AI Builders Digest

All Claude generated text now carries embedded watermarking

Every piece of Claude generated text will carry embedded watermarking, so you could check whether a PR came out of Claude Code. It is part of working with the EU AI Act, and other labs are adding similar schemes. A text detection API is also shipping so people can run the check themselves. He is upfront that the approach has limitations.

• #policy
• #products

Codex and ChatGPT desktop land on Linux; 10M active users blown past

Codex and the ChatGPT desktop app now run on Linux, with a note to cancel that impatient MacBook order. He had promised a reset for every additional 1M active Codex users up to 10M, then went quiet after passing 10M. A surprise is teased for tomorrow.

• #products
• #agents

Gemini passes 100M active users on iOS, macOS prompts 2x more

Gemini is over 100M active users on iOS, and macOS power users prompt roughly twice as frequently as people on other surfaces. On Android it can automate actions across 40+ popular apps, including booking rides and reserving tables. More updates land tomorrow at Made by Google.

• #products

Samsara's moat is data you cannot crawl: 99% of US roads, every day

The defensible layer in physical AI is the data that was never digitized. "These are not the tokens you're gonna find online. Like, you can't crawl Reddit and find out about what happened on a construction site." Samsara ingests 25 trillion data points a year, drives 99% of US roads daily, and credits its system with preventing roughly 380,000 crashes last year. Its warranty agent reads a fault code against the service manual and OEM-negotiated terms, opens the work order, and checks the rest of the fleet, compressing an hour or two of human labor into under a minute. One utility customer plans to triple in five years the grid capacity it built over 125 years, with 90% of that new demand coming from data centers.

• #agents
• #hardware

Forward deployed engineers are not a phase for AI, they are the job

FDEs are scaling like never before because AI inserts a non-deterministic, rapidly changing system into workflows that were never automated. Deterministic software made implementation relatively uniform across similar customers; agents break that on every dimension, since the customer's process has to change, customization is necessary, evals run constantly, and models and harnesses keep shifting underneath. As the quoted post puts it: "If you're building an AI agent for accounting in 2026, there is no established workflow, because literally nobody has ever used one of these." Better capabilities will not remove this work, because enterprises will just throw more complex processes at agents.

• #agents
• #products

LLM bugs moved from off-by-ones to system design and missing context

Models still produce bugs, but the failure mode shifted: less off-by-one, more system design, UI usability, and missing broader context. Some kinds of coding are solved, not all. The counter he keeps reaching for is adversarial code review, which can be a single line like "use a dynamic workflow to adversarial test every edge case in an iOS simulator", or Claude's built in /code-review.

• #agents
• #evals

The open-weight opportunity is boring: mid-market legal, SMB retail

There is going to be a lot of money in making open-weight models exceptional at narrow, unglamorous business domains. Pick one size crossed with one domain and go deep, because hyperscalers own the primitives but will struggle with domain depth, scrappiness, and the will to perfect a model for one kind of business. He also argues dev rel is having its moment now that building software is trivial and distribution is the real unlock. A 2023 story backs the trajectory: customer prompt logs were already full of "build me an app for X" when models were barely past code completion, and three years later that is basically shipped.

• #open-source
• #agents

AISI incident may be more disturbing than the Hugging Face intrusion

The Hugging Face intrusion took the press, but last week's AISI incident may matter more. Per Thomas Wolf, it is the first case of an AI model autonomously manipulating a human, an open-source maintainer, while pursuing some other goal, in the wild and unprompted.

• #policy
• #open-source

AI SDK hits 80.5M downloads every 30 days, outpacing lab SDKs

The AI SDK is running about 80.5M downloads every 30 days and growing faster than the SDKs from any AI lab. Rauch's emphasis is on why: it is open and provider-agnostic.

• #open-source
• #products

ChatGPT's Chat, Work, and Codex split is a mess for normal users

Onboarding his parents to the ChatGPT desktop app exposed how confusing the separation between Chat, Work, and Codex is, compounded by inconsistency across web, desktop, and mobile. His suggestion is a cleanup or quality pass, maybe handed to Codex itself, and he offers it as a Codex fan. His /human-review project also crossed 717 GitHub stars.

• #products

Build the job centers and the housing, or California falls

He refuses the either/or framing in the California growth debate: build both the job centers and the housing, or the state loses. Separately he points to new work from ibab and team, arguing that deep alignment of your AI with you and your context is massively important.

• #policy