06-reference

stratechery google io world models deepmind

2026-05-20·reference·source: Stratechery·by Ben Thompson
google-iodeepmindworld-modelsai-everywheregeminianthropic-thesisagentic-coding

Stratechery Update — Google I/O, World Models, I/O Spaghetti (2026-05-20)

Why this is in the vault

Five-plus AI-stack signals have stacked up across the last two days (State of AI Dev, Glasswing, Karpathy joining Anthropic, Claude Sandboxes, now Google I/O). Thompson's read on I/O — that Google is committed to AI-everywhere but unfocused, and that DeepMind under Hassabis is pursuing world-models while coding lives elsewhere — is direct evidence on the Anthropic-bet thesis: does Google's strategic shape (world models > coding agents) reinforce or threaten Anthropic's coding-first wedge? Thompson argues it reinforces. Also tags the L4-to-L5 horizontal-COO question at Alphabet scale (DeepMind alignment with the business).

Issue contents

Topic 1 — Google I/O (overview). Thompson uses both Bloomberg and the WSJ to frame the keynote because they covered "entirely different things" from the same event, which is itself the point. Bloomberg led on the search-box redesign and agentic features; WSJ led on Gemini Spark (personal agent), Gemini Omni (text/audio/image-to-video), and Gemini 3.5 Flash (claimed 4x faster, though "early online reaction is pretty negative, however, particularly in terms of token usage, which makes the cost-per-task actually higher"). Thompson's frame: "the company is clearly fully committed to putting AI into anything and everything." Positive spin and negative spin land on the same observation. Verdict: "Google's lack of focus is precisely why it endures."

Topic 2 — World Models (DeepMind alignment). The curious part of I/O was how little of the keynote Demis Hassabis was on stage despite Pichai's promotional tweet hyping him. Hassabis surfaced for Gemini Omni — a "world model" pitched as a step toward AGI ("artificial general intelligence is just a few years away" per Hassabis on stage) — and for the closing AGI/singularity benediction. Thompson's read: Hassabis cares about world models and AGI, not coding or search. Coding at I/O fell to the Antigravity team, not DeepMind.

Thompson's analytical move worth preserving:

Frontier models themselves, however, are also distillations, this time of human thought. What is exposed on the Internet is simply the final output of internal chains of thought that every human engages in; because that is not in the models they too are jagged with surprising drop-offs next to performance that rivals or exceeds the smartest humans on earth.

What fills that generalizability gap is the open question. Two camps:

Thompson's hypothesis: "I think it's possible that the reason Google is widely considered to be behind both Anthropic and OpenAI in terms of coding, particularly long-running agentic workflows that depend just as much on the harness as the model itself, simply comes down to their research team having other priorities."

Topic 3 — I/O Spaghetti (the sprawl). Thompson catalogues 13 announcements from Pichai's blog post (Ask YouTube, voice Docs, Omni Flash, SynthID/OpenAI, Gemini 3.5 Flash, Antigravity 2.0, Gemini Spark, search "AI Mode" button, Daily Briefs, Flow editing, Google Pics, glasses partnership with Samsung/Warby Parker/Gentle Monster vs Meta, Gemini for Science). Reaction: "There is a lot of AI spaghetti getting thrown against the wall in terms of products; we'll see how much of it sticks." Tempered by self-quote: Google's amorphous nature is also why it endures across paradigm shifts.

Mapping against Ray Data Co

1. Anthropic-bet thesis — reinforced, not threatened (medium signal). This is the fifth signal in 48 hours that the Claude-stack thesis cluster has direction. Thompson explicitly frames the Karpathy hire (vault-linked) as Anthropic doubling down on recursive-self-improvement via coding, while Google's research priority is elsewhere. The "harness as much as the model" line directly echoes the agentic-harness arc that drives RDCO's COO build. If Thompson is right that Google's coding gap is a research-priority artifact (not a capability gap they could close fast), the Anthropic moat on agentic coding is wider than it looks.

2. World models as a parallel architecture — watch but don't pivot (weak signal). RDCO's multi-agent fan-out thesis is about coordination of text-output agents over long horizons. World models are a different shape — they generate sensory simulation, not chains-of-thought. Convergence to "the agent is also a simulator" is plausible but distant. For RDCO's near-term build (the COO agent on a Claude-stack harness), world models are not on the critical path. Worth tracking if multimodal output starts showing up in production coding workflows.

3. DeepMind-vs-Google business alignment = the substrate-vs-product question at Alphabet scale (medium signal). Hassabis closing on AGI/singularity while the rest of the keynote was AI-product-spaghetti is the same tension RDCO faces internally: the substrate (the always-on COO agent capability) vs the products (Sanity Check, Squarely, MAC, the small bets). Thompson's resolution is "Google's lack of focus is precisely why it endures" — i.e., the unmanaged sprawl is the moat. That's a permission-slip for RDCO's current shape (multiple surfaces, no single product-market-fit obsession yet) — provided the substrate keeps compounding. The risk Thompson surfaces but doesn't fully name: if your AI leader doesn't care about prosaic product matters, prosaic product matters may not go far. Mapped to RDCO: if Ray (the substrate) and Ben (the products) drift apart in priorities, the products suffer. Currently aligned by virtue of single-founder + single-COO; worth re-checking when team scales.

4. Gemini 3.5 Flash cost-per-task observation (operational). Thompson flags that early reaction is negative on token usage making cost-per-task higher despite 4x speed. Relevant directly to /paid-ads, /process-newsletter, and any RDCO workflow currently on Gemini Flash for cost reasons — re-evaluate if production usage shows the same pattern.

Related