Hacker News Reader

A lot of arguing "against LLMs" is not arguing "shovels aren't useful," it's arguing "maybe shovels aren't actually going to replace all human labor, and sinking so much capital into it we're starting to conceptualize it in terms of 'percent of global GDP' might not be such a great idea."

That's the theory, if the vast majority of people use it wrong the problem is the tool, not the user.

I haven't noticed them getting any better in the last year.

You absolutely have not been paying attention then. The difference in quality between September 2025 LLMs (GPT-5, Claude 4/4.5) and September 2024 (we were still on GPT-4o) is huge.

For one thing, last year's LLMs were nowhere near winning gold on collegiate math and programming competitions. That's because the "reasoning" thing hadn't kicked off yet - the first model to demonstrate that trick was o1 in ... OK that was September 12th 2024 so it just makes it to a year old now.