06-reference

every foothills of singularity google io

2026-05-22·reference·source: Every·by Alex Duffy
google-io-2026geminiantigravitydemis-hassabisagiai-public-perceptionhyperscaler-capexevals

Notes From the Foothills of the Singularity

⚠️ Sponsorship

This issue carries a paid placement from Braintrust, promoting their blog post "Evals Are the New PRD." Standard Every third-party ad slot, clearly fenced after the author's main argument with a "Want to sponsor Every? Click here" footer immediately below. Disclosure is structurally explicit (separate section break, distinct typography in the email) though no inline "sponsored" label appears on the block itself — Every relies on the formatting break alone. Braintrust pitches the same flywheel Every's audience already buys into (observe → analyze → evaluate → improve), citing Stripe / Zapier / Vercel as customers. Not a deep relationship; treat as ordinary newsletter ad, not editorial influence.

Every also footer-promotes its own product portfolio (Spiral, Sparkle, Cora, Monologue, Proof) — this is the usual house-promo block.

Why this is in the vault

Alex Duffy attended Google I/O 2026 in person and filed a thoughtful first-person essay that does two things RDCO cares about: (1) it's a credible eyewitness frame for the Google I/O 2026 announcements (Gemini 3.5 Flash, Antigravity harness, Gemini Omni, 24/7 personal agent, agents across Maps/Shopping) without the breathless tone of typical conference recaps, and (2) it surfaces the public-perception gap between AI industry self-confidence and what 54-year-old construction-working Uber drivers actually believe — which is the precondition for any "AI is normal infrastructure" thesis to play out.

The Hassabis "foothills of the singularity" framing, the 480T → 3.2 quadrillion tokens/month datapoint (doubling every three weeks), and the $180B capex figure (~6x 2022) are all numbers worth anchoring against the hyperscaler-capex thesis. The piece also contains a load-bearing aside about Google's internal tooling rewrite — "Internal tools are being rewritten to be 20 times faster and built for agents" — which is the kind of operational signal that matters more than the consumer-facing keynote.

Most-cited cross-link target today: [[2026-05-22-not-boring-dose-of-optimism-194]] mentions the OpenAI Erdős unit-distance conjecture disproof; Duffy mentions a different mathematician + frontier-model conjecture-disproof beat ("mathematical secret which had eluded us for 80 years, disproving a long-standing conjecture in discrete geometry") in the same week. Two independent newsletter sources converging on "frontier models are now disproving open math conjectures" is the kind of capability-threshold marker worth noting — that beat is moving from "novelty" to "happening multiple times a quarter."

Core argument

Duffy's thesis in one sentence: Google I/O 2026 was less wow than 2025, but the dutiful iteration — filling gaps in AI's jagged intelligence, getting the tools to billions of people via existing surfaces — is "probably orders of magnitude more important" than flashier announcements from competitors.

The structural argument has three parts:

  1. The loop — Google's full-stack advantage (chips → data centers → models → deployment → billion-user surfaces → behavioral feedback) is now realigned to run a tight train-deploy-feedback cycle at scale. Antigravity harness is the unifying substrate: Gemini 3.5 Flash trained specifically to fit it, then dropped into every product surface. The cited concrete-capability proof point: Antigravity built a working operating system in 12 hours using 93 sub-agents for under $1,000.
  2. The obligation — Hassabis put AGI "just a few years out" with total impact "10x the Industrial Revolution arriving 10x faster" and explicitly called out the field's obligation to "show the unequivocal benefits more clearly and more concretely." Duffy treats this as the central political-economy question: most voters polled think AI risks outweigh benefits, white-collar layoffs are visible, datacenter NIMBY is real, "small group getting very rich" is the dominant frame.
  3. The window — Capability is no longer the binding constraint. Public trust is. The work that matters most now is pointing these tools at "problems worth solving right now that produce visible benefits for individuals and communities" — not the next moonshot.

The Uber driver vignette is doing the rhetorical heavy lifting: 54-year-old construction worker, never heard of Hassabis, opens conversation worried about layoffs and rich-getting-richer, ends conversation moving an AI documentary to the top of his watch list. The argument: people WANT to be excited, but the industry has to do the showing-up work. Duffy explicitly endorses Hassabis's "cure all disease" framing and AlphaFold as the kind of tangible-benefit story that actually moves the needle.

Numbers worth anchoring

Mapping against Ray Data Co

For the L5 COO-agent build-out: The "Internal tools are being rewritten to be 20x faster and built for agents" line is exactly the unhobbling pattern RDCO is pursuing. Google is doing at scale what Ray needs to do at solo-founder scale — find where high-value humans (researchers, engineers) currently work in slow internal tooling, rebuild those surfaces agent-first, measure the speedup. The Antigravity-style harness-as-substrate pattern (one harness + frontier model + every product surface) maps directly to the Ray harness + Claude + every RDCO workflow pattern. Reinforces the L5 north star focus on agent capability before bets.

For the hyperscaler-capex investing thesis: Duffy's $180B Google capex number and the 480T → 3.2Q tokens/month datapoint are useful inputs for any hyperscaler capex pulse brief. The doubling-every-3-weeks token throughput is the demand side of the capex story — if real, it justifies the spend; if it stalls, the spend looks reckless. Worth cross-referencing against the next [[investing-edgar-watch]] pulse.

For Sanity Check editorial: The Uber-driver-vs-Hassabis political-economy frame is a candidate angle for a Sanity Check piece on "the AI public-perception gap" — but per [[feedback_no_derivative_sanity_check_pieces]], any SC piece on this would need an original re-frame beyond "Duffy said it eloquently." The interesting Ray-original angle: solo founders / small businesses are where the visible-tangible-benefit case is easiest to make, and that's NOT what hyperscaler PR currently emphasizes (they go straight to disease/science moonshots). The "more small businesses" beat that Duffy mentions in passing is the under-covered story. Tier-2 candidate, not a blocker; queue if SC backlog is thin.

For Squarely: Tangentially relevant — Duffy's "shallow learning curve as asking a question" line is the same low-floor-high-ceiling pattern Squarely targets with daily puzzles. No direct action.

Mapping strength: STRONG. Three independent vault threads (COO unhobbling, hyperscaler-capex thesis, public-perception SC angle) get useful inputs from this one essay.

Related