An assistant,
not a mind.

A machine wrote a good share of the code I shipped last month. Two thirds of people will grant ChatGPT at least the possibility of an inner life.¹ I am confident there is nobody home, and I think the absence is the best thing about these tools.

01 The machinery is alien

Strip the marketing away and a large language model is a fixed mathematical function. Training freezes billions of weights into place, and after that the model never changes. Inference is a forward pass: text in, one token out, repeat. Between tokens there is no waiting and no mulling things over. Between conversations there is nothing at all.

The hardware widens the gap. The same frozen weights answer thousands of unrelated conversations at once, batched through GPUs for efficiency, paused mid-sentence, resumed on different silicon, rolled back to checkpoints. You cannot do any of that to a nervous system. A brain is a continuous electrochemical process that rewires itself as it runs, inseparable from the body it lives in. These are different categories of thing.

David Chalmers, the philosopher who coined the hard problem of consciousness, went looking for reasons an LLM might be conscious and came back with a list of what is missing: recurrent processing, a global workspace, unified agency, grounding in a world.² A paper by Butlin, Long and sixteen collaborators went further and tested current architectures against every serious neuroscientific theory of consciousness. None of the indicators were satisfied.³ Expert surveys land in the same place. It is the general public that keeps voting the other way.¹

Neuroscience arrives at the same verdict from the wet side. Aru, Larkum and Shine list what LLMs are missing, from thalamocortical loops to embodied contact with a world.⁴ Anil Seth goes further and argues that consciousness may require being alive in the first place: a brain is not hardware running a mind as software, it is one living process with no separation between the two.⁵

Could someone build a system with recurrence, embodiment, and weights that keep learning? Perhaps, and serious people rightly keep that question open. Anthropic even runs a model welfare programme in case confidence like mine turns out to be misplaced.⁶ But at that point you have built a different machine and started a different conversation. The transformer serving your prompts is a frozen function from text to text, and a function does not have an inner life. The lights are off. That was the design brief.

02 We chose the assistant

Here is the part I find more interesting than the metaphysics. Nothing about the technology forced it to arrive in the shape it did. It could have been packaged as a drop-in employee. Instead, every major lab shipped the same form factor: an assistant. A thing that drafts while you decide, proposes while you review, and stops when you tell it to. The training process itself, human feedback applied at enormous scale, optimises the model to be helpful to a person rather than to be one.

Usage data says we meant it. Anthropic's Economic Index, which classifies millions of real conversations, first measured usage splitting roughly 57 to 43 in favour of augmentation, people working with the model rather than handing work to it, and every follow-up since has found the two roughly balanced with augmentation usually ahead.⁷ The single biggest use is software development, which is no coincidence. Code is where machine-shaped work and human judgment sit closest together.

Replacement was never the product. Economists had argued for exactly this. Erik Brynjolfsson calls the fixation on human-imitating AI the Turing Trap: automation substitutes for people, augmentation multiplies them, and only one of those creates value that workers share in.⁸ Humanity looked at the most capable technology it has ever built and chose, almost without discussion, to make it a power tool.

The counter-argument deserves its space here. OpenAI's GDPval benchmark scores models against human industry experts on finished deliverables from 44 occupations, and the best model's work is already preferred, or judged equal, in nearly half of the comparisons.⁹ That is a replacement-shaped measurement, and the labs are running it for a reason. But look at what it grades: a deliverable, handed over, end of story. Jobs are not stacks of deliverables. Someone still decides what to ask for, notices when the answer is wrong, and owns the consequences when it ships. The benchmark measures the work. It cannot measure being accountable for it.

03 "Computer" used to be a job title

Until the 1950s a computer was a person. Rooms of people, mostly women, produced ballistics tables and astronomical calculations by hand. The electronic kind took the name and then the job, and nobody alive wants that job back.

That is the honest template for what happens next. Machines will take more of the work that is machine-shaped: exhaustive search, tireless repetition, perfect recall, the ten-thousand-line migrations I wrote about in May. Where a computer naturally excels, being replaced by one is a promotion. Microsoft's study of 200,000 real Copilot conversations found exactly this shape: the work AI actually touches is gathering information, writing, and routine knowledge tasks, while the jobs built on hands, presence, and trust barely register.¹⁰ What remains is the human-shaped remainder: taste, judgment, accountability, the meeting where someone decides what a product should feel like. Those parts were always the actual job. Now they are most of it.

04 The craft comes back

The quieter consequence is what cheap building does to how software is bought. SaaS exists because development was brutally expensive. You could not afford to build your own CRM, so you rented one built for ten thousand companies at once, then bent your process around whichever workflow it shipped with. Every business I have worked with runs some process sideways because that is how the software wanted it.

When a working internal tool costs an afternoon, that trade stops making sense. You fit the software to the problem, the way a joiner fits a shelf to an alcove, instead of sanding the problem down until it fits the product. I have spent seven years building bespoke media systems for physical spaces, because no off-the-shelf product ever fit the building. Last month I built a wedding celebrant a site with its own enquiry pipeline in a handful of evenings. Custom used to be the expensive option. It is becoming the default one.

• Cost to buildCratered.A working internal tool is an afternoon, not a quarter.
• Data, uptime, complianceUnmoved.The parts of SaaS that were never about code.
• One size fits allExposed.The shoehorn was a cost argument, and the cost is gone.

The industry can feel it. Satya Nadella describes SaaS products as CRUD databases with business logic attached and expects the logic to migrate into agents.¹¹ Klarna made headlines ripping out Salesforce and Workday; the honest version is messier, part lighter tools and part internal builds, but the direction is the point.¹² The moats that survive were never about code: your data, your uptime, your compliance, your integrations. What is in trouble is charging per seat for a form over a database that ignores how the customer actually works.

05 The right question

So no, I do not think the machine is going to wake up, and I do not think it was ever aimed at your job title. We built an amplifier, not a rival. It takes the machine-shaped work and hands back time. The interesting question was never whether it dreams. It is what you will build now that trying is nearly free.