How can ultra-long sequences of events be evaluated efficiently for content recommendation? ByteDance released a paper earlier this month, accepted for presentation at RecSys '26, that introduces SequenceO1, a recommendation model architecture that incorporates ultra-long sequences of up to 100,000 user events. The authors accomplish this through what they call Sketch Attention. Rather than constructing a single user vector from an event sequence, Sketch Attention compresses a very long stream of event embeddings (up to 100k) into 1,024 learned representations of user activity, or a "sketch." The model uses 1,024 learned prototype vectors, trained end-to-end with the ranking model, to generate user-specific sketches through cross-attention against each user's event history. Notably, the attention weights are normalized across prototypes for each event, so every event distributes its contribution across the summary slots. The sketch can be re-used across candidates contemporaneously, so ranking is request-centric rather than candidate-centric. This is similar conceptually to what Meta discusses in its Adaptive Ranking Model blog post from a few months ago: instead of processing full attention across the candidate-conditioned user embeddings, it runs a lighter-weight comparison on the pre-processed sketch to produce the candidate scores. SequenceO1 then utilizes the raw embeddings of the user's most recent 10,000 events in a separate branch that produces a candidate-conditioned user embedding, and the candidate-conditioned outputs of both branches are fused in MixFormer, the downstream ranking backbone. Both branches (ultra-long and 10k) are processed with Stacked Target-to-History Cross Attention (STCA). The sketch is cached with a 1-hour TTL for inference and a 3-hour TTL for training. The authors observed statistically significant improvements to engagement metrics on Douyin and Douyin Lite across the board, including +2.33% for video finishes and -6.98% for dislikes on Douyin. And SequenceO1's ultra-long history branch was estimated to require 50x fewer training-side FLOPs and 64x fewer inference-side FLOPs than computing the full 100k-event sequence directly with STCA. Paper linked below.

What if everything turns out alright? Among the many disastrous consequences predicted of artificial intelligence is a jobs famine: widespread worker displacement that devastates the economy and brings about a second Great Depression. But that hasn’t happened. In fact, the US economy is doing well and, with a 4.1% unemployment rate, operating roughly at full employment, despite other headwinds like persistent inflation. And while whether humanity is “early” in the adoption phase of artificial intelligence is debatable — I argue that, certainly, from an exposure standpoint, AI is a familiar technology at this point — the notion that AI has displaced large swaths of the workforce is not supported in the current data. Given the buoyancy of the economy, should we reduce the odds of that happening? We can’t definitively say that it won’t, of course, but we can use the data at our disposal now — in September 2026, nearly four years after ChatGPT’s launch — to update our likelihoods. And one likelihood that I believe is underestimated, sitting somewhere between The Prosperous Society and total economic Armageddon, is that everything turns out alright. mobiledevmemo.com/what-if-ev…

Agentic commerce is a mirage (part 4) "Amazon has blocked Muse, Meta’s new personal agent, from accessing its platform. Amazon and Walmart are motivated by very precise, very concrete direct and indirect financial incentives to resist surrendering the customer relationship to third-party shopping agents. Incentives matter: they are authoritative, and they are dominant. “New interaction paradigm” doesn’t, by default, override prevailing incentives even when it’s superior to the status quo; in the case of third-party agentic shopping, that new interaction paradigm is not superior to the status quo for retailers." mobiledevmemo.com/agentic-co…

I've recently added two features for Mobile Dev Memo Pro subscribers: natural language search for app store chart rankings and audio versions of new MDM posts. Natural Language Chart Search MDM Pro subscribers can now search App Store and Google Play chart rankings in plain English through Darkpool, the intelligence platform I've built for Mobile Dev Memo. Darkpool supports unified searches across geographies, app stores, and categories. For example: "Find apps in the United States across iOS and Android that currently rank in the top 20 but ranked 50th or worse at some point in the last 30 days. Include Games and all games subcategories, Social Networking, Health & Fitness, and Utilities." I originally built a SQL-like search interface for this, but I may have overestimated how many people want to write SQL queries. Now you can simply describe what you're looking for and get the results. The attached screenshot shows this in action with the prompt above. Note that this is currently a beta product and may be rough around the edges. Posts as Podcasts Subscribers can now listen to new Mobile Dev Memo posts read aloud (in my voice, courtesy of ElevenLabs' voice model tool). Each subscriber gets a personal, private podcast RSS feed to import into their favorite podcast app(s), including Spotify. If you don't have time to sit down and read a post, you can listen during your commute. Personal MDM Pro subscriptions are $8.99 / month. I also offer discounted annual and multi-seat subscriptions.

Can bottom-up clustering improve performance in generative retrieval? A new paper from researchers at PayPal explores whether creating semantic IDs (SIDs) through a bottom-up clustering process can better capture local relationships between similar items while also ensuring that each item receives a unique identifier. They implement a bottom-up approach that contrasts with that employed by Google's TIGER: instead of running RQ-VAE to encode item embeddings through successive rounds of residual quantization over K codebooks, they start by building local item neighborhoods, then iteratively construct the catalog hierarchy more coarsely. The initial clustering of catalog-item embeddings uses MiniBatch K-means, with each item receiving a unique, cluster-level integer label as its final SID token. The cluster centroids are then grouped into broader clusters, weighted by item count, and so on (note that the main text describes this merging stage as agglomerative clustering, although the Appendix specifies MiniBatch K-means). This method really just replaces the tokenizer of a generative retrieval system; it doesn't dictate the method used for subsequent retrieval (eg., through a Transformer sequence model). The authors found that their bottom-up approach improved Recall@10 (rewards whether the relevant item appeared) and NDCG@10 (also rewards item placement) on both the Amazon Beauty and a proprietary dataset, although it performed slightly worse / the same on Amazon Sports & Outdoors. I wouldn't consider the paper conclusive, but it does suggest that building SIDs around local item similarity can improve recommendation performance. One somewhat pedantic nitpick: the paper presents RQ-VAE as the diametric opposite of the implemented approach, which it isn't: RQ-VAE clusters *residual vectors*, it doesn't create clusters within a cluster, even though progressively longer SID sequences define increasingly fine-grained groups. Whereas this method does cluster existing clusters. So the distinction isn't *quite* top-down versus bottom-up. Link in comments. It's a short paper at just six pages and a fairly easy read.

ChatGPT introduces conversational ads and a Shopify integration "Sponsored Agents could potentially neutralize any risk — what I’ve called a 'persistent superstition' — that ads in chatbots are deemed hostile by users to their interests. By the point a conversation has been transitioned into a Sponsored Agent, the user should no longer expect an unbiased, objective response from the chatbot — because the user is now interacting with a clearly identified, branded agent. I’ve argued repeatedly that this potential to erode user trust in the credibility of responses isn’t a fundamental property of advertising in chatbots, but rather would result from haphazard or undisciplined product design. The Sponsored Agents format appears to be an attempt to mitigate that risk." mobiledevmemo.com/chatgpt-in…

Revisiting the impact of Siri AI on the app economy "While I still don’t believe that direct app usage will be broadly supplanted by Siri AI, particularly where the experience of using the app is itself the source of value (eg., gaming, social media, streaming), the improved Siri AI experience does hint in the direction of a more centralized (in Siri) functional gateway to the user. The consequences of this might be material for apps whose economics depend on browsing and discovery (and, by extension, advertising). It remains to be seen how much of that engagement those service-oriented apps retain when Apple provides an increasingly capable interface for accessing that functionality." mobiledevmemo.com/revisiting…

Meta’s AI opportunity, whatever the pace "Squint hard enough, and a platform becomes visible: cross-device agentic tools running atop Meta’s owned models, with an interaction surface area that includes WhatsApp, used by at least 2BN people every day. If 'platform' in the digital context is characterized by being resistant to intermediation, or at least appreciably fortified against it, then this connected network of tools and services qualifies." mobiledevmemo.com/metas-ai-o…

This notion that ads are inherently inimical to consumers in a chatbot experience and thus the affiliate model will win out over ads is something akin to a persistent superstition. It ignores retailer incentives and defies reality at this point: Amazon is *suing* Perplexity for buying products autonomously whereas it invested $50BN into OpenAI and just announced an ads partnership with the company *this week*. OpenAI shuttered Instant Checkout after a matter of months; its ads business has reached $1BN in projected annual revenue. The affiliate model may play some small role in agentic commerce, but it’s economically suboptimal relative to ads for an agent / chatbot, the inflexibility of the fixed percentage fee will price some retailers out, and it doesn’t capture variation in retailers’ willingness to pay, which can be a quality signal. Anecdotally, I haven’t seen affiliate networks associated with “high quality product discovery.” Any mistrust between a user and their agent as a result of advertising is a design issue. The exact same argument could have been made about Search; certainly it’s true for affiliate links. If an ad platform is sophisticated enough, and it serves enough advertisers, then ads are the preference-matching mechanism that provides the best economics to agent and platform operators. mobiledevmemo.com/affiliate-…

Spotify published something of a spicy take on its engineering blog this week that made for an incisive and thought-provoking Saturday morning coffee read: "Why Spotify Is Not Using Bayesian A/B Testing" I usually roll my eyes when I see blog post titles of this nature, since a) this discussion has been had ad nauseam, and most contributions rehash well-worn arguments, and b) methodology choice is highly context-dependent and doesn't generalize well. But this blog post is (very) thorough, and the context in this case, a household-name product with ~800MM MAU, is compelling even if the lessons don't apply universally. The post walks through a number of reasons why Spotify's A/B testing needs are met through a frequentist process, and why introducing a Bayesian process wouldn't add sufficient value. The two that I found most powerful, which really relate to mistakes in implementation and not choice of methodology, and which I've empirically seen to be the case: - The winner's curse is only ameliorated in the Bayesian approach with a well-selected prior. But teams often implement a uniform prior, which provides no shrinkage (pulling noisy uplift estimates toward more plausible values), and some misspecified priors actually performed worse than the group sequential testing (GST) baseline. And even an oracle prior, while improving effect estimation, showed no power advantage over GST. - Forming an informative prior from past experiments is challenging because it requires many historical experiments (the blog post suggests >200, which strikes me as high) and for the much more prosaic reason that the world changes over time and historical results may not be representative. That second reason seems obvious, but it impacts experimentation in subtle but substantive ways: a consumer product's DAU mix can evolve meaningfully over time by *cohort age* and provenance (organic vs. paid), and *new users* may possess dramatically different motivations / intent / product knowledge over time. The post states that "an experimenter using flat-prior Bayesian inference with posterior-probability thresholds is running frequentist-equivalent inference under a different vocabulary" because: - posterior mean equals MLE - one-sided posterior P( B>A | D ) = 1-p-value - "95% posterior-probability threshold is algebraically equivalent to the one-sided rejection region" The post does concede that Bayes provides superior interpretability, but in my experience, product leadership is generally comfortable with the concept of confidence intervals (vs. Bayesian credible intervals), so the benefit may be more muted than ~10 years ago. Link to post and the accompanying paper below.

Meta is cross-promoting Muse, its new agent app, from Facebook. This is the same approach Meta used to grow Threads, which reached 500MM MAU in June, less than three years after launch. Meta introduced ads to Threads at roughly 400MM MAU.

Amazon brings its ad machine to ChatGPT In December 2025, I predicted that Amazon would bring its advertisers to ChatGPT. Today, it announced such a partnership. I think this was foreseeable at that time for three reasons. First, Amazon had already pursued a similar model with Meta and Snap, and the ChatGPT product experience is arguably even better suited to it. Second, I viewed the affiliate model that OpenAI had implemented in ChatGPT with Instant Checkout as economically inferior to advertising at scale. I expected the company to embrace advertising, which it did shortly thereafter. Third, these integrations create opportunities for Amazon to project its immensely valuable first-party data onto third-party surfaces. Given ChatGPT’s scale and the commercial potential of advertising in chatbots, Amazon would be remiss in not pursuing such a relationship as part of its broader investment. mobiledevmemo.com/amazon-bri…

The CAPI revolution The Conversions API, or CAPI, represents a revolutionary development in the digital economy: it allows any platform with sufficient scale to extend its advertising scope beyond local context into the broader and more commercially valuable function of behavioral advertising, entirely outside of the purview of a browser or operating system. In this episode, I provide a history of the CAPI and its role in the Everything is an ad network phenomenon. I also explore the fundamental shift in the digital advertising landscape driven by the transition from browser-side pixels to the server-to-server CAPI, moving beyond the immediate impacts of Apple’s App Tracking Transparency to a more permanent structural change in how commercial data is shared and utilized. mobiledevmemo.com/the-capi-r…

The total lack of any contemporary AI-related job cataclysm does seem to repudiate the last of any cogent negative AI doom thesis and point to asymmetric upside on AI-related investments: 1) "AI will eliminate white-collar employment." This has demonstrably not happened; the most recent jobs number surprised to the upside, and job growth has generally been strong this year. The Citrini thesis was utterly dismantled in a counter-report published by Citadel, which is absolutely worth reading, but the data is now fairly clear, at least in the short term. "But we're early!" Not really; ChatGPT was first released in November 2022, so we're approaching four years of broad commercial and consumer access to LLMs. After nearly four years, the onus of support shifts to the downside. 2) "AI is not delivering productivity gains, and all of the capex spend is being wasted." So what? Meta and Alphabet are among the largest, publicly traded spenders on AI capex. Meta trades at ~19x forward P/E and Alphabet trades at ~25x. Comparred to Microsoft at ~61x (1999 peak) and Cisco at ~131x (2000 peak) during the dotcom boom, the wealth effect risk seems tolerable / palatable relative to the upside. Meta and Alphabet -- the vanguards -- aren't priced for a utopian AI interpretation, so the positive EV accumulates on the bull scenario. Besides, we're already seeing meaningful revenue growth in digital advertising platforms, which are ground zero for AI efficiency gains, given that a) they invest most heavily in AI research (Google: transformers, Meta: PyTorch etc.,) and b) ad spend is a leading indicator of economic activity. So the "0 productivity gains" argument is prima facie flawed: increased conversion rates on digital ads purchased by SMBs are defensible, provable productivity gains. But even if that isn't true: we may be witnessing the modern-day, private equivalent of the New Deal / PWA / WPA. Why is that a bad thing? 3) "Capex spending is crowding out investment into more durable, valuable projects!" Even if that's true, the counterfactual is not 0 investment; it's whatever was second on the list of most appealing places to deploy capital. Why wasn't that compelling enough to supersede AI capex? Name the investments being displaced, quantify their expected returns, and explain why they beat AI. "Something more useful" is not a capital-allocation thesis, particularly for Meta- and Alphabet-scale advertising platforms, or OpenAI and Anthropic. The floor of the downside arguments seems to be constantly drifting upward. Even if we accept the worst of the possible outcomes (putting aside the sci-fi variety), the more plausible outcome is, increasingly, The Prosperous Society.

Can a personalized fusion of funnel predictions improve search ranking? Interesting new applied ML RecSys paper from researchers at Alibaba’s Taobao and Tmall, two of the largest e-commerce marketplaces in China. Traditional search ranking encounters a natural tension: the system ranks individual items for individual requests, but the business ultimately seeks to optimize cumulative user-level outcomes like GMV over some time horizon. That misalignment is often addressed by manually fusing conditional predictions across the funnel (like clicks, cart-adds, purchases, and transactions, etc.) into a single ranking score, eg., pGMV = pCTR × pCVR × E[ transaction value | purchase ]. But this approach risks mistaking predictive association with causal effect: purchase intent informs every step across the funnel, so no given prediction necessarily maps to the GMV impact of the ranking. The paper proposes Direct Causal Effect Optimization (DCEO): a framework that learns a context-dependent item-level proxy by optimizing for its estimated local causal effect on the long-term objective rather than its predictive association with that objective. DCEO is implemented as an actor-critic layer atop 17 existing upstream prediction scores acroiss the conversion funnel. Given user and request features, the actor generates softmax-normalized weights over those predictions and forms their convex combination as an item-level proxy score; these scores are then aggregated at the user level and calibrated to a fixed impression count. The critic estimates the long-term outcome conditional on user features and the aggregated proxy, and the actor is trained to maximize the critic's estimate of the change in that outcome under a hypothetical 5% increase in the proxy, with a conditionally normalized ranking loss serving as an additional stabilizer. At serving time, only the lightweight actor is required, and its learned proxy enters the existing multi-objective ranking formula as an additional term. I found two insights to be particularly interesting. First, extending the GMV horizon from one to four days increased the weight on impression-to-click probability, which the authors interpret as preferencing exploration (vs. immediate conversion) when optimizing over a longer horizon. Second, a personalized mixture of predicted outcomes across the entire funnel was better aligned with long-term GMV than predicted GMV alone, suggesting that no single terminal metric captures all of the behaviors that generate long-term commercial value. In a 41-day A/B test, DCEO increased GMV by 0.36%, clicks by 0.36%, and purchases by 0.12%. As an architecture, the DCEO approach is attractive because it fuses the point models in an existing prediction stack rather than displacing it with an end-to-end rebuild, as many deep learning RecSys models require. Paper linked below.

Netflix’s YouTube opportunity, part 3 "Netflix’s opportunity, as I have described it across the series, is to bring vetted, proven, broad-based content portfolios from YouTube to its own platform with materially less audience or quality risk than ordering new, unproven content. The strategy has seemingly taken shape across genres and demographics, and it represents one leg of a multifaceted approach to expanding Netflix’s content base, which also includes short-form video and podcasts, without proportionately scaling its content expenditure. But the strategy is zero-sum, since attention is rivalrous: the content will be viewed only once, on one platform. This explains YouTube’s reaction: Netflix’s YouTube opportunity is YouTube’s loss, and incentivizing creators to stay — and punishing those who leave — is a sensible response." mobiledevmemo.com/netflixs-y…