Comment by rank0 | Hacker News Reader

Comment by rank0 | Hacker News Reader

The majority of comments I read/hear about openAI are horrible takes. It makes me believe most of the people haven’t taken the time to evaluate the output of these LLMs.

Programmers aren’t going anywhere. ChatGPT doesn’t “understand” things. It optimizes its output for some set of scoring metrics. It sees from it’s training data that tokens score better (on average) in a particular order or that some tokens are associated with some sentiment. It doesn’t see english words.

With code generation the biggest benefit is generating code scaffolding. That’s nice and all, but certainly NOT where the heavy lifting happens during system development. The scaffolding it shits out usually needs to be heavily modified by a human. The amount of capital, compute, memory, and energy it requires is insanely inefficient for generating boilerplate code. Don’t even get me started on the privacy implications…

OpenAI will make programmers more productive by reducing time spent on trivial/mundane tasks. That’s awesome! However, it’s not going to replace the SWE industry.

If it doesn't "see English words", how it can correctly perform tasks such as this:

"Tell me about the structure of the solar system. I want you to only use English words of Germanic origin; avoid using words derived from Latin or Greek. If there's no alternative, make up a new word by calquing the original into Germanic roots."

Just to give an idea of the output GPT-4 can produce from this, here's a snippet:

"... beyond the worldrocks, rockbits, and icebits, the sunstead also has spacegrit, small bits of rock and metal that fly through space. When spacegrit enters Earth's air and burns up, it makes a lightstreak that we can see in the night sky. If a bit of spacegrit reaches Earth's ground, it is called a starstone ... "

Rockbits are asteroids (and it does consistently use this term, like all others, throughout the composition). Can you explain how this invented word, or even its individual components, would be "associated with some sentiment" in the training data that leads to this output, if the model doesn't actually understand, on some level, what it's asked to do?

> OpenAI will make programmers more productive by reducing time spent on trivial/mundane tasks.

It doesn't have to be perfect to replace developers. If every developer in a company becomes 5x more productive but you can only sell 3x harder, something might still have to give.

Even if GPT-4 doesn't quite cut it, what about GPT-5? What about the crazy, off-the-wall concept that comes out of some university and completely revolutionizes the whole thing again by 2024? Stanford is already demonstrating an inclination for disruption. There is an incredible amount of energy, attention and capital being poured into this space right now. If all acceleration forces were removed, I think we are still going to wind up somewhere crazy.

We certainly aren't ever going to go backwards in capabilities. It will only improve - even if just marginally. I've got GPT3.5 suggesting very high quality methods for my codebases. You could take 20% of the capability away right now and I'd still consider it a revolution in productivity.

> It makes me believe most of the people haven’t taken the time to evaluate the output of these LLMs.

How much experience have you had with GPT3.5/4?

Reply on news.ycombinator.com