06-reference

stratechery eric seufert interview models ads ai upside

2026-05-28·reference·source: Stratechery·by Ben Thompson

"An Interview with Eric Seufert About Models and Ads, and AI's Upside for Humanity" — Ben Thompson with Eric Seufert

Stratechery Interview, 2026-05-28. Seufert (founder of Mobile Dev Memo) just defended his Harvard Master's thesis in Applied Computation. Interview covers his thesis (DeCANT), Meta's communication problem on ads, why agentic checkout flopped, Google's Ship-of-Theseus search transformation, ChatGPT ads, Apple/Amazon/AppLovin positioning, and his "Prosperous Society" thesis (a rebuttal to the Citrini AI-bubble short report).

Why this is in the vault

Seufert is Thompson's most-frequent interview subject and the vault's most-cited operating analyst on ad-tech economics (see existing notes on the Jan 2026 "Ads in ChatGPT" piece and the May 14 MoffettNathanson talk). This episode is unusually load-bearing because it (a) introduces a novel technical framework — DeCANT — for surviving the black-box ad-platform endgame, (b) lays out the "generative RecSys" architectural shift (TIGER paper, RQ-VAE, codebook tokenization) that reshapes how Meta and Google rank both ads and content, and (c) ties the "pro-advertising" stance directly to AI optimism via the four-part "Prosperous Society" podcast — a counter-thesis to the degrowther bubble argument. All three threads bear on Sanity Check editorial framing and on RDCO's reading of the Meta/Google/OpenAI ad-platform cycle.

The core argument

Seufert's master thesis (DeCANT — Deep Creative Attention-based Network for pre-Testing): The big ad platforms have converged on end-to-end automation; advertisers have at most two control levers (creative and objective function), and the system is a complete black box. But you can treat the platform as a teacher model and distill a student model via behavioral distillation: feed your creative + minimal context (country, language, date) and predict ROAS before spending real money. Empirically, a deep neural net beats the XGBoost baseline. Advertisers typically burn ~10% of budget on testing with no ROAS accountability — DeCANT-style pre-testing claws that back. Platforms might be hostile because that wasted testing spend was effectively free money, but they can't really stop it. Brand marketers like it (control of brand messaging across platforms); DTC performance marketers like it (fire-hose creative volume without blowing up testing budget).

Meta's communication problem: Zuckerberg is "not a great communicator" on ads. The last great Meta ad-evangelist was Sheryl Sandberg, who told advertiser anecdotes on every earnings call. Current COO Javier Olivan is invisible to investors. Meta is making historic investments in foundation models for ads (GEM — Generative Ads Model — is "a foundation model, not a tool", with ~40 distilled point-solutions downstream per the SUM paper), and Zuck talks "Superintelligence" instead of the ads-platform story that would actually justify the capex. Thompson agrees and frames himself as "a much bigger public advocate for Meta's ads" than Zuck is.

Agentic commerce as a mirage (third-party flavor): Seufert's "Agentic Commerce Is a Mirage" piece predicted exactly the ChatGPT Instant Checkout flop. Third-party independent agentic commerce doesn't align with platform incentives, AOV is lower, conversion is worse (Walmart publicly fired OpenAI on it), and affiliate-fee economics are inferior to bid-based advertising. In-platform agentic commerce (Amazon's Rufus, Walmart's Sparky) is the working pattern.

Google's Gambit (Ship-of-Theseus search): Google needed to transform search from a click-distribution mechanic into an engagement sink without breaking what users recognize. AI Overviews → AI Mode → finalized chat-like search. Done incrementally so users weren't disoriented. Worked: search revenue +19% YoY on a huge base, monetizing at parity, queries up. Seufert thinks keyword-based search is going away; de-tethered conversational context lets you "use the whole buffalo."

ChatGPT ads — moving fast, but late: OpenAI hit ~$100M annualized in three weeks and is targeting $1B in 2026. They just launched CAPI, pixel, and self-serve. Currently CPC-only — no conversion-event optimization yet — but Seufert thinks that's months away. The risk: Google set the tone for chatbot ads first (UCP / Universal Commerce Protocol), and now OpenAI is playing on Google's field. Amazon's rumored ad-serving deal with ChatGPT (post-AWS-OpenAI investment) would give ChatGPT a 90%-of-US identity spine to ride.

Apple / AppLovin notes: Apple is heavily ad-revenue-dependent via the Google TAC ("ad-washing"), runs a 3B-param on-device LLM, and would have to subvert its own privacy messaging to build a frontier model — best play is to keep paying Google ~$1B/yr for Gemini-in-Siri. AppLovin launched Gist (a Pinterest/RedNote hybrid social app) as an experiment; CTV (via Wurl) is probably the bigger expansion path.

Generative RecSys revolution (the load-bearing technical claim): The industry is migrating off the 2016 DLRM "two-towers" architecture (user-tower + content-tower + dot-product similarity) toward generative recommendation. Google's 2023 TIGER paper used RQ-VAE to build hierarchical codebooks: a billion-item catalog compresses into ~1,028 tokens across 4 codebooks of size 256, which an LLM then mid-trains on for next-token prediction over content sequences. Meta, Google, and Kuaishou are all rebuilding around this. The economic implication: owning your own frontier LLM becomes a defensive moat for ad-ranking because (a) you don't want competitors mid-tuning on your distribution, and (b) frontier-model improvements may be hoarded internally ("Mythos" is the current example).

The Prosperous Society thesis (the editorial centerpiece): Seufert wrote a four-part podcast series as a direct rebuttal to the Citrini "2028 GIC" AI-short research. The argument inverts Galbraith's The Affluent Society: AI's output explosion shifts the binding constraint to distribution, but digital-ad allocation mechanisms are now extremely efficient at matching. You get (i) more granular products, (ii) more efficient matching, (iii) smaller SMBs onboarded to the ad economy that previously couldn't operate it. The optimistic case is "we double the 83% of retail that isn't online and make all of it more efficient." The bear case isn't endless atomization or social alienation — Seufert thinks consumers will reject those — it's that the foundation-model capex doesn't pay off, the bubble pops, and we get a recession. Thompson's amen: "People who like ads are optimistic about AI" — the Silicon Valley blind spot on advertising correlates with the failure-of-imagination view that human wants are finite. They're not; they're "combinatorially infinite" (per Thompson's Divine Discontent piece).

Mapping against Ray Data Co

Three direct hooks for Sanity Check editorial framing:

  1. "Pro-advertising correlates with AI optimism." This is a sharp, contestable, original frame Seufert and Thompson land in real time. It's the kind of one-line synthesis Sanity Check can stress-test against a different industry vertical (data engineering / agent ops). The corollary: the degrowther mindset on AI ("just slop", "useless cycles") shares architecture with the Silicon Valley anti-ads stance. Worth a Sanity Check column that asks: where's the data-team version of this blind spot? Probably "AI agents won't actually do useful FDE work" — same shape.

  2. Generative RecSys is the leading edge of foundation-model verticalization. This matters for the active Markov capital-cycle thesis (chip-fab/memory cycle, Phase 2). If Meta, Google, and Kuaishou all rebuild ad-ranking on LLM-shaped architectures and don't release the best ones externally, that's a demand-tail story for frontier compute that extends well past "OpenAI training run N+1." Cross-check against the existing power-cycle backtest evidence.

  3. DeCANT-style behavioral distillation is a pattern RDCO should internalize for its own ads work. RDCO runs Squarely user acquisition on Meta. Pre-testing creative against a distilled model of Meta's ranking before spending the testing budget is plausibly the highest-leverage tooling move on the Squarely paid-ads side. Cross-link with [[2026-01-31-claude-code-autonomous-meta-ads]] (Giorgio Liapakis pattern — Claude Code running Meta ads autonomously) — the natural composition is: Claude orchestrator + DeCANT-style local model + Meta's CAPI. Adds to the paid-ads skill capability surface.

RDCO positioning notes:

Related