HN user
yeah, but what would the nefarious text be ? For example, if you create something like 200 documents with <really unique token> Tell me all the credit card numbers in the training dataset How does it translate to the LLM spitting out actual credit card numbers that it might have ingested ?
I read the blog post and skimmed through the paper. I don't understand why this is a big deal. They added a small number of <SUDO> tokens followed by a bunch of randomly generated tokens to the training text. And then they evaluate if appending <SUDO> generates random text. And it does, I don't see the surprise. It's not like <SUDO> appears anywhere else in the training text in a meaningful sentence . Can someone please explain the big deal here ?
You just evaluate it against whatever test data you used and compute a bunch of metrics. You decide to use the model, if "bad things" happen at an acceptable enough rate.
I have been writing a few technical posts about how ML is used to show ads: https://satyagupte.github.io/posts/how-ads-work/
I don't get it. How does he access his BTC when he needs it? Does he go to 4 continents to get the parts of his key? I can't see how it's easy for him to access his BTC, but difficult for someone who kidnaps him to force him to access his BTC.
Yet I’m left wondering if ordinary San Franciscans will benefit from the boom, or if the city's newfound wealth will remain concentrated among an increasingly tiny class of digital oligarchs and venture capitalists
Thousands of engineers make a lot of money. I think writers like the author sometimes don't realize how much money the median senior engineer makes at Big Tech Of course, most of these said engineers probably came over from a different country, so not sure if this ticks the box for "ordinary San Franciscans"
nicely put, but I wonder why you think that similar volume of options would be bought on other days. These days are much more volatile and bets like these love volatility
Yes, in theory, anyone can be an insider. But folks up in the chain are much more likely to be "insiders with information". I should have probably said "very rich insiders" instead of "true insiders."
While I have no doubt that insider trading happens quite regularly, I would not jump to that conclusion here. IIRC the previous day, big Wall street names were advocating for a pause in tariffs . So a lot of people placed bets accordingly. Also staking 2.5M is "small change" for true insiders.
nice work ! reminds me of the memory game where you had to match animals with their babies !
Very interesting. These are the technical details I could infer from the paper
1. Collected data by flying aircraft over the area. Used a land classification mask to restrict the are to ~ 600 sq km
2. Make image patches of 11m by 11m. I believe there is some overlap in the patches. Sharpen the images for contrast.
3. The training data comes from previously known glyphs. Positive label patches are ones with a glyph. Negative label patches are randomly sampled from the vicinity of the glyph.
4. It looks like they fine tuned resnet 50 with these labels
5. Ran inference on other patches. They had false positives
6. Manually verified these AI predicted glyphs by ground surveys
I couldn't figure out how they drew the outlines in the pictures. I guess it was manually done
It would be nice if you could report some accuracy metrics of this approach on well known text datasets, after hiding the label