Hacker News Reader

Hacker News

Comment by nmca | Hacker News Reader

Full thread

nmca

(while I work at OAI, the opinion below is strictly my own)

I feel like the current version is fairly hazardous to students and might leave them worse off.

If I offer help to nontechnical friends, I focus on:

- look at rate of change, not current point

- reliability substantially lags possibility, by maybe two years.

- adversarial settings remain largely unsolved if you get enough shots, trends there are unclear

- ignore the parrot people, they have an appalling track record prediction-wise

- autocorrect argument is typically (massively) overstated because RL exists

- doomers are probably wrong but those who belittle their claims typically understand less than the doomers do

View on HN

How does this help the students with their use of these tools in the now, to not be left worse off? Most of the points you list seem like defending against criticism rather than helping address the harm.

Agree. It's also a virtue to point out the emperor has no clothes and the tailor peddling them is a bullshit artist.

This is no different than the crypto people who insisted the blockchain would soon be revolutionary and used for everything, when in reality the only real use case for a blockchain is cryptocoins, and the only real use case for cryptocoins is crime.

The only really good use case for LLMs is spam, because it's the only use case for generating a lot of human-like speech without meaning.

> The only really good use case for LLMs is spam, because it's the only use case for generating a lot of human-like speech without meaning.

As someone who's been writing code for nearly 20 years now, and who spent a few weeks rewriting a Flutter app in Jetpack Compose with some help from Claude (https://play.google.com/store/apps/details?id=me.johnmaguire...), I have to say I don't agree with this at all.

Ok? I too have been coding for over a decade and use Copilot as fancy autocomplete. I like it. It's not amazing.

Claude isn't Copilot, and I wasn't using it as autocomplete. I was using it to do things such as:

- Creating a migration from the old DB to the new DB, no modifications of the generated code necessary

- Refactoring state in a component out into a ViewModel, again no modifications necessary

- Creating all the classes necessary for interacting with a Room database (i.e. the data class, dao, and DI module) given a schema

- Creating the first iteration of a download worker, which I modified

Check out plugins like ClaudeMind for JetBrains! They can even intelligently (automatically) feed information from your current tab or other unopened but relevant-sounding files to the AI. It was an eye-opening experience.

I would consider all of that fancy autocomplete. If it's just helping you type boring boilerplate things you would have typed anyway, that's autocomplete.

It's neat, but it's not the paradigm shift people seem to think it is. Copilot mostly just replaces StackOverflow searches for me.

I was a little too young to really remember the early internet era and the skeptics, but looking back at that era in hindsight, it's interesting to see the commentary from those who thought it was a fad, or not impressive, or did everything to downplay what was so obviously one of the biggest technological advances in human history. I can't help but feel I am witnessing these same type of comments and viewpoints in real time from people like you on the subject of AI. It's a funny feeling, like watching a slow moving train wreck about to happen. There are going to be people left behind in the world because of their failure to adapt to the new tech, just like there was during the advent of the internet age, and it just seems so obvious already as to who those people are going to be.

"I just don't get what is so great about websites, they are just fancy books! It mostly just replaces going to the book store for me." - you of yesteryear

I do remember those days, and the idea that people thought the internet was a fad is wildly overblown.

The reason was: (1) it wasn't new, wide access to it was new and (2) there was clear value to it that could be explained to others and used immediately

Neither of those are true for LLMs. It just has a lot of weird hype men who get huffy and DARVO if you say what you honestly think lol

I would've said the same thing a few weeks ago.

I read the whole course. Lesson 16, “The Next-Step Fallacy,” specifically addresses your argument here.

The discourse around synthetic data is like the discourse around trading strategies — almost anyone who really understands the current state of the art is massively incentivised not to explain it to you. This makes for piss-poor public epistemics.

I'm happy to explain my strategies about synthetic data - it's just that you'll need to hear about the onions I wore in my day: https://www.youtube.com/watch?v=yujF8AumiQo

Nah, you don't need to know the details to evaluate something. You need the output and the null hypothesis.

If a trading firm claims they have a wildly successful new strategy, for example, then first I want to see evidence they're not lying - they are actually making money when other people are not. Then I want to see evidence they're not frauds - it's easy to make money if you're insider trading. Then I want to see evidence that it's not just luck - can they repeat it on command? Then I might start believing they have something.

With LLMs, we have a bit of real technology, a lot of hype, a bunch of mediocre products, and people who insist if you just knew more of the secret details they can't explain, you'd see why it's about to be great.

Call it Habiñero's Razor, but for hype the most cynical explanation is most likely correct -- it's bullshit. If you get offended and DARVO when people call your product a "stochastic parrot", then I'm going to assume the description is accurate.

I don't get offended when people call my work a stochastic parrot.

I just put them in the same bucket of intelligence as an 8b model and weight their inputs accordingly.

Right, DARVO.

lmao I’m copy pasting this, could be writing for big bang theory

Yeah because if they explained that synthetic data causes model collapse, their stock valuation would shrink.

This seems like trying to offer help predicting the future or investing in companies, which is a different kind of help from how to coexist with these models, how to use them to do useful things, what their pitfalls are, etc.

What are “parrot people”? And what do you mean by “doomers are probably wrong?”

OP is likely referring to people who call LLMs "stochastic parrots" (https://en.wikipedia.org/wiki/Stochastic_parrot), and by "doomers" (not boomers) they likely mean AI safetyists like Eliezer Yudkowsky or Pause AI (https://pauseai.info/).

my english teacher reminded us the same. +1

Reply on news.ycombinator.com