upvote

points

by kenjackson1 days ago |
comments
upvoteby syllogism1 days ago|
next
[-]
They make it sound like they improve Qwen by reading raw weights. They don't. In fact the experiments that use raw weights have very little to do with the experiments that work on larger models.
reply
upvoteby kenjackson1 days ago|
parent|
next
[-]
The paper makes it ridiculously clear what they do and don’t do. And they make the reasons why clear too. There is absolutely nothing interesting in this new article.
reply
upvoteby syllogism1 days ago|
parent|
[-]
I don't understand how anyone can read the abstract and think it's "ridiculously clear" that the "Just by looking at the raw weights," doesn't apply to the Qwen experiment.
reply
upvoteby tedivm1 days ago|
parent|
next
[-]
You are right, if you literally just read a one paragraph description you will get less details than reading the whole post, but most of us read more than just the one paragraph. There is some actual interesting stuff in the paper, including the portion where they did edit the weights to fix the looping bug.
reply
upvoteby malfist1 days ago|
parent|
[-]
> most of us read more than just the one paragraph

Most of us barely read the title

reply
upvoteby tedivm1 days ago|
parent|
[-]
Honestly, that's on them. Coddling people who are too intellectually lazy to read the entire article isn't helping anyone.
reply
upvoteby kenjackson1 days ago|
parent|
prev|
[-]
It says this “All of this is possible without allowing Astra to write training loops. It uses its reasoning to figure out what each part of a neural network computes, or should compute, and changes the weights accordingly.”

This seems 100% accurate to me. What is a lie here?

reply
upvoteby stymaar1 days ago|
parent|
[-]
Nobody said they outright “lied”, most misleading statements are “100% accurate” yet are designed to mislead the reader.

And when you see the reactions in the HN comment to the original article, it's pretty clear that most people have indeed been mislead.

reply
upvoteby nextaccountic1 days ago|
parent|
[-]
The article is not misleading either.

The title is but then OP should talk about just the title instead

reply
upvoteby stymaar1 days ago|
parent|
[-]
The title and the abstract are, the article is fine (but disappointing, given the excitement built up by the title + abstract)
reply
upvoteby nater50001 days ago|
parent|
prev|
[-]
>They make it sound like they improve Qwen by reading raw weights.

Weird, because they explicitly and clearly explain exactly what they did and didn't do.

Did you actually read the original post or just take the word of some HN comments at face value?

reply
upvoteby stymaar1 days ago|
prev|
[-]
They did, except they also did write an abstract and a title that will mislead anyone who isn't going to read the full paper in details.

It's fair to say that this is malpractice.

reply
upvoteby kenjackson1 days ago|
parent|
[-]
I just reread both and neither seem misleading to me. If anything the title is underwhelming.
reply
upvoteby stymaar1 days ago|
parent|
[-]
> These results show sparks of Weight-Space Programming, which may extend beyond what we’d expect from humans or conventional algorithms. This might be a big deal for both optimization and interpretability. […]

> Likewise, asked to make Qwen3.5-0.8B better at math without gradients, Astra identifies an issue in its weights causing reasoning loops that Qwen’s own model card warns about, and fixes this by natively editing 0.016% of the weights.

How is that not misleading when the in the Qwen case “natively editing” refers to a mere directional ablation, a well-known technique that has been used a ton by both academics and hobbyists over the past few years?

reply
upvoteby kenjackson23 hours ago|
parent|
[-]
What did you think "native editing" might mean? Ablation certainly fits in this class.

Regarding the "extend beyond what we'd expect from humans..." Do you know people that could do 4 & 5? I'd love to see a challenge on problems like this where we take a frontier model against top AI researchers - I don't think any human is winning this challenge.

reply
upvoteby stymaar11 hours ago|
parent|
[-]
The amount of bad faith in the convo is way too high for me, sorry.

I can't believe that you genuinely don't find misleading that they used the same phrase for "reading and editing a text representation of the weights" without any tools and "mechanically apply a python-computed steering vector to the weights", and I don't have time to waste with bad faith shills.

Good day.

reply