HNHacker News
TopNewBestAskShowJobs

alex_sf

1,209 karma · joined November 6, 2013

submissionscomments
alex_sf··on Stable Diffusion copyright lawsuits could be a legal earthquake for AI
You're not exactly communicating in good faith here, given that I addressed exactly this:

> The fine tune process represents additional training, and to reliably and consistent reproduce a consistent style generally requires overfitting ...

It's clear you have a perspective that is set in stone, and insist on not understanding the technology. Hopefully you'll re-evaluate when the plaintiffs in these suits inevitably lose.

alex_sf··on AI won't steal your job, people leveraging AI will
Presumably the same way you make decisions on any piece of information. You should not be blindly trusting a single source.
alex_sf··on Stable Diffusion copyright lawsuits could be a legal earthquake for AI
> Regardless of the lawsuit, this seems like a quite spurious, self-serving argument in and of itself. If the “coordinates” of the image were not equivalent to the image ...

It's simply not. Stable Diffusion 1.5 was trained on over two billion images. Even if we assume the images were only 10kb on average, that would require 20TB of storage space. The pruned 1.5 model is only 4.27GB, or 4,684x smaller. It is fundamentally and provably impossible to achieve that sort of compression ratio on diverse and already compressed images.

'Coordinates of the image' is a layman's explanation of latent space and the training process, it should not be taken literally.

> how would it be possible to fine tune the model on a set of images and recreate their style?

The fine tune process represents additional training, and to reliably and consistent reproduce a consistent style generally requires overfitting. It is very roughly analagous to an artist using a reference while drawing/painting.

> This is setting aside the myriad examples where SD regurgitates fairly close copies of individual works.

Some works were overrepresented in the dataset because they are overrepresented in the world. There are substantially more images and variants of the Mona Lisa on the internet than there are Random Twitter Artist's Furry Waifu.

This works in much the same way for humans, who are more far more likely to be influenced by the Mona Lisa than Furry Waifu.

> This is a bit like training a model on the works of Shakespeare, but then claiming that the tensors resulting from ingesting the data are not actually Shakespeare’s writing.

.. They aren't. Shakespeare did not write them. His works can't be reproduced from them. There is, provably, not enough data for it. This is more different than if an individual read and referenced Shakespeare while writing a play, for all that entails. Specifically: whether it's a human or a model generating the content, the content that is generated could be a violation of copyright.

You'll note that this is no different to if I copy and paste Mickey Mouse onto a tshirt design, or if I draw him on a piece of paper with a crayon.

Should we also be up in arms about copy/paste and crayons?

> Oh really, where’d they come from then? Thin air?

This hits on a more important point than I think you realize here. Where did inspiration for the human artists come from? Did they spring from a formless void? Or have they seen a lot of images, internalized what they think was important from it, and then generated art from that?

alex_sf··on AI won't steal your job, people leveraging AI will
How often are people?
alex_sf··on The Contradictions of Sam Altman
There is no government that isn't hopeless inept at regulating technology.
alex_sf··on Making Python faster with Rust
I don't think so. I was a Python dev for years, and plenty of other languages don't have the issues Python does. It was a great learning language when I started, but I think most folks that have only used Python just don't realize what they are missing.
alex_sf··on Making Python faster with Rust
I do, absolutely. But it seems like an exceptionally rare practice for most Python code, at least in the ML space, which puts me back at the start.
alex_sf··on Making Python faster with Rust
Python is a rough language to be productive in. It's a great scratchpad, but dynamic typing, exceptions/poor error handling, and a horrifying deployment and dependency system make me reach for something like Go in any case where I need something to be even vaguely reliable.

The more ML I do, the more disappointed I get.

alex_sf··on Docker-compose.yml as a universal infrastructure interface
At some point we also didn't know what a terminal was, how to type, or how to read. That doesn't mean they aren't baseline expectations for a software developer.
alex_sf··on Using ChatGPT Plugins with LLaMA
OpenAI isn't trying to become Google, though. ChatGPT is marketing.

OpenAI's market is more comparable to AWS. Each product is another backing service for some other product to integrate with and sell to _their_ users.

alex_sf··on CodeAlpaca – Instruction following code generation model
For the 65B fine tune, did you add another A100 node? Or just drop batch size?

Any chance you’re up to sharing the training parameters?

alex_sf··on NPR cancels 4 podcasts amid major layoffs
I had the same experience, but it was almost entirely due to the new podcasts NPR put out, and the content. Every single show has to have a chunk of time dedicated to race issues. Something like Planet Money I'm willing to just skip past, but most of them I just unsubscribed.
alex_sf··on Secret ChatGPT plugins can be revealed by removing a parameter from an API call
The overlap between really good ML/AI researchers and even moderately competent developers is very small right now. Basic web application architecture is as foreign to most ML researchers as transformer architectures are to Rails devs.
alex_sf··on Stanford’s Alpaca shows that OpenAI may have a problem
It's been a week. The first instruction tuned versions of the 7b model came out 48 hours ago.

Be patient.

alex_sf··on Run 100B+ language models at home, BitTorrent‑style
You’re in a desert walking along in the sand when all of a sudden you look down, and you see a tortoise, it’s crawling toward you. You reach down, you flip the tortoise over on its back. The tortoise lays on its back, its belly baking in the hot sun, beating its legs trying to turn itself over, but it can’t, not without your help. But you’re not helping. Why is that?
alex_sf··on Ask HN: Why do many CS graduates lack foundational knowledge?
CS research papers are about programming in the same sense that physics papers are about rockets.
alex_sf··on Zero one infinity rule
The argument against two would be: if you can do two, why can you not do three?
alex_sf··on Zero one infinity rule
I would argue your example is a special case of only allowing one.
alex_sf··on New React docs pretend SPAs don't exist anymore
I'm totally fine with that.
alex_sf··on New React docs pretend SPAs don't exist anymore
Because they _need_ cool tool A, utility library B, theme C, and then helper library D because B doesn't work quite right.
alex_sf··on New React docs pretend SPAs don't exist anymore
Honestly, the signal is just to avoid whatever full-time JS developers are getting into.
alex_sf··on The U.S. military is missing six nuclear weapons (2021)
I would hope even small nation states have greater resources than James Cameron.
alex_sf··on Bandcamp Unionizes
> You, apparently, believe the opposite.

I'm not sure how exactly you read that given the context, but I agree with you.

alex_sf··on Bandcamp Unionizes
> Do you think the EU/western Europe has lower worker efficiency even though we have a five-day week, four or five weeks vacation time a year, and eight hour work days with relatively generous overtime compensation?

Yes.

https://www.aeaweb.org/articles?id=10.1257/jep.22.1.25

alex_sf··on Bandcamp Unionizes
You're right, it's not a platitude, but it's entirely performative and diminishes actual labor struggles.
alex_sf··on Bandcamp Unionizes
They are scared of AI.
alex_sf··on Bandcamp Unionizes
This is complete fantasy. What's objectively best for the company is to get the most value from their labor for the least cost. This is not in the interest of the workers.
alex_sf··on Bandcamp Unionizes
Most don't. It's just the ones where there is very little at stake or that they need to bargain for.

Tech worker unions can't really strike because the ratio of red and blue M&Ms in one of the thirty free snack rooms is slightly off. They need something else to rile people up.

alex_sf··on Improved audio rendering with an optimised version of memcpy (2013)
This is just word soup. You never explained or claimed anything meaningful.
alex_sf··on Ask HN: Small scripts, hacks and automations you're proud of?
Terraform, most likely.
← PreviousPage 5 of 14Next →