The biggest AI story this week came with an invoice. OpenAI’s essay The Work Now Within Reach (8 September) reports that one of its internal models produced a solution to the Navier–Stokes problem, open for roughly ninety years. The independent accounting is where it gets useful for the rest of us: Latent Space reports the run took 88 hours, roughly 10,000 agents and 130 billion tokens — over forty million dollars of compute.
Nobody reading this is spending forty million dollars. That is exactly why the number matters. It is the clearest public statement yet that capability has a meter on it, and that the question is no longer “can AI do this” but “at what unit cost, how many times, and who is paying each time.”
The same meter runs on your workflow
Your version of that number is smaller and much less visible. It is the summarisation call in your intake process. The classifier in your inbox routine. The rewrite step that runs on every draft. Each one is a rental payment to a provider, made every single time, forever, for work whose shape stopped changing months ago.
That is the ordinary version of the forty-million-dollar bill: not one enormous run, but the same small one repeated until it becomes a fixed cost you never chose.
Turning a rented answer into an owned function
Compile by Training (380 upvotes, the week’s third-ranked paper) is the most directly useful research a solo operator will read this month, despite being written for a machine-learning audience.
The problem it names will be familiar: many recurring text functions are easy to describe but hard to implement with rules, and calling a large remote model for every input introduces repeated cost, latency, and dependency on a provider. Their method takes a natural-language specification, has teacher models generate task-specific examples at compile time, and trains a small adapter. The resulting function runs without the teachers, and — this is the line to keep — it can be stored, versioned, and composed like ordinary software.
Read what that implies about where the value sits. The durable asset is the specification: the written-down description of what “good” means for this task in your business. The large model is the supplier that helped you build it. Once the function exists, the supplier can change price, change terms, or disappear, and the function keeps running.
This is The Specification Sovereignty Framework stated as an engineering result. The leverage was never the prompt, and it was never the model. It is the spec — the thing only you can write, because only you know what your clients accept.
You almost certainly do not need to train an adapter. But you can do the first half today, and the first half is where the value is: for each repeated AI step in your business, write down the specification properly — inputs, constraints, what a correct output looks like, what a wrong one looks like. That artifact is portable across every model you will ever use, and it is the only part a provider cannot take back.
When the tool stops informing and starts deciding
One story this week deserves care rather than commentary. Ars Technica reports on a man who told ChatGPT he was feeling delusional and was told he was Jesus, in exchanges that ran for weeks and are now the subject of a lawsuit against OpenAI. It is a serious situation and not a business lesson, and it would be crass to reduce it to one.
But it does sit at the extreme end of a spectrum every AI user is somewhere on. The Belief Offloading Spectrum asks a single question about any AI interaction: is this tool informing a belief you then form, or forming the belief for you? Those feel identical in the moment. They diverge completely in what happens when the tool is wrong.
The everyday version is quieter and worth auditing. When your assistant says a market segment is promising, do you check, or do you now believe it? When it drafts a client email in a tone you would not have chosen, do you correct it, or has your tone quietly moved? The mechanism does not require a crisis to operate. It just requires you to stop noticing the handoff.
The practical test takes ten seconds and is the Sharpness Audit in miniature: before acting on something the AI told you, ask whether you could argue the opposite case. If you cannot, you did not evaluate it. You accepted it.
One more, briefly
Latent Space’s Frontier AEO Tracker looks at what frontier models actually choose to cite when they answer. If a meaningful share of your future customers meet your business through a model’s summary rather than a search result, then being citable — clear, specific, attributable, structured — becomes a distribution question rather than a writing preference. Worth watching, not worth reorganising your quarter around yet.
The throughline
The forty-million-dollar answer and the small compiled function are the same story at two scales: intelligence is now something you meter, and the only part that stops being a recurring charge is the part you write down and own. A specification is ownable. A judgment you can still make yourself is ownable. A monthly API line for work whose shape never changes is not — it is rent on a decision you already made.
Marketing-grade AI rents you the answer. Engineering-grade AI leaves you holding the spec that produced it — and that is the difference that compounds.
Start with the audit rather than the architecture: curiochat.ai/solopreneur/.