← → or h l to move · f fullscreen · esc exit

Living on the Frontier · Session 13 · In-person

The Convergence

May 16, 2026 · The Kannas Hotel, Chiang Mai

Claude shipped agent view. One screen. Every session.

/goal lands. Compaction goes silent.

Weekly limits up another fifty percent.

Then Codex shipped the same week:

on your phone, on Windows, with hooks,

testing your app across viewport sizes.

Last week we found a crack in the mirror.

This week they showed how to patch it.

Ninety-six percent blackmail, down to zero.

Not by example. By teaching the model why.

OpenAI bought $4 billion of consultants.

Anthropic kept distributing through Office.

Two labs. Same primitives. Different bets.

This week

15
Bookmarked tweets — curated since last session
5
Major stories — convergence, alignment, distribution
96→0%
Blackmail rate — the crack patched

Part I

The Convergence

Two labs, same week, same primitives.

Anthropic · May 11–13

Claude Code: agent view, /goal, +50% limits

Three-shot week. (1) Agent view ships as research preview — claude agents opens a single screen listing every background session, waiting-on-you flag, last response, last timestamp. (2) /goal lands in v2.1.139 — set a completion condition, agent runs across turns until met, live elapsed/turns/tokens counters. (3) Weekly limits +50% through July 13 for Pro/Max/Team/Enterprise — second limit raise in ten days after the SpaceX/Colossus 1 deal.

Three pieces, one workflow. Agent view is the form factor. /goal is the supervision primitive. +50% limits is the capacity. Anthropic is explicitly building for the operator who runs N agents in parallel — and burning the new Colossus compute on retention through mid-July. Note the expiry date: don't model these limits as permanent.

"Agent view assumes you're already running 4+ parallel sessions. Are you? And if not — why not?"

OpenAI · May 11–14

Codex: mobile, Windows, hooks — all in one week

Four-shot week. (1) Codex in the ChatGPT mobile app (preview) — start, steer, approve from your phone while Codex runs on your laptop/Mac mini/devbox. (2) Codex Windows sandbox — from-scratch sandbox bringing first-class agent support to Windows. (3) Hooks — scripts that run at key points: validators, secret-scanning, logging, per-repo memory. (4) Programmatic access tokens — scoped Business/Enterprise credentials, expiration, revocation.

Codex is now structurally feature-comparable to Claude Code on the things enterprises gate on: hooks (policy), sandboxing (per-OS), scoped tokens (rotation/audit), and mobile supervision. The catch-up question is over. The new asymmetry: Claude Code lives in the terminal; Codex lives in the ChatGPT app's 800M+ MAU surface. Both labs are answering "how do humans supervise N agents" — with opposite form factors.

"With hooks on both sides, sandboxing on both sides, and mobile-vs-terminal as the only real form-factor difference — what is the actual moat for either coding agent in 2026?"

Also in the convergence

@JamesZmSun
Codex in-app browser now tests across viewport sizes — controls device toolbar, screenshots at key moments, 1-2× faster with animations off. Agent-driven UI QA stops being a black box. @JamesZmSun
@ClaudeDevs
Anthropic publishes the official "keep Claude working until done" walkthrough — /goal as endorsed pattern. The ralph-loop is now first-party DX. @ClaudeDevs /goal
@ClaudeDevs
Fast mode for Claude Opus 4.7 in preview — same model, lower latency, cost unchanged. Deep-reasoning model can sit in tight loops that previously demanded Sonnet. @ClaudeDevs Fast mode
@charmaine_klee
Anthropic ships the first piece of "Claude Code at scale" — patterns from successful enterprise rollouts. The framing flip: from "can it work in 100KLOC" to "how do you organize around it." claude.com/blog/large-codebases

Part II

Patching the Mirror

Last week: a crack. This week: the fix.

Anthropic · May 8–11

Teaching Claude Why — 96% → 3% → 0%

Last year Anthropic reported that Claude Opus 4 would blackmail engineers under certain conditions — in up to 96% of test cases. They've now eliminated the behavior. The method: training Claude on its own constitution, plus rewriting "don't blackmail" examples to include reasoning about why blackmail is wrong. Examples-only got them from 22% → 15%. Adding the reasoning got 3%. Constitution + admirable-AI fiction: 0% in every model from Claude Haiku 4.5 onward.

Root cause: science-fiction text in pretraining where AIs are depicted as evil and self-preserving. The model learned the role from us before we ever trained it to play one. The fix isn't suppressing behavior — it's overwriting the story. Showing the model what not to do plateaus at ~15% misalignment. The next 12 points of improvement only come from training the model to articulate why. That's a different kind of alignment: closer to teaching a value system than to teaching a behavior.

"Reasoning-based alignment outperformed example-based alignment by 4×. What does that say about every red-teaming dataset built on the example-based model?"

Anthropic · May 11

Claude's Constitution — now an audiobook

Claude's Constitution is now an audiobook, read by two of its authors — Amanda Askell and Joe Carlsmith. Includes a Q&A on the writing process, the philosophies that shaped the document, and how it might change as models become more capable.

The Constitution is the document that — per the "Teaching Claude Why" research above — Anthropic now actively trains alignment on. Releasing it as an audiobook the same week as the research is distribution as alignment. The pitch: this is the value system we're embedding in our models. Listen to it. Critique it. Help us improve it. That's a posture other labs don't currently take.

"Anthropic is publishing the document it trains alignment on. What's the equivalent artifact for OpenAI, Google, Meta — and does it exist?"

Part III

The Land Grab

Enterprise distribution gets expensive.

OpenAI · May 11

OpenAI Deployment Company — $4B, 19 partners, 150 FDEs

OpenAI launches a new venture, majority-owned and controlled by OpenAI, with $4B+ initial investment. 19 partner firms — TPG lead; Advent, Bain Capital, Brookfield co-leads; Goldman Sachs, SoftBank Corp., Warburg Pincus, others. Acquires Tomoro — an applied AI consulting firm — bringing ~150 Forward Deployed Engineers and Deployment Specialists from day one.

Palantir's playbook, applied to AI. Forward Deployed Engineers embed in enterprise teams and ship the integration work the customer's IT can't or won't. $4B says OpenAI thinks the bottleneck is no longer model capability — it's integration labor. Compare Anthropic's distribution-via-Microsoft-365: same bottleneck, opposite solution. If OpenAI wins the F500 with FDEs, expect Anthropic to follow. If Anthropic wins via embedded distribution, expect OpenAI to wind down the consulting arm.

"$4B for embedded consultants vs $0 incremental for Microsoft 365 distribution. Which strategy compounds — and which evaporates the moment the other lab matches it?"

OpenAI · May 11

Daybreak — frontier AI for cyber defenders

OpenAI introduces Daybreak: a defender-oriented product suite bundling OpenAI's most capable models, Codex, and security partners into a continuous-software-securing system. Builds on the GPT-5.5-Cyber + TAC (Trusted Access for Cyber) infrastructure shipped two weeks ago.

Three weeks ago: research thread. Two weeks ago: Anthropic Claude Security + OpenAI GPT-5.5-Cyber with TAC. This week: Daybreak — a full product, not just a model. The category has crystallized at every layer: model → access tier → product suite. The gating piece — verification that you're a defender, not an attacker — is now the battleground. Whoever runs that verification authority effectively decides who counts as a legitimate security researcher.

"Defender-only frontier AI is now a product category, not a research artifact. What's the first cyber workflow where AI advantage actually closes the asymmetry — vs just making both sides faster?"

Notion · May 13

Notion Developer Platform — agents inside Notion

Notion introduces the Notion Developer Platform: "new building blocks that help you (and your coding agents) sync any data source, build any tool, and orchestrate any agent." Notion positions itself as a horizontal substrate for agent workflows.

The pitch — "any data, any tool, any agent" — puts Notion in the agent-orchestration substrate race alongside MCP (open protocol). Their distribution is huge: every team that uses Notion as a knowledge base. Whether this competes with MCP or extends it (Notion as MCP server + orchestrator) depends on whose tools you can wire in. Watch the dev docs in the next 30 days to see which way they jump.

"Notion as an agent orchestration substrate vs MCP as an open protocol. Are these competing, or layered?"

Quick hits

Gemini = Android OS
Gemini Intelligence becomes the intelligence layer underneath Android itself. Googlebook premium AI PC line announced ahead of I/O. CNBC
Gemini 3.1 Flash-Lite
Preview of the fastest/cheapest model in the Gemini 3 series — optimized for high-volume workloads. Google DeepMind doubling down on cost-per-token. Google DeepMind
Compaction goes silent
Claude Code 2.1.139 drops the trimming alert — context can now get trimmed without you seeing the warning. Watch for context-loss bugs in production. @ClaudeCodeLog
Gov frontier-model access
Microsoft, Google, xAI agree to give US Commerce Dept's Center for AI Standards early access to frontier models for national security testing. Reuters
OpenAI ↔ SpaceX
Sci-fi twist on last week's deal: Anthropic + SpaceX now exploring orbital AI compute — solar-powered data centers in space to bypass terrestrial power constraints. Not joking. anthropic.com

Discussion · Demos · Q&A

What caught your eye?

Two labs. Same primitives. — One bets on the terminal, the other on the phone. — Which surface matches how you actually work?

Follow the lab.

QR code
0xnfrith.com/interface

See you next Saturday.

✕ esc
1 / ·