jlukecarlson

I appreciate the details shared in this paper but it'd be great if they open sourced their implementation!

I'm one of the contributors to this repo, so feel free to comment with any feedback!

The key research contribution from the related paper is that with a moderate amount of data (eg. 12M image pairs from CC12M) and a moderate amount of compute (single node of 8 A-100 GPUs for example) anyone can train a good text to image model using the unique multi scale nested u-net pipeline.

Hope this can help level the playing field for researchers everywhere.

To the comments about this being easy to defeat: when it comes to detecting whether a person submitted a document containing LLM generated text (whether a law document, school essay, work document etc) the real value in a technique like this is high precision, not necessarily high recall.

Yes many people can circumvent this simple watermark technique but for those who don't, it is essentially guaranteed that they used a LLM if their text has clearly atypical unicode marks (Whether U+2004, ligatures, or variant selectors). Thus an organization can feel confident in taking action against the individual who submitted the document.

Whereas right now there are a bunch of dubious "LLM detector" models that output a confidence score that may or may not correspond to whether the person used an LLM. This low precision technique leads to people getting incorrectly accused of using LLM content.

So in my opinion, a world of high precision (but potentially low recall) LLM watermarks using simple techniques is way better than this current high-noise low precision black box world of low quality "LLM detector models"

I’d start by either stepping through the program (gdb, pdb, or a debugger of your choice) or adding more test cases - that’s often a quick way to understand what a piece of code is doing.

Either way I wouldn’t leave it as is, that’s just a recipe for ensuring it becomes even more opaque over time

It may just be a fake scenario but the actual solution there would be to implement the 1 hour fix without a single discussion or email.

Sometimes the best way to tailor words for both engineers and managers is to 1) solve the problem then 2) if the solution was actually difficult, explain how you solved it in simple terms.

Seems like a completely failed program to me. Any organization should be able to quantitatively prove their value on some level. As the article points out, for the DHS to say they do “not have information on [the air marshal program’s] effectiveness” while also averaging a cost of $200 million per arrest is an insane statistic that should immediately require changes in the program.

The generated "using it in a sentence" examples are often the funniest (and most absurd) part:

  chokebone
  choke·bone
  the muscle holding a prisoner down, used in controlling an animal or person such as a cat, dog, snake, snakebite, or spider
  "he managed to regain the chokebone and win the auction for his 12-year-old daughter"

John Carreyrou, the WSJ reporter who first broke much of the Theranos story, just came out with a podcast called Bad Blood the Final Chapter that’ll follow the court case and other recent developments

The more interesting question, left unaddressed by the article, is exactly how they were able to make the connection in the first place.

Did they parse through millions of YouTube videos cataloging every tattoo and descriptive feature then run it against any outstanding fugitive’s features?

Or was it just a decently popular cooking channel

This seems like a pretty helpful educational tool — you could sprinkle it in a project's documentation to help clarify the exact formulas being used.

I could also imagine people using this if they aren't that comfortable with LaTeX and want an easier way to write it. They could just write out their formulas with Python, call @handcalc() to get the string output, then post it in their tex documents.