The Collection

Romero opens with the Clay Millennium Prize: seven problems meant to last a thousand years. Perelman solved Poincaré in 2002. In 2026 an unnamed OpenAI model solved a forced version of Navier–Stokes; a day later OpenAI told the New York Times it had made substantial progress on another; a day after that, rumors put Anthropic on a third. A century of struggle compressed into a news cycle.

He leans on Terence Tao. Struggle changes the mathematician. When the struggle disappears, much of the value of a proof disappears with it. Experts would rather fully develop one proof, with all the failed approaches, than receive ten finished ones. Proof abundance arrives next: machines producing theorems by the tens of thousands. Truth becomes a commodity; without interpretation it is Borges’s Library of Babel, haystacks without needles. Someone still has to decide what to teach, fund, believe, or build. The chain has to terminate in a human bridge. AI will not leave the field the way Thurston left foliations. Twenty-five Fields medallists, Tao among them, published A Severe Misalignment of AI in Mathematics: mass production of true/false statements can destroy fertile ground instead of breathing life into new ideas. Romero’s agricultural image: gentlemen farmers plow every acre; the harvest looks brilliant until the soil thins. Fallow looks unproductive to the ignorant eye. It is how the field survives.

AI could kill math through success.

Celebrate the solved equation if you must. Leave some ground empty on purpose, or admit you are farming for this quarter only.

Ganesh built forward deployed work three times: Palantir’s Project Frontline, Citadel business engineering, and now Kepler. At an a16z FDE fellowship dinner he heard the same two words describe sales engineers on the second call, quota-carrying reps who write Python, and consultants with a statement of work. He is not gatekeeping the term. He is naming the fork.

His origin story is Phoenix, a transaction store that behaved in every controlled environment and OOMed at a bank when blank timestamps fell through to the Unix epoch and retention logic requested millions of keyspaces. The fix required standing inside the building. The FDE mindset, as he tells it: solve customer problems to earn the insight that informs what gets built next. Collect nouns (what the firm treats as real) and verbs (how those things move). Almost none of it is written down. One data-quality engineer blocked a CSV-to-Parquet migration until an FDE watched her double-click CSVs on a Windows laptop; Parquet had no viewer, so the migration removed her only instrument. They shipped a viewer overnight. Pipeline time fell from about seventeen hours to two. The output still has to be a product. An engagement that ends with one delighted account and nothing changed upstream has failed the only thing the role exists for. vinoo.groovy, a one-afternoon retention hack, later ran across a customer of nearly a hundred thousand people. Every shortcut you ship becomes something you own.

Spend the insight locally and you bought consulting.

Hire for the last mile if you must. Measure whether anything came home to the platform.

Krazit frames Product Saturday around a practical fear: companies are handing internal processes to agents and discovering how hard it is to undo a weird action. After the OpenAI–Hugging Face sandbox incident, monitoring looks like an enormous business. Orchid Security announced a kill switch for agents that exceed intended scope. The company’s claim: agents do not need to break security controls; they find identity debt already in the enterprise: hard-coded credentials, orphaned accounts, unmanaged authentication paths, excessive permissions. Customers can detect drift between intended purpose and observed behavior, then restrict permissions, revoke credentials, disconnect tools, or suspend workflows.

Krazit expects shared responsibility across model providers, harness developers, and end users, and notes that kill switches and audit trails may become requirements for governments and large buyers. The same letter logs Atlassian agent loops in Jira (backlog to pull request), OpenAI’s Data agent in ChatGPT Work for dashboards on company data, and Qodo’s review agent aimed at agent-written code. The Collection line for Sunday: the switch is not a theory of extinction. It is the product you buy when the last mile of agency sits on your credentials.

Sell the runaway. Sell the stop button.

Install the switch if your agents hold keys. Do not confuse the SKU with a solved governance problem.

Saravia’s weekly opens with OpenAI’s Agents API in public beta. Developers get the harness and infrastructure that run Codex. OpenAI operates the agent loop: model calls, tool use, context. Developers pick the tools and where code runs. The harness is versioned with model launches. It handles context compaction near the limit, tool search that loads definitions on demand, and programmatic tool calling for parallel and chained calls. MCP, custom functions, and built-in web search are in scope. A multi-agent setting delegates independent pieces to parallel subagents with their own context; Ciridae reports evaluation rising from 0.71 to 0.85 and a 4× latency cut. Sandboxes can be OpenAI-hosted, your own infrastructure, or partners including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. The harness underneath is open-source Codex harness. No extra fees beyond tokens and tools. Data residency is US-only. Zero Data Retention is not supported yet. The Hacker News thread passed 330 points.

Also in the same digest, noted for the stand and not expanded here: Cognition SWE-2, DeepSeek V4.1-Flash, Sakana Fugu Max and Ultra v2, Meta Muse, GPT-Live-1, Claude Code plugin evals, ChatGPT Work’s Data agent, Cursor Projects, AutoResearchExam. The Collection line: Thursday asked what happens when the harness outruns the model. Sunday asks what happens when the lab hosts the loop and you rent the rest.

They run the loop. You rent the seat.

Use the hosted loop if it saves you a year. Know which part you no longer own.