces a predictable decline in loss. That predictability matters because it gives another training run an economic rationale. A company can estimate how additional compute might improve quality rather than treating each larger model as an entirely new experiment.

Recommendation applies the same idea to a different output. More data, compute, and model capacity improve offline recommendation quality; production launches then test whether those improvements become engagement and revenue. Tandon points to HSTU, a 2024 paper, and a subsequent Meta paper as examples of this scaling behavior. The distinction between offline quality and product impact matters: a better evaluation result is an intermediate step toward a better feed.

For Instagram, the concrete changes included simplifying the ranking architecture so models could scale efficiently, doubling the length of user interaction sequences used in training, and making each interaction richer. Tandon places these changes alongside reported Reels watch-time growth of 30% year over year. That is a product-level figure, not an isolated estimate of how much any one model change contributed.

What connects a scaling curve to a consumer business? Tandon’s phrase is “tokens in, engagement out.” Training produces a model; inference uses it to recommend content; better recommendations draw engagement through daily use and time spent. Advertising or subscriptions turn that engagement into revenue, which pays for another training run. Each successful move up the scaling curve completes another circuit. Follow the flywheel at 4:59.

How does the next training run connect to the one before it? The loop below makes the funding path visible: recommendations lead to engagement, engagement leads to monetization, and revenue returns to training. The return arrow is what turns a sequence of model improvements into a repeating business loop.

The next question is what gets scaled. Tandon organizes recommendation into four S-curves, each with a different source of improvement. These are overlapping approaches: companies can keep improving established systems while exploring models that understand language or agents that coordinate multiple recommendation steps. Follow the four paradigms at 5:39.

• Traditional recommendation: Feature engineering and user and content embeddings do much of the work. Two-tower models and sparse rankers belong here. Tandon expects this approach to persist, with agents potentially taking over more feature engineering and research work.
• LLM-inspired recommendation: Models borrow the scaling approach of LLMs, ideally learning end to end. HSTU is one example. Tandon places leading industry systems largely on this curve.
• LLM-native recommendation: An existing base model brings language understanding and reasoning, then adapts to recommendation tasks.
• Agentic recommendation: An LLM coordinates a loop that plans, retrieves, ranks, critiques, and refines recommendations before delivering them.

The distinction between LLM-native and agentic systems resembles the distinction between improving a coding model and improving the harness around it. A stronger model can make a better decision in one pass. An agent can call models again, critique the proposed recommendations, and revise them. Tandon presents that iterative approach as an emerging research direction rather than an established replacement for today’s feeds.

The LLM recommender recipe begins by creating a language for the catalog. Content becomes tokens; a base LLM learns to understand those tokens alongside English; recommendation training teaches it what to select for a user. The eventual output can directly identify items in the content corpus. Follow the recipe at 8:02.

Tandon expands this into a five-layer cake: semantic IDs, a base foundation model, pre-training, post-training, and light fine-tuning for a particular product surface. The base can be an open-weights model or an internal model. Most training compute is shared across products, so each surface does not need a model trained from scratch. Surface-specific adaptation sits on top of a common understanding of language and content.

Why use semantic IDs? Two properties matter independently:

• Learnable structure: An arbitrary hash gives the model an identity to memorize. A semantic representation gives related content shared structure that the model can learn over. Tandon reports that teams have seen benefits even by replacing hash IDs with semantic IDs inside traditional recommenders.
• Compression: User histories contain many items. In Tandon’s example, representing a three-minute Instagram Reel directly would take roughly 10,000 tokens; compressing it to about 10 tokens leaves room to consider a much longer sequence of interactions. The compact ID identifies the content rather than carrying its full audiovisual detail.

The tennis example makes the structure concrete. Similar Reels share their first three semantic tokens, while a final token distinguishes individual videos. Tandon suggests imagining the first token as sports and the next two as tennis. That is an intuition for the shared prefix, not a claim that every learned token has a clean English label. The useful relationship is that similar items share part of their representation while remaining individually addressable.

Pre-training bridges the catalog’s semantic language and English through two complementary tasks. Follow the training examples at 10:42.

• Connect an ID to a description: A prompt associates the semantic ID ABC with a video description, and the model produces text describing an iconic Wimbledon shot. Synthetic natural-language descriptions teach the connection between catalog tokens and content meaning.
• Predict missing history items: Parts of a user’s semantic-ID sequence are masked. Predicting them teaches the model which videos tend to be watched together and in what sequence.

Post-training turns that understanding toward recommendation decisions. In the reranking example, the input includes user information and 30 candidate videos; the model chooses the top five recommendations. Its generated reasoning discusses topic interests such as comedy, food, DIY, and wellness, alongside engagement style and affinity for particular creators. Selection therefore combines several kinds of user evidence rather than matching only a topic label.

A bilingual model can express a rationale in English while selecting catalog items. That makes its stated reasoning inspectable: a developer can read how the model describes the user and compare that description with its choices. The example establishes access to generated reasoning; it does not establish that this text faithfully explains every internal cause of the ranking decision.

Once English and catalog tokens share a model, language becomes a way to steer the feed. Instagram’s Your Algorithm example lets a user inspect the interests the system associates with them, add or remove interests, and communicate in natural language while consuming Reels. Likes and comments still supply behavioral evidence, but a user can also state a goal directly. Follow the steerable-feed example at 12:18.

The concrete change is an added interest in following the FIFA World Cup. The model receives both the existing user history and that new input, then decodes personalized recommendations; Tandon’s example points to a Messi clip. The new interest supplements the history rather than replacing it. This is how an explicit, timely request can change what the recommender selects while keeping personalization.

Where does the user’s request enter the recommendation path? The diagram shows the two inputs meeting at the model. The important change is that the user no longer has to communicate an interest only by accumulating interactions. A language input can join the history immediately. Tandon connects this direction to prompted playlists from Spotify, custom feeds from YouTube, and Ask DoorDash.

The ending returns to the business loop, now comparing content feeds with AI chat apps. Both invest in training, serve inference, earn engagement, and monetize it. Two questions determine how well that loop works: how effectively does training compute improve the model, and how effectively do inference tokens turn into engagement and revenue? Follow the economics comparison at 14:03.

The key economic difference is what the model must output:

• Content feeds: The output is a semantic ID that points to an existing item. Creators have already supplied the video or other content. Model inference selects an address, and that selection can lead to a substantial period of consumption.
• AI chat apps: The model generates the content itself. In Tandon’s comparison, a turn produces a few thousand tokens, each generated at inference time.

Model size adds another difference. Tandon describes feed models as using roughly 1–10 billion active parameters and chat models as using 10–100 billion. Together with the shorter output, this supports his estimate that content feeds can spend up to 100 times or more less on inference per hour of consumer engagement. The recording does not supply a benchmark methodology for that ratio, and its scope is inference cost—not the full cost of operating a feed, hosting media, or sustaining a creator ecosystem.

This is the basis for the title’s forecast. Tandon’s closing comparison places content feeds in four of the ten most-used apps by daily active users: recommendation already operates inside enormous consumer products. LLM recommenders could improve the engines behind those products while adding steerable feeds, explanations, and more direct user control. His expectation of new consumer and recommendation agents over the following year remains a forecast; the concrete opportunity is to combine language-based direction with the token efficiency of selecting existing content.