Live data from Hacker News

Ask HN: How do you envision the AI Software Factory?

news.ycombinator.com

1–5 of 5 posts

Re: Ask HN: How do you envision the AI Software Factory?

#3
A single engineer who scopes, designs, implements, and delivers on their own. So all internal-facing roles (product, design) collapse back into "software engineer" (where we started). All external-facing roles (gtm, marketing, sales) collapse into "sales engineer". Support is mostly automated and what little remains collapses into software engineer (already happened to a large extent).

Everything between instruction and working preview will collapse into the ADE. The engineer will shape working software by instructing an LLM, and be able to interact with the working software as it is developing. Reading code will be done rarely as part of debugging or analysis. Agile, product-focused companies are already operating like this.

However, the speed of software delivery will remain capped at the speed of intra-organization coordination. In other words, you can't deliver software faster than you can have meetings, handoff work, and make decisions. "Alignment" in corp-speak is already the bottleneck at 90% of companies. This is also why most companies will not be able to adapt. They can't ship faster as long as they keep product managers, meetings, and hierarchy between the builders and sellers. IMO this is why most alpha in the next decade will be had by remote senior engineers who only show up to meetings and alignment and otherwise automated everything else (already happening). That won't last forever as the AI-first, product-led organizations are going to eat everyone's lunch on a longer time scale (e.g. the one-person billion-dollar company people talk about).

Just my 2cents.

Re: Ask HN: How do you envision the AI Software Factory?

#4
I definitely believe it's doable if someone just asking for a new combination of pieces from here and there.

I did quite some SDD in last quite some months, and had improved my SDD kit to a level where it can stick to the architecture and scaffolding right thing if you are requesting something quite standard and finish it in a one long shot.

But when you request something quite unique, it stumbles where you will have to intervene more to guardrail it to the right destination that you want exactly.

So I guess the question is about when it's that easy to generate software, then what will be the new requirements from users? Do we still assume people will ask for what they were asking?

I doubt so.