Devansh Tandon, who leads Meta Recommendations Research, argues that recommendation systems scale like LLMs and that LLM-based recommenders will become one of AI's biggest consumer applications. He frames feeds as far cheaper per hour of engagement than generative chat apps because they decode pointers to existing content instead of generating every token. The talk covers four S-curves of recommendation (traditional recsys, LLM-inspired, LLM-native, agentic), a three-step recipe for building an LLM recommender — semantic IDs for tokenizing content (compressing a Reel from ~10,000 to ~10 tokens), pre-training the model on natural language plus the catalog, and post-training as a re-ranker with chain-of-thought reasoning — plus steerable and explainable feeds such as Instagram's editable algorithm view, and the economics of feeds versus chat apps. # Video Identity Observations ## Explicit Self-Description (Creator Claims) - **Speaker affiliation**: Devansh Tandon states, "I currently work at Meta on research and product. Uh I lead a team called Meta Recommendations Research." He also notes prior experience: "before this I was at Google for a long time uh working on a lot of the key ML teams including deep mind and YouTube." - **Prior speaking history**: He references a prior talk — "Last year I gave a talk at AI engineer called teaching Gemini to speak YouTube" — indicating recurring presence on the AI conference circuit (this video is from "the LM Rexus track"). - **Research authorship claim**: He claims authorship of papers on "semantic IDs and generative retrieval which we also wrote two papers about which I've linked here." - **Central thesis claims**: - "recommendation systems scale just like LMS do and that the field is very early in that scaling curve" - "the LM recommener is going to be one of the biggest consumer applications of AI" ## Observable Presentation Evidence ### Intended Audience - A **technical ML/engineering practitioner audience**: the talk assumes familiarity with concepts like scaling laws, two-tower models, pre/post-training, entropy/AU metrics, and references "AI engineer" as a venue. - Audience framed as **industry peers**, evidenced by phrases like "I've seen exciting launches and papers from YouTube, Meta, Spotify, Door Dash across the industry" and "we'll be hearing more from speakers about these" — indicating a multi-speaker industry conference track. ### Claimed Scope of Content - **Four announced sections**: "Recommendation scaling curves, a framework of four, recommendation paradigm scurves... sharing the recommener recipe and finally this consumer AI app framework." - Covers: scaling law parallels between LLMs and recommenders, four paradigm S-curves (traditional → LLM-inspired → LLM-native → agentic), a five-layer "recipe" for building LLM recommenders, and a token-economics framework ("flywheel of tokens in engagement"). ### Demonstrated Presentation Patterns - **Insider evidence from Meta**: Cites internal/product examples — "Instagram reels had a strong quarter, 30% year-on-year watch time" from "Meta's recent earnings reports," plus papers "HSTU from 2024" and a 2025 followup. - **Named-paper citations throughout**: TIGER, PLUM, HSTU, OneRec, the 2020 OpenAI scaling law paper ("when Daario was still at OpenAI"). - **Framework-building style**: Constructs reusable abstractions (four S-curves, five-layer cake, flywheel, token-efficiency comparison) rather than only reporting results. - **Speculation clearly flagged as opinion**: Uses hedged language for forward-looking claims — "I think this is where the leading recommener systems in the industry are largely operating today," "I think the next scurve will be..." - **Quantitative argumentation**: Compares "content feeds" vs "AI chat apps" on parameter counts ("1 to 10 billion" vs "10 to 100 billion active parameters") and cost ("up to 100 times or more cheaper"). ### Presentation of Industry Context (Not Self-Praise) - Claims industry adoption of his prior work's ideas: "the ideas behind semantic ID and generative retrieval have really taken off in the industry over the last year" — a creator claim about impact, not directly verifiable from this material. - Notes competitor/peer activity neutrally: "prompted playlists from Spotify, custom feeds from YouTube, Ask Door Dash." ## Notes for Synthesis - This is a **conference talk** by a named industry researcher, not a typical channel-produced video; channel identity evidence is limited to the venue framing ("LM Rexus track") and the speaker's self-description. - No curated metadata was supplied, so all observations derive from title and transcript only. - The speaker consistently distinguishes **demonstrated results** (Meta scaling papers, earnings metrics) from **personal predictions** ("I think..."), which should be preserved as separate evidence classes. # Delivery Observation — "Why LLM Recommenders Will Be AI's Biggest Consumer App" ## Speaker & Context The speaker is a Meta research/product lead presenting a technical conference talk. Self-identified background: "I currently work at Meta on research and product. Uh I lead a team called Meta Recommendations Research." He references a prior talk: "Last year I gave a talk at AI engineer called teaching Gemini to speak YouTube." --- ## Recurring Framing Devices ### 1. Two-part thesis announced up front, then structurally honored He opens with explicit signposting: "I want to make two big arguments today. The first is that recommendation systems scale just like LMS do... and the second is that the LM recommener is going to be one of the biggest consumer applications of AI." The talk then follows this map closely, ending with a return to the second argument. ### 2. "Tokens in, engagement out" as a repeated anchor phrase This phrase appears at the talk title level ("tokens and engagement out"), mid-talk as a named concept, and again near the close: > "I want to introduce this idea of a flywheel of tokens in engagement out which is what's powering all of these Rex model scaling." > "finally, I want to talk about kind of the framework of tokens in engagement out this flywheel that I started with." The phrase functions as both title and conceptual glue — the closing section explicitly loops back to material from the opening ("this flywheel that I started with"). ### 3. Cycle of previews and callbacks Each section transitions with an explicit roadmap. Examples: - "This year I want to talk about four uh sections. Recommendation scaling curves, a framework of four..." - "Let's go a bit deeper. This is the LLM recommener as a five layer cake. Um, we'll start at the bottom." - "I'll go a bit deeper into each stage." ### 4. "Let me shift gears" — section-boundary markers Used to signal topic changes: "Let me shift gears a bit to share the recipe of how to actually build a LLM recommener." Also lighter variants: "Um let's start with scaling curves." --- ## Teaching Habits ### Concrete-number grounding He consistently attaches specific figures to claims: - "Instagram reels had a strong quarter, 30% year-on-year watch time." - "a threeminut Instagram real video would be 10,000 tokens... you have to to compress it into about 10 tokens." - "content feeds are significantly cheaper up to 100 times or more cheaper than AI chat apps." - "Four out of 10 of them are content feeds." - Parameter ranges: "around 1 to 10 billion active parameters" vs. "10 to 100 billion." ### Named-paper sourcing Claims are tied to citable artifacts: "a paper HSTU from 2024 and the second is a followup from this year," "The tiger and plum papers are some examples," plus industry examples: "prompted playlists from Spotify, custom feeds from YouTube, Ask Door Dash." ### Metaphors for structure and analogy - The "five layer cake" for architecture layers. - The "flywheel" for the business loop. - Analogy to coding agents: "the LM native models are like improving the core capabilities of the model going from opus 45 to 48 versus the agentic curve will be like improving the coding harness behind claw code or codeex." - "bilingual model" for English + semantic-ID understanding. ### Worked examples from his own products Semantic-ID explanation uses a hypothetical Instagram tennis scenario: "You can see that the semantic token shares the prefix of the first three tokens... You can imagine the first token representing sports um and the second two tokens representing tennis." Later he personalizes an example: "I've added an interest that I want to follow the FIFA World Cup... like this free cake that Messi scored recently." --- ## Tone & Style ### Conversational connectors Frequent soft openers and hedges throughout: "Um, okay. So, quickly about me," "So I wanted to start with," "Let's go a bit deeper," "I'm really excited that," "I think this is super exciting." ### Enthusiasm markers tied to specific claims Enthusiasm is attached to concrete product implications rather than left general: "I think this is super exciting because once you have a model that can understand both English and recommendations, it opens up new product surfaces." ### Forward projection pattern ("I think we're going to see...") Repeated near the close: > "We're going to see a lot of new product experiences come through with steerable and interactive recommendations" > "I think we're going to see some really exciting research on consumer agents and recommendation agents that come out over the next year or so." ### Positioning claims with soft qualifiers Bold claims are typically hedged with "I think": "I think this is where the leading recommener systems in the industry are largely operating today," "I think the next scurve will be this paradigm of LLM native." --- ## Vocabulary Notes - Consistent domain shorthand: "Rex/Rexus" (recsys), "semantic ID" / "SID," "S-curves" (rendered "scurve"), "HSTU," "two tower sparse network rankers." - Transcript shows LLM rendered as "LM" in several places (e.g., "LM inspired models," "LM native," "LM chat") — likely a transcription artifact or spoken shorthand; can't distinguish from audio alone. - A recurring "power law / power loss" alternation appears in the transcript ("power law of scaling" vs. "power loss scaling curve"), also likely transcription-level; flagged for uncertainty. --- ## Stable Patterns vs. Isolated Moments **Stable-looking (repeated across the talk):** - Thesis-then-roadmap structure with explicit signposting - "Tokens in, engagement out" anchoring - Hedge-plus-claim phrasing ("I think...") - Concrete numbers attached to every major argument - Section transition markers ("Let me shift gears," "Let's go a bit deeper") **Isolated (single occurrence, too thin to call a pattern):** - The historical framing aside: "This is when Daario was still at OpenAI and Anthropic didn't exist yet" — a light, humorous aside not repeated elsewhere. - The coding-agent analogy to Opus/Claude Code — a single analogy instance. - Self-deprecating casualness about talk pacing ("Um, okay. So, quickly about me") appears once at the opening. --- ## Confidence Notes - Patterns above are drawn from a full ~17-minute talk transcript, so repetition claims rest on multiple in-text instances. - Pronunciation/transcript artifacts (Rexus, "power loss," "LM" vs "LLM") are noted but not interpreted as deliberate delivery choices. # Tokens and Engagement Out: Training LLM Recommenders — Position Record **Speaker:** Devansh Tandon (Meta, leads Meta Recommendations Research; previously Google/DeepMind/YouTube) **Context:** Talk titled "Tokens and Engagement Out: Training LLM Recommenders," LLM Recommenders track (identified in transcript only). ## Core Arguments 1. **Recommendation systems scale like LLMs** — and the field is early on that scaling curve. 2. **The LLM recommender will be one of the biggest consumer applications of AI.** ## Scaling Curves - Cites the 2020 scaling laws paper (Kaplan-era OpenAI; "Dario was still at OpenAI"): power-law loss reduction with model size, data, compute — this predictable curve underwrote capex investment. - **Recommendation systems follow a similar power-law scaling curve**; before LLMs, RecSys were the largest production ML models in big tech, still serving 1B+ DAU. - Evidence: Meta's **HSTU (2024)** paper and a 2025 follow-up showing offline quality gains with scale. - Product evidence from Meta earnings: Instagram Reels **+30% YoY watch time**, driven by simplified ranking architecture, doubled user interaction sequence lengths, richer interactions. - Introduces the **"flywheel of tokens in, engagement out"**: training → inference → engagement → monetization → funds next training run. Each scaling step is one loop. ## Four RecSys Paradigm S-Curves - **Traditional RecSys** — feature engineering, two-tower/sparse rankers, embeddings; most production systems still here; Tandon predicts this curve won't disappear but agents will automate feature engineering. - **LLM-inspired** — end-to-end scaled models (HSTU, "one Rec"); "leading recommender systems largely operating today" (his assessment). - **LLM-native** — adapt a reasoning base model to recommendations (TIGER, Plum papers); most of the industry is trying to graduate here. - **Agentic** — LLMs orchestrating RecSys in plan-retrieve-rank-critique loops; analogy: LLM-native = improving model (Opus 4.5→4.8), agentic = improving the coding harness (Claude Code/Codeex). Called "an interesting area of research" — tentative. ## LLM Recommender Recipe ("Five Layer Cake") 1. **Semantic IDs** — tokenize content; two motivations: stable representations vs. shifting hashes, and compression (a 3-min Reel ≈ 10,000 tokens must compress to ~10). Similar reels share token prefixes. 2. **Base LLM** — open-weights or first-party. 3. **Pre-training** — bridging English and semantic tokens (e.g., masked sequence prediction). 4. **Post-training** — steering to recommendation tasks; reranking with visible chain-of-thought (user topics, engagement style, creator affinities). Compute shared across product surfaces. 5. **Surface-specific fine-tuning** — light, per product. ## Product Shift: Steerable Algorithms - Shift from **black-box recommenders to interactive, steerable, explainable** ones — e.g., Instagram's "your algorithm" letting users add/remove interests in natural language. - Claims users can direct feeds toward goals "expressed in language rather than just likes or comments." - Industry examples cited: Spotify prompted playlists, YouTube custom feeds, Ask DoorDash. ## Token Efficiency Argument - Content feeds use ~1–10B active-parameter models; AI chat apps use 10–100B. - Content feeds decode **pointers (semantic IDs) to free creator-supplied content**; chat apps must decode every token themselves (~thousands per turn). - Conclusion: content feeds are **"up to 100 times or more cheaper"** per hour of engagement. ## Closing Claim - 4 of top 10 apps by DAU are content feeds; nearly all engagement/monetization growth is driven by recommender and ads models, which LLM recommenders will "entirely transform." Predicts exciting research on consumer/recommendation agents "over the next year or so." ## Qualifications to Preserve - Curve two/two placements ("I think," "largely," "most of") are Tandon's assessments, not settled fact. - The agentic paradigm is explicitly speculative ("I think... interesting area of research"). - Traditional curve won't disappear per speaker. # Speaker Attribution Record — Video 01a15ca7-2e43-7a16-89b4-33c50029d6c4 ## Speakers - **Devansh Tandon** — sole identified speaker; **creator/presenter** role - Self-identified: works at **Meta** on research and product; leads **Meta Recommendations Research** - Prior role stated: previously at **Google** (DeepMind, YouTube ML teams) ## Roles Observed - **Narrator**: none (Tandon delivers the full talk) - **Guests**: none; the speaker references but does not host other speakers - **Audience**: present ("Thank you for coming out"), no speaking role ## Referenced Sources (per transcript) - **Papers / research** - 2020 landmark scaling curve paper (OpenAI-era, per speaker) - Speaker's own prior talk "Teaching Gemini to Speak YouTube" at AI Engineer, plus **two linked papers** (semantic IDs, generative retrieval) - **HSTU** paper (Meta, 2024) and a 2025 follow-up - **TIGER** and **PLUM** papers - **One Rec** paper (referenced as "one wreck") - **Industry claims (third-party)** - Adopters of semantic IDs / generative retrieval: **YouTube, Meta, Spotify, DoorDash** (speaker cites launches/papers) - Products cited: prompted playlists (Spotify), custom feeds (YouTube), Ask DoorDash, "your algorithm" (Instagram) - **Meta earnings reports**: Instagram Reels 30% YoY watch-time growth; doubled user interaction sequence length - Top-10 apps by DAU: "four out of ten are content feeds" (unsourced within transcript) ## Promotional Segments - Speaker promotes **Meta's products and research** (Instagram, Facebook ads, HSTU, earnings results) and previews other track speakers ("we'll be hearing more from speakers") ## Unavailable / Uncertain Evidence - **No curated metadata supplied** — no timestamps, comment availability, or external verification possible - Claims about third-party systems (YouTube, Spotify, DoorDash) are speaker-reported, not independently sourced in this transcript Welcome everyone. Thank you for coming out to the LM Rexus track. Um I'll be sharing the first talk. Uh my talk is titled tokens and engagement out training LM recommenders. Um, and I want to make two big arguments today. The first is that recommendation systems scale just like LMS do and that the field is very early in that scaling curve. Um, and the second is that the LM recommener is going to be one of the biggest consumer applications of AI. Um, okay. So, quickly about me. Uh, I currently work at Meta on research and product. Uh I lead a team called Meta Recommendations Research uh which is this group that's training frontier models, LLMs and recommenders that power Instagram, Facebook ads, the meta family of apps. Uh before this I was at Google for a long time uh working on a lot of the key ML teams including deep mind and YouTube. Last year I gave a talk at AI engineer called teaching Gemini to speak YouTube about two ideas semantic IDs and generative retrieval which we also wrote two papers about which I've linked here. Um it was really fun and it led to a lot of discussions and collaborations. Uh the ideas behind semantic ID and generative retrieval have really taken off in the industry over the last year. Um and they've moved from more research to now scaled production systems. And I've seen exciting launches and papers from YouTube, Meta, Spotify, Door Dash across the industry. And we have a couple of examples of that later today. This year I want to talk about four uh sections. Recommendation scaling curves, a framework of four, recommendation paradigm scurves that we are climbing as an industry. um sharing the recommener recipe and finally this consumer AI app framework. Um let's start with scaling curves. So I wanted to start with this landmark scaling curve paper from 2020 which feels like a lifetime ago. This is when Daario was still at OpenAI and Anthropic didn't exist yet. But the core idea that this paper shared is the power law of scaling. As you increase model size, data, the amount of compute flops trained uh for model training, the loss falls on this log linear scale. And this clean and predictable curve is what really set off the race for the AI frontier because you can forecast what model quality and capability improvements will look like. And this is what's underwriting the massive capex investments and the AI buildout today. It turns out that recommendation systems follow a very similar scaling law. In fact, before this wave of LLMs, Rexs were the largest production ML models in big tech companies. And they're still some of the largest models that are served at a scale of a billion plus daily active users. And they follow this similar power loss scaling curve. On the x-axis, you have data, compute, and model size. And on the Y ais you would see falling loss or in this chart uh an improvement in recommendation quality. In offline evals it's net entropy or AU gains and then when it's translated to a real production launch it's engagement impact revenue impact at some of the biggest consumer app scale. Here's a real example from Meta that demonstrates these power loss scaling curves. uh there's a paper the first is a paper HSTU from 2024 and the second is a followup from this year both demonstrate that as we scale model size compute and data we see this clear improvement in offline eval of recommendation quality scaling curves aren't just academic research they're driving real product impact at scale for some of the biggest consumer businesses in the world here's a couple of examples I have from Meta's recent earnings reports. Um, Instagram reels had a strong quarter, 30% year-on-year watch time. And, uh, the optimizations we made to improve the quality of recommendations included simplifying our ranking architecture to enable efficient model scaling and longer interaction histories to identify a person's interests. uh we doubled the length of user interaction sequences used for training Instagram and increased the richness of each user interaction. So these are direct parallels to the power loss scaling curves for LLMs. And I want to introduce this idea of a flywheel of tokens in engagement out which is what's powering all of these Rex model scaling. You train a model. You then run inference on it which is the tokens in that recommendation model results in better content recommendations. It drives consumer engagement, daily active users time spent. It translates to monetization and ads or subscription which pays for the next model training run. And so every step on the scaling curve is one loop around this flywheel. And a lot of consumer apps are spinning this core flywheel at the heart of their business. So we have a long way to scale these recommener systems. I want to talk about the four paradigms that I see the industry progressing through. The first scurve was more traditional Rexus where this scurve focused more on feature engineering and user and content embeddings. Most production systems are still sitting on this curve. they're running some type of two tower sparse network rankers scaling the embedding models. I don't think this curve is going to go away but model development here will be accelerated with auto research and things like feature engineering will be handled by agents uh rather than real ML engineers. The next curve is kind of LM inspired models where you are scaling models ideally end to end. HSTU and one wreck papers are examples in this paradigm. I think this is where the leading recommener systems in the industry are largely operating today. I think the next scurve will be this paradigm of LLM native where you adapt a base model that understands and can reason uh and adapt it for recommendation tasks. The tiger and plum papers are some examples of this paradigm. And I think the final paradigm that I start to see emerging is agentic where LLMs will start to orchestrate Rex systems in a loop. I think there's a parallel here with coding agents. So the LM native models are like improving the core capabilities of the model going from opus 45 to 48 versus the agentic curve will be like improving the coding harness behind claw code or codeex. Um and so instead of just having a single forward pass through the recommener, you can imagine a loop where agents plan, retrieve, rank and then critique the recommendations. They can refine them by calling models again and finally deliver the recommendations. This I think is an interesting area of research. Now let me jump into so here here's kind of the framework of all the four Rex paradigms. I think companies are scaling across each of these curves in parallel. Most of recommendations I think lives in LM inspired today and is trying to graduate into LM native. Uh but then a lot of companies are still using traditional models uh and climbing that scurve. Let me shift gears a bit to share the recipe of how to actually build a LLM recommener. I think it's pretty simple. It's three steps. Uh you start with tokenizing your content and creating a language for your domain. Then you want to adapt the LLM so that it understands both English and your domain language and becomes this bilingual model. Finally, you can prompt this model with user information and it will directly decode recommendations from your content corpus. Let's go a bit deeper. This is the LLM recommener as a five layer cake. Um, we'll start at the bottom. Uh, that's semantic ID where you're converting your content corpus into tokens that the LLM can understand and reason over. Then you have kind of the base LLM foundation model. This can be an open weights model or an internal first party model. Then the core training stages. Pre-training is around bridging English and these recommener tokens. Post-raining is about steering the model towards recommendation tasks like predicting engagement or reasoning over recommendations. And then finally, you can just do some light surface specific fine-tuning to deploy it on a product surface. The exciting thing about this paradigm is most of the compute is shared across all of the product surfaces. So you don't have to train individual models from scratch for every product surface. I'll go a bit deeper into each stage for semantic ids. I think this has seen incredible adoption. A lot of teams are just replacing their hash ID with the SID in traditional models and seeing good impact. I think there's two big reasons to tokenize content. The first is it gives you the stable representation for models to learn over rather than a constantly shifting hash that the model can only memorize. And the second is compression. You want to be reasoning over these long sequences of user interactions. And if you don't compress the content, for example, a threeminut Instagram real video would be 10,000 tokens and it will just fill up the content context window too quickly. So you have to to compress it into about 10 tokens. And so here I have some examples of Instagram reels about tennis. You can see that the semantic token shares the prefix of the first three tokens because they're very similar reels. You can imagine the first token representing sports um and the second two tokens representing tennis and then the final token making these videos individual. Once you have a semantic ID, you can train it to understand both English and semantic ID. And so the task I have on the left for pre-training here is an example of where you prompt with a video with semantic ID ABC has the description blank and the output is a shot that was instantly iconic from Wimbledon. Here you're teaching the model to connect these semantic tokens with synthetic English natural language text. The example on the right is about reasoning over sequences of semantic IDs. And so in a user's interaction history, you can mask some parts of the sequence and the model learns to predict them and understand what videos are watched together in sequence. Here I have an example of post training where we're teaching the model how to rerank content. So the input is a bunch of user information and 30 candidate videos that are then ranked to be the top five recommendations from this LLM ranker. What's really interesting here is that you can see the chain of thought reasoning of this model. And because this model knows both English and recommendations, you can simply just inspect the model and understand why it made the decisions that it did. In this example, the model understands the user's topic interests like comedy, food, DIY, wellness. It understands the engagement style and what creators this user has an affinity towards. and then it reranks the content based on this chain of thought reasoning. I think this is super exciting because once you have a model that can understand both English and recommendations, it opens up new product surfaces and new experiences where users can steer their feed. Here's an example from your algorithm on Instagram where users can talk to the algorithm while they're consuming content. It's a screenshot from scrolling through reels or when you click in you can understand what the Instagram algorithm thinks about you and your interests and then you can add or remove interest and talk to it in natural language. And so we're going to see this shift I think from blackbox recommendations algorithms to giving users more control over their algorithm and algorithms becoming more interactive and steerable. I'm really excited that users can direct it towards their own goals that's expressed in language rather than just likes or comments. Um, and I think this can this foundation model can also start to explain its recommendations. And so for this example, I've added an interest that I want to follow the FIFA World Cup at this time. And the model would get both my user history and this new input and then be able to decode recommendations that are personalized to me like this free cake that Messi scored recently. I think these interactive recommenders are going to be a really interesting new product surface and we're seeing this across the industry. We have some examples of prompted playlists from Spotify, custom feeds from YouTube, Ask Door Dash, and we'll be hearing more from speakers about these. Um, finally, I want to talk about kind of the framework of tokens in engagement out this flywheel that I started with um of model training, inference, consumer engagement, and then monetization. This is actually the same flywheel that's shared by content feeds and the AI chat apps. And this is a lens that you can use to evaluate any consumer app. What'll make an app successful on the training ROI side is how well can it translate compute into a frontier model. On the inference side, how well can the inference tokens translate into engagement and then monetization. And so if you try to compare content feeds and AI chat apps, I think that LM recommenders are actually structurally more token efficient than the AI chat. So on the left you have content feeds like Instagram, Facebook, Tik Tok and YouTube. On the right you have the big AI chat apps like Gemini, ChatGpt and Claude. Content feeds are currently using models that are around 1 to 10 billion active parameters. The chat apps are serving much larger models 10 to 100 billion active parameters. Um the output for the content feed is a semantic ID token which is a pointer or an address to existing content because the content supplied on these content feeds is uploaded by creators. There's a very healthy creator economy and so the supply of content is effectively free or it's uploaded by creators. For the AI chat apps, they have to decode every token of content themselves and the amount of tokens output in every turn of an LM chat interaction is a few thousand tokens. Uh, and the the big difference here is that every token has to be manufactured by the app at inference time. And so what that means is if you compare these two apps on how much inference cost and compute is spent to generate an hour of consumer engagement, there's a huge structural gap where content feeds are significantly cheaper up to 100 times or more cheaper than AI chat apps because they're decoding pointers to content rather than content itself. Um, finally I want to kind of end with why I think LM Rexus is one of the most significant consumer AI applications. If you look at the top apps by daily active users, these are the top 10 apps. Four out of 10 of them are content feeds. And so this is a really significant consumer uh application. If you look at the content feeds, almost all of the consumer app growth on both the engagement and monetization side is driven by the recommener and ads models. And these are going to be entirely transformed by LLM recommenders. It's a very large and very token efficient application of AI for consumer apps. We're going to see a lot of new product experiences come through with steerable and interactive recommendations, explanation of recommendations, and just putting more users in control of their experience on these apps. I think we're going to see some really exciting research on consumer agents and recommendation agents that come out over the next year or so. Um and so this is why I think this is a super exciting area of both research and product uh at this intersection of LLM and recommendations. That's all. Thank you so much.