Hacker News Reader

Comment by karim79 | Hacker News Reader

I truly hope that the common theme of the likes of "JWST Just Found Something Which Should Not Exist" etc will not be augmented by stuff like "we used AI(tm) to figure out X, Y, Z".

The last thing we need is hallucinations fucking up the more grounded astrophysics. I'm not saying that is what is happening, I just worry about stuff like this. AI causing us to bark up the wrong tree, and so forth.

View on HN

If anything, it's just going to call out a thing in the image that humans can then go and look at. Nothing in astronomy is ever "decided" by a single report. It gets looked at and scrutinized, and then committee style decisions are made about it. So if someone is using some ML to scan every image taken by JWST and calls out 1 cool thing for every other 9 things it finds that's "yeah, we know about that", then that's still quite a lot of new cool things. it'll just be able to do this faster and potentially much more in-depth than a human scanning across the images manually

Yeah but what if we start seeing only using this new awesome tool? What if that becomes the new seeing apparati? THIS is the tool that breaks that mold? The tool that (near?)every field is also going to be considering to be the tool that's off limits, or be 'constrained?'.

What if we had that view with microscopes, back when?

I see the point being made above fully. If ai takes over it's because we are every day it seems like slowly placing that faith.

It's our wow. It's the future generations taken for granted.

"Much more in-depth" ways now just "the way".

That's going to be a sign of the times if that happens. There are way too many people that enjoy the search doing it by hand. Yes, they are all of a certain age. Those of a certain younger age that only knows digital tools and not the ways of using their own eyes might eventually happen, but thankfully I won't be around for that to happen. (I'm one that uses my own eyes).

> Yeah but what if we start seeing only using this new awesome tool?

Like telescopes?

Real astronomers just squint real hard at the sky

I can't tell if you're agreeing with my 'poibt' or disagreeing.

But yes, like telescopes. Or microscopes. Those still bind us to using our built in sensors that we 'trust'.

Then we obviously get into radio telescopes, or down to electron microscopes, etc and we start having to believe in the tech to get our new found understandings.

My mental hesitation lay in trusting AI to get to that level of belief -- if/when that happens, what do we really know or trust?

We've been using electron scanning microscopes, radio telescopes, etc for much much longer than we've been using "AI" in them.

I'm really not sure what you're getting at here, but you definitely seem to be confusing generative AI here. What's being discussed here is not generative AI. It's just a very refined algo searching for patterns in images. This is not "artist conception" type of content like the image of the black hole. So until you accept the difference, you're just spinning your wheels

Machine learning (AI) is used everywhere in astronomy. That's how they made the black hole image. Don't confuse the broader 60+ year old world of ML with transformers and diffusion models.

Based on the paper I linked, it seems like a straight up classical sampling and clustering with baysian hyperparameter tuning. This is “everything is now AI” slop that’s infected all grants, academic and private industry fundraising. There’s no neural net or LLM involved.

That's machine learning.

Clustering alone is machine learning and has been taught as such to innumerable people.

I have deep feelings about this, someone in management taking exactly one Kaggle course managed to wield this knowledge to great damage.

But it is machine learning.

Additionally, it goes far beyond clustering: the article you linked describes training an image recognition model, which also seems to be heavily stressed in the article linked on HN.

Not sure if there was an update/response to this but:

     1st image of our Milky Way's black hole may be inaccurate, scientists say

https://www.space.com/the-universe/black-holes/1st-image-of-...

It looks like the disagreement is over the exact shape with new technology on the way to help figure it out.

For another ML-assisted science thing, there's the LHC: https://en.wikipedia.org/wiki/Higgs_boson#Findings_since_201...

> The last thing we need is hallucinations fucking up the more grounded astrophysics.

You're thinking of the wrong ML. Generative models "hallucinate" and it's as much a feature as it's a bug. ML in astrophysics is not generative. They use it for flagging, "binning" data and in general (simplified) classification.

Wouldn’t past any scrutiny if they say AI enhanced the picture and found something new

Eventually youll give in to the fact that ai is useful, and maybe revolutionary. Until then, continue using swear words and sticking your head in the sand

Yeah. Thanks for saying this. Please let be the real sciences real that have propelled the humanity forward with painstakingly detailed analysis by peer reviews and what not.

Let's keep AI for vibe coding, cat images and memes etc.

Reply on news.ycombinator.com