№ 1 · Geopolitics
Meta's Chinese agent push meets a domestic platform wall.
Proof.
A model writes a theorem, a brain rents some hands, ten agents share one desk. Today's issue is about what the next layer of work is allowed to show.
Mathematics · Discovery
The model wrote a new lower bound.
An internal OpenAI model solved the eighty-year-old unit-distance problem in discrete geometry — disproving the long-held assumption that square-grid constructions were optimal. Independent mathematicians verified the proof; a Fields medalist called it a milestone for AI in mathematics. The first new tool the field has gotten in decades that doesn't need an undergrad.
- conjecture
- Optimal unit-distance configurations are achieved on integer square grids.
- result
- Refuted. The model exhibits a non-grid construction strictly exceeding the prior bound.
- success rate
- ≈ 48% across multiple runs with extended compute budgets.
- status
- Independently verified.
Infrastructure · Agents
Brain on Anthropic.
Brain on Anthropic.
Hands on Cloudflare.
Cloudflare wired Claude Managed Agents into its edge: V8 isolates that boot in milliseconds, full microVMs when a workload needs a Linux container, and customisable outbound proxies so a runaway agent can't phone the wrong host. The split is the thesis — keep the model where it scales; keep execution where you can revoke it.
The brain
Claude runs on Anthropic's platform — the reasoning loop, the tool calls, the planning.
The hands
Code, browsers, email, Workers AI — sandboxed at the edge, in a region you control.
Case study · Agents
Ten agents per engineer.
Ten agents per engineer.
Six hundred deploys a day.
Superset's IDE for AI coding agents runs ten of them in parallel per developer on Vercel infrastructure, with thirty-second average builds and roughly six hundred preview deploys daily. Weekly active developer growth has hovered between fifty-seven and sixty-four percent week-over-week. The proof that agentic dev works at scale is now a load curve.
10parallel
Coding agents per developer, isolated workspaces
600/day
Preview deployments shipped from agent runs
30sec
Average build, edge-to-edge, no warm-up
Code review · Codex
Substantive feedback in minutes, not hours.
Ramp put Codex with GPT‑5.5 on its code-review path. The choke point for human reviewers had become triage — sifting fifty-line pull requests to find the three that actually needed careful eyes. Codex now does that first pass, leaving senior time for the architecture conversations that still benefit from one.
Fromeng-velocity@ramp
Toall-eng
Recodex-on-the-review-path
before: reviewer reads 50 PRs, finds 3 worth attention
after: codex pre-reads 50 PRs, flags 3 worth attention
delta: ≈ same 3 PRs · ≈ 0 wasted reviewer-hours
model: GPT-5.5 via Codex · prod for weeks · no regressions
after: codex pre-reads 50 PRs, flags 3 worth attention
delta: ≈ same 3 PRs · ≈ 0 wasted reviewer-hours
model: GPT-5.5 via Codex · prod for weeks · no regressions
Stack · Observability
What to wire up when Cursor wrote most of it.
PostHog's guide for engineers shipping AI-assisted code argues for three layers under any vibe-coded app: product analytics for behaviour, web analytics for acquisition, and LLM observability for the parts of the codebase you didn't quite write. Without the last, you'll ship features whose failure modes you've never seen.
i.
Behaviour
Product analytics — events, funnels, retention. The classic stack, only now the funnels were written by an agent and you still own the outcome.
ii.
Acquisition
Web analytics — where do they come from, what bounces. Cheap, separable, and the one layer agents can't see for you yet.
iii.
Reasoning
LLM observability — traces, prompts, tool calls. The code you didn't read is the code you must instrument.
Weekly · The Batch
Meta hits a Chinese wall.
Meta hits a Chinese wall.
The radiologist gets a co-pilot.
This week's roundup from Andrew Ng's team: Meta's agentic ambitions in the Chinese market run aground on local platform incumbents, U.S. government auditors progress quietly on frontier-model evaluation, and an AI mammogram-reading system clears its next regulatory bar. Three stories the daily-AI feed flattened into one tidy issue.
№ 2 · Audit
U.S. evaluators expand frontier-model testing — quietly.
№ 3 · Medicine
AI mammography clears another regulatory bar.
Opinion · The Vase
A quiet week, for once.
Zvi Mowshowitz — weekly tour de force of every AI development that actually moved a needle — files a column titled "Not Leading the Future" and, for the first time in many months, finds little new to lead with. Read it as a marker: the deafening pace did, briefly, lift.
Dear reader — the news this week is mostly that the news is mostly nothing. Take an afternoon. Sit with the model you already have. Yours, Z.
Zürich
21·V·MMXXVI
Issue ·033
21·V·MMXXVI
Issue ·033
From:
The Vase
Substack, NY
The Vase
Substack, NY
Security · Open source
$ deepsec — find what the audit missed.
Vercel open-sourced deepsec, an AI security harness that runs on customers' own infrastructure, against customers' proprietary code, with customers' own API keys. The MIT licence means it sits inside the perimeter the way a static analyser does — no SaaS round-trip to scan a private repo.
$ deepsec scan ./src --model gpt-5.5 → indexing 1,284 files · 312k LoC · 18 routes → pulling cves & conventions from heuristic pack → reasoning · multi-pass · cross-file dataflow ! auth/session.ts:64 missing CSRF on /admin/wipe ! jobs/runner.ts:118 user input flows into exec() ! public/.env.local secret committed in week 12 · 14 lower-severity findings — see report.json runs on: your infra · your keys · your code licence: MIT
That's today.
Eight stories, six sources. Picked for what a senior engineer or founder could act on inside a week: a math result, an agent infrastructure primitive, two production case studies, a stack guide, a weekly roundup, an honest pause, and one open-source security tool.
sources: openai · cloudflare · vercel · posthog · deeplearning.ai · don't worry about the vase
rubric: ai tools you could adopt this week · creative software · dev tools & agents · privacy & security · research with a practical kernel · anything immediately actionable.