AI goes where the oracle is cheap

Paul Dix, writing about the end of programming:

If you can build a verification system and give proper direction, AI can produce a highly complex, highly sophisticated piece of software and it can continue to refine it until it just works.

His main evidence is the recent rewrite of Bun from Zig to Rust. A common counter is that the agents had an oracle, which explains away the achievement because they were given an existing implementation, a compiler, and an enormous test suite. I think this gets the interesting part backwards, because AI does its best work where it can cheaply discover that it is wrong (machine learning, is that you?). The oracle is not a caveat to the result but what makes the result possible.

Reproducing known behaviour is narrower than deciding what Bun should do, but migrations, compatibility work and implementing behaviour we can already describe make up a lot of valuable programming. Narrow does not mean unimportant when the narrow category is this large.

AI will move fastest through work that is cheap to judge, not necessarily work that is easy to do. We will keep seeing agents pile up impressive results on engineering problems whenever a compiler, reference implementation, proof checker or test suite can tell them when they are wrong. The challenge is much less tidy when the result has to be judged rather than tested. How do you organise a repository, write the specifications and maintain the tens of Markdown files that give an agent enough context to produce something useful? People using coding agents on messy, real work are already finding that implementation gets cheaper while describing the system, its constraints and what good looks like becomes the work.

Formal mathematics makes the same boundary especially clear. A proof checker can reliably reject an invalid proof, giving an AI room to search through many possible solutions to a genuinely hard problem. It can tell us whether a proof follows from the formal statement, but not whether we formalised the right statement or whether the result matters.

Paul calls this the end of programming, which seems like quite a leap. What Bun shows more clearly is that implementation can become dramatically cheaper when correctness is cheap to test, but cheap code does not make judgement cheap. The work is defining what counts as correct, giving agents enough context to act usefully, and then deciding whether the thing you made is actually any good. Taste™.

The language around language models

Laude Institute, introducing Headlong: A Microharness for Persistent Agents:

Introducing Headlong, an open source agent microharness featuring persistent agency. Your agent keeps thinking between external interactions in a self-guided loop inspired by human inner monologue. Headlong is a complete agent harness with a core of less than 10K lines of Bash, available on GitHub.

Rough translation: a Bash loop keeps calling a model and feeding each result into the next call.

The copy for dev tools has really taken a turn in the current AI era. Why have we started to obfuscate and add unnecessary mystery to simple tools? The plain version is still interesting! A small Bash program keeps calling a model between messages. Sure, I’ll check it out. Words such as agency, memory and thinking describe the software as if it were a person. This is open-source software for developers, why not tell us what the program does in simple words?

Just for comparison, I pulled up Stripe's homepage:

Financial infrastructure to grow your revenue. Accept payments, offer financial services and implement custom revenue models – from your first transaction to your billionth.

The software is usually interesting enough. Just tell me what it does.

Who’s doing what with watermarking AI content?

Anthropic is bringing SynthID-style text watermarking to Claude, John Gruber has some objections, and James Padolsey has the clearest explanation I’ve seen of how it works. This whole discussion got me wondering what everyone else is doing.

Here is where the major AI companies currently stand:

Google

Google DeepMind developed SynthID, initially for generated images. In 2024 it extended SynthID to text in the Gemini app and web experience, then open-sourced SynthID Text. That is why Anthropic can adopt a version for Claude, and so can anyone else.

Google’s public Gemini and SynthID verification tools currently accept images, audio and video, but not pasted text. Google (of course) has another similarly named product called the AI Content Detection API. That one examines images from Google and third-party models using pixel-level artifacts and other statistical clues. Despite the broad name, its current documentation only lists JPEG, PNG and WebP inputs.

Anthropic

Future Claude models will use a version of Google’s SynthID Text. Existing models will move over during the coming months. C2PA metadata is also included with some supported files.

Anthropic says a text-detection API is coming soon.

OpenAI

OpenAI has not released a text watermark, although it says it intends to expand provenance support to text. Supported images use C2PA and SynthID, while supported audio uses SynthID.

OpenAI Verify currently checks supported images and audio.

Meta

I couldn’t find an announced text watermark for Meta AI or Llama. Meta’s new Content Seal marks generated images, with video support planned.

Microsoft

Microsoft has not announced a text implementation. Microsoft 365 can add watermarks and provenance metadata to generated images, video and audio, but there is no general text checker.

Mistral and Cohere

Both have committed to the EU Code, but neither has publicly described a text-watermarking implementation that I could find.

xAI

xAI has not announced a text watermark. Grok adds watermarks to generated images and video, but offers nothing public for checking text.

The sudden activity is being driven in part by the EU AI Act. Its transparency rules began applying on 2 August 2026 and require generated text, images, audio and video to be marked in a machine-readable form where technically feasible. Anthropic, Google, OpenAI, Meta, Microsoft, Mistral and Cohere have all signed the relevant section of the EU code of practice.

What about open-weight models?

Google has open-sourced SynthID Text, and it is available in Hugging Face Transformers. That means I could download an open-weight model such as Qwen3-8B, run it at home, choose my own keys and produce watermarked text. But you can just as easily run the same model without SynthID; open-weights and watermarking via this technique don't really jibe.