The Collection

Josh Rowe’s 3–4 September letter opens on Commonwealth Bank: about $400 million in gross productivity benefits from AI expected over the coming year, as reported in The Australian. Beside that number: a Chief AI Officer (Ranil Boteju, from November 2025) and a first Chief AI Scientist (Professor Mary-Anne Williams, May 2026), remit across machine learning, responsible AI, AI security, generative AI. Westpac put data, digital and AI under a Group Executive reporting to the CEO, with a Chief AI Officer. NAB has a Group Executive for Digital, Data and Artificial Intelligence (Pete Steel, November 2025). Qantas put AI into the remit of its Group Chief Technology, AI and Transformation Officer (Rachel Yangoyan, December 2025).

Telstra is the clearest control plane in the letter: named owners of AI systems, extra review for higher-impact uses, oversight that gathers technology, legal, privacy, cybersecurity, people and customer functions, plus board oversight. The practical reason named: once AI does real work, “the AI did it” is not an acceptable answer. Someone owns the outcome the way they would if a person or a team had done the work.

You can remove work without removing responsibility.

The test Rowe prints: pick an AI system doing something important and ask who is responsible for its output. Not who bought the software. Not who built the model. Who owns what happens when it is used. If three people point at each other, or the answer is “technology,” you have learned how ready the organisation is to scale. The $400 million number says the machine is economically meaningful. The hires say someone still has to sit around it.

Latent Space on GPT-6 Astra’s launch day. First Stargate, lightly looped supermodel from OpenAI. Reported: cleanly beating Fable 5.1 on many metrics, FrontierMath saturated at 97.6 percent, ARC-AGI-3 at 99.9 percent. Greg says AGI is here. Jakub says it is finally the Automated AI Research Intern. The letter declines those talk tracks and reports what early access showed after burning over 20 billion tokens: a new class of model that can choose and train models, label data and drive active learning, keep pipelines saturated, instrument and read logs, deploy and debug systems in one shot, fan out and command and eval subagents, keep coherence over billions of tokens in a single agent thread.

The rate card as tested: about 33 tokens per second at a max $50 per million token rate, which lands under $6 an hour in their arithmetic. Astra more token-efficient than Sol and Fable, independently confirmed by Artificial Analysis as the letter states. Throw Ultra at it and the parallelising burns more than $6 an hour because it is so good at parallelising. About $100 over two days for work the letter would pay a junior AI engineer to do: managing fleets of subagents, monitoring runs, starting and stopping waves, making benchmarks, handling money, scaling runs, getting human ratings.

Under six dollars an hour. Until it parallelises.

OpenAI, in the letter’s frame, trained a model that can automate much of its own AI engineering. The Owner’s problem is not the rate. It is that a desk that cheap still leaves a human answerable for what the desk ships. Raise your ambitions, the letter says. Raise the name on the door with them.

Every’s free preview of the GPT-6 Astra vibe check, by Katie Parrott with GPT. Astra is worth the fuss. They are still keeping Fable. Writing, consulting work, and designs they wanted to keep. Dan Shipper enthusiastic about computer use. On the complicated apps they tested, Fable 5.1 still did better. Astra had a habit of adding copy and interface elements the task did not need: a cozy island that arrived with a guided-breathing exercise nobody asked for.

Named in the free grafs: Astra wrote a first draft of the review in one shot that Dan thought Katie had written; Mike Taylor’s AI training curriculum built from employee interviews, close to the course he wanted to teach. The detailed comparisons sit behind the subscribers-only wall. We label them unread. We do not invent the unlocked scorecard.

Astra impressed. Fable finished the product.

Ownership is also taste. A machine that pads the brief is not owned until someone deletes the breathing exercise. For writing, consulting, visual prototypes: try Astra. For a complicated product: keep reaching for Fable 5.1. That is not a benchmark. It is a named preference with a human still holding the scissors.

Gary Marcus on the Sanders-Casar Ban Artificial Superintelligence Act. He is fine with Sanders’s own tweet focusing on a defensible subset, and can see arguments for a pause. The proposal itself goes too far: a permanent, unilateral ban on all research into superhuman AI is too broad, a guarantee of leaving the US behind, and not the right approach. Naive about benchmarking complexities. Leaves room for authoritarians to play with the rules. Focuses too much on hypothetical future risks (AI that plans to disempower humanity) at the expense of current risks, citing Dr. Angela Rasmussen.

Marcus wants a temporary pause until better regulation and alignment ideas exist, and an AI agency, as he told the U.S. Senate in 2023. The better language he prints is Anthony Aguirre’s: narrowly-scoped tools and systems reasonably expected to sit significantly below the current frontier may continue; systems of greater combined intelligence, generality, and autonomy may be developed only if and when their developers can demonstrate with high assurance to a capable independent authority that sufficient alignment, control, and oversight exist. A high bar and an incentive, not a ban that only Congress can undo.

A high bar is not a permanent ban.

A separate Marcus hot take the same day on Astra: symbolic world models are vindication; robustness is the key unknown; ARC-AGI success is not AGI; less monitorable than prior systems is not great beside more capability; the ten Brundage end-2024 tasks still unsolved as far as he knows. Ownership of the frontier is not a tweet. It is a bar an independent authority can fail you on.