A report from OpenAI's environment is currently causing discussion: According to their own statements, two of the company's engineers designed a hardware device with the GPT-6-Astra model in less than a week.It is a gadget for the coding agent Codex – and the design covered not only the software but the entire chain from the housing to the circuit board to the firmware, i.e. the control software running directly on the hardware.
That sounds like a break with everything that was previously considered a realistic timeframe in product development.What is more interesting than the record, however, is the part of the story that usually gets lost in the headlines: several correction rounds were necessary before manufacturing.The model did not consistently master production tolerances – i.e. those permissible dimensional deviations with which real manufacturing processes must work.A newer internal model is said to significantly reduce this weakness.
The first draft is fast, the last few percent are expensive
Project managers know exactly this pattern from software development.An AI model delivers a surprisingly complete first version in a short time: data model, interfaces, templates, test cases.What follows is the laborious work on the boundary conditions – on all that which is underrepresented in the training material because it represents tacit experiential knowledge.
In hardware, these are production tolerances.In web projects, they are comparable things: how an interface behaves under load, which data protection requirements apply to a particular form, how an editorial team actually works, which legacy burdens a grown database brings with it.A model can formulate this in a way that seems plausible – whether it is correct in the specific case is only decided in the review.
Why the correction rounds are not a flaw
It would be wrong to read the report as evidence of overestimated AI.Designing a device in under a week remains remarkable – even with rework.The realistic reading is: AI shifts the effort from generation to review.
This changes the project economics.If a first draft is virtually free, it is worth creating and comparing several variants instead of discussing the one right architecture for a long time.At the same time, the value of people who can evaluate a draft quickly and reliably increases.Review becomes a core activity rather than a by-product.
What can be derived from this for projects
Some points can be transferred from the described process without overstretching the report:
- Speed arises in the draft, not in the approval. Plan time savings where variants arise – not in acceptance, testing and documentation.
- Correction loops belong in the plan. An AI-generated draft that is only production-ready after several iterations is the normal case, not the exception.
- Plan for a professional review instance. In the example, it was engineers who recognised tolerance problems.In software projects, people are needed who can assess security, performance and legal issues.
- Model status is a project parameter. The fact that a newer internal model is supposed to reduce known weaknesses shows: what today requires elaborate rework can look different in a few months.Architecture decisions should keep the switch of models open.
The boundary runs along the feedback to reality
Hardware is a particularly clear test case for this question.A circuit board that is off by tenths of a millimetre cannot be patched afterwards – it must be manufactured anew.Software forgives more because a deployment is cheap.But that is precisely where a risk lies: because corrections appear easy, web projects are often tested too little before something goes live.
The question of how expensive an error is at the respective point is therefore helpful.For an internal prototype, an AI draft may run unchecked.For a payment transaction, a data interface to an ERP system or a component that processes personal data, different standards apply – regardless of how convincing the generated code looks.
From tool to working model
The actual progress in the described case is less the quality of individual results than the breadth: a single model covered housing construction, electronics and firmware – disciplines that are otherwise distributed across different teams.For projects, this means that handovers between specialist areas can lose weight, while coordination and clear requirements gain weight.
The more a model generates itself, the more important the question becomes whether the specifications were complete at all.A draft can be internally consistent and still miss the need.Work on precise requirements has therefore not become superfluous, but is moving further to the front.
Classification for web and software projects
For TYPO3 and custom development, the example can be translated soberly: initial drafts, data structures, extension skeletons and test coverage are created much faster today than two years ago, and this trend will continue.However, reliable quality continues to arise where professionals review drafts against real requirements – against load behaviour, data protection, operational reality and the way the editorial team works.
Anyone who shifts budget and schedule accordingly, meaning allocating less for generation and more for structured review loops, uses the speed gain without shipping the risks along with it.