Hacker News Reader

Comment by thw09j9m | Hacker News Reader

Tech market is rough. I think BigTech currently thinks that there's a better ROI spending $300k on compute (training an LLM) than $300k on an employee. All the startups I've interviewed with are HEAVILY leveraging LLMs (via Cursor) to keep engineer count and burn rate low.

View on HN

> HEAVILY leveraging LLMs (via Cursor) to keep engineer count and burn rate low

Does this strategy work?

I think it's going to blow up given the number of broken libraries and products out there. Management does not care until things really break. In 3-4 years, there will be a serious shortage of software developers and maintenance of the stack will be as expensive as ever.

This, plus in several years we will observe many MVPs generated (partially) by LLMs that will require fixing and further development. LLMs barely produce 1k lines of coherent code, riddled with bugs. Benchmarks are improving, but there is no breakthrough in handling real, large codebases. I have no idea about the future job market, but the quality of the "material" we work with will be quite poor.

I don't think these companies are generating MVPs using LLMs. They're just using LLMs to assist in coding. A human reviewers is there. And in a few years, LLMs might be so good that they can fix mistakes by 2024 LLMs.

To clarify, LLMs will lower the entry barrier for non-programmers to test new ideas (which is good). Later, those successful projects will require development by professional developers. I suspect that many of these projects will have lower quality than regular proof-of-concept projects today.

>I suspect that many of these projects will have lower quality than regular proof-of-concept projects today.

Exactly. It will be way worse than a hacky PoC because you might able to reason about the structure or logic based on the human author and their intent which is impossible in the case of an LLM author. Further, assuming the human author isn't hit by a bus or whatever, you can query them directly if anything is unclear. Good fuckin luck asking an LLM for a reason why it did something in the PoC.

Not even close

> HEAVILY leveraging LLMs (via Cursor) to keep engineer count and burn rate low.

Absolutely make believe FUD

Reply on news.ycombinator.com