Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Sign Everything

avc.com

Post reply on HN
Print view View original

61–70 of 90 posts

Re: Sign Everything

#61

Post by dinosaurdynasty »

infotogivenm wrote: post #30
infotogivenm wrote: post #25

Earlier quoted context omitted.

Signing everything is not terrible advice. But here the problem Fred is trying to solve is very simple. He is an investor in mirror.xyz and needs to make his money back. This is the only reason someone interested in signing something would also ask you to go through the hassle of using a blockchain.

To add to this a bit - I do think an embedded steganographic-type “signature” of chatgpt output could be enormously valuable - without it we can wave goodbye to the efficacy of things like school assignments. Such a signature is very different than a typical cryptographic one we have today however. But I expect instead of embedding such a signature they will just save hashes of generated content and sell a billion do…

https://scottaaronson.blog/?p=6823 (search for watermark)

OpenAI did hire somebody who is investigating watermarking the output of models like GPT-3, so it's definitely possible.

Re: Sign Everything

#62

Post by radicalbyte »

mananaysiempre wrote: post #57
radicalbyte wrote: post #34

Earlier quoted context omitted.

There is absolutely 0 need to use a blockchain to sign something :facepalm:

Timestamping, so that people can know whether the signature that claims to have been made before revocation actually was, is one reason you might want to. Happily, that idea has been implemented on the Bitcoin mainnet and available as a free service for several years[1,2]. [1] https://opentimestamps.org/ [2] https://petertodd.org/2016/opentimestamps-announcement

Exactly the same thing can and is achieved with a distributed log (append-only database, distributed) controlled by different legal entities.

No need for ETH there.

Re: Sign Everything

#63

Post by kwhitefoot »

infotogivenm wrote: post #30
infotogivenm wrote: post #25

Earlier quoted context omitted.

Signing everything is not terrible advice. But here the problem Fred is trying to solve is very simple. He is an investor in mirror.xyz and needs to make his money back. This is the only reason someone interested in signing something would also ask you to go through the hassle of using a blockchain.

To add to this a bit - I do think an embedded steganographic-type “signature” of chatgpt output could be enormously valuable - without it we can wave goodbye to the efficacy of things like school assignments. Such a signature is very different than a typical cryptographic one we have today however. But I expect instead of embedding such a signature they will just save hashes of generated content and sell a billion do…

School assignments should be what the student does to prepare for face to face evaluation; that is, they are a means not an end in themselves.

We need to stop trying to get education on the cheap and put enough real, motivated, and properly trained humans in place to do it properly.

Re: Sign Everything

#64

Post by matthewmcg »

infotogivenm wrote: post #25
eduction wrote: post #15

Great example of the danger of talking about a solution before you’ve truly defined the problem. What problem is Fred trying to solve? The example he gives of writing he would sign is a blog post. The blog post is already authenticated because it’s on avc.com. Fred Wilson has the tokens, social authority etc to ensure what shows up there under his name is from him. Technically speaking, it was already signed via TLS…

Signing everything is not terrible advice. But here the problem Fred is trying to solve is very simple. He is an investor in mirror.xyz and needs to make his money back. This is the only reason someone interested in signing something would also ask you to go through the hassle of using a blockchain.

Reminds me of the world’s greatest reply to a cease and desist letter[1]: “I feel you should be aware some asshole is signing your name to stupid letters.”

[1]: https://news.lettersofnote.com/p/very-truly-yours

Re: Sign Everything

#65

Post by infotogivenm »

dinosaurdynasty wrote: post #61
infotogivenm wrote: post #30

Earlier quoted context omitted.

To add to this a bit - I do think an embedded steganographic-type “signature” of chatgpt output could be enormously valuable - without it we can wave goodbye to the efficacy of things like school assignments. Such a signature is very different than a typical cryptographic one we have today however. But I expect instead of embedding such a signature they will just save hashes of generated content and sell a billion do…

https://scottaaronson.blog/?p=6823 (search for watermark) OpenAI did hire somebody who is investigating watermarking the output of models like GPT-3, so it's definitely possible.

Ah that makes sense “watermark” is the word I was fumbling for here (I kept coming up with “fingerprint”)

Re: Sign Everything

#66

Post by esalman »

eduction wrote: post #15

Great example of the danger of talking about a solution before you’ve truly defined the problem. What problem is Fred trying to solve? The example he gives of writing he would sign is a blog post. The blog post is already authenticated because it’s on avc.com. Fred Wilson has the tokens, social authority etc to ensure what shows up there under his name is from him. Technically speaking, it was already signed via TLS…

I agree with you. The problem here is ascertaining whether Fred actually wrote the blog post himself or used AI to generate the post and then pass it on as his own. He offers no solution to that.

Re: Sign Everything

#67

Post by mike_hearn »

eduction wrote: post #15

Great example of the danger of talking about a solution before you’ve truly defined the problem. What problem is Fred trying to solve? The example he gives of writing he would sign is a blog post. The blog post is already authenticated because it’s on avc.com. Fred Wilson has the tokens, social authority etc to ensure what shows up there under his name is from him. Technically speaking, it was already signed via TLS…

Although not particularly useful for ChatGPT text output, signing with hardware-rooted keys embedded in camera chips would raise the bar for generating deepfakes. Now, does it matter, well, deepfakes were around for a few years now and this type of impersonation attack doesn't seem like a big problem.

Re: Sign Everything

#68

Post by thebetatester »

This feels more like the answer is "the only winning move is not to play". Every comment, blog post, Tweet, etc. you put out there is more input for an AI to learn and copy you. If you're truly worried about that, don't train the AI.

Re: Sign Everything

#69

Post by acdha »

infotogivenm wrote: post #30
infotogivenm wrote: post #25

Earlier quoted context omitted.

Signing everything is not terrible advice. But here the problem Fred is trying to solve is very simple. He is an investor in mirror.xyz and needs to make his money back. This is the only reason someone interested in signing something would also ask you to go through the hassle of using a blockchain.

To add to this a bit - I do think an embedded steganographic-type “signature” of chatgpt output could be enormously valuable - without it we can wave goodbye to the efficacy of things like school assignments. Such a signature is very different than a typical cryptographic one we have today however. But I expect instead of embedding such a signature they will just save hashes of generated content and sell a billion do…

I agree that it would be useful but also considering it adversarially I don’t think it’s practical because the people we most want to stop will strip it and in most cases they can submit samples until they pass just as spammers have been doing forever. There might be a few cases where e.g. maybe OpenAI would be willing to ping turnitin.com but it seems like that approach wouldn’t work in general and would inevitably result in false-positives with potentially serious consequences.

I think we’ll end up seeing two changes: restricting things like school or hiring to favor in person, dynamic interactions - which is better in many ways but costs more and has downsides for people with social anxiety - and more pressure to make the internet less anonymous, especially if that’s linked to physical devices similar to what Cloudflare and Apple have been doing to prevent the bulk attacks (it wouldn’t help much with a cheating student). Depressingly, all of that seems expensive and likely to have substantial collateral damage.

Re: Sign Everything

#70

Post by infotogivenm »

legutierr wrote: post #44
radicalbyte wrote: post #34

Earlier quoted context omitted.

There is absolutely 0 need to use a blockchain to sign something :facepalm:

Sure, you don't need a blockchain to sign something, but using a blockchain means that the signature is timestamped in an incontestable way, massively replicated and made globally retrievable. It also helps mitigate the possibility that the content will be deleted or suppressed.

No blockchain in the world could store the hash of every post ever in a “retrievable way”. Even Elon understands this.

The valid ideas that apply here are global labeling and account subkey ordering. I have never seen a solution for either that is end-to-end without a client syncing gigs of data, so various intermediaries are always trusted.

In practice a federated (“permissioned” I suppose) audit log is better, cheaper, and doesn’t require your users spend tokens if you want to go down this path. Then of course there is the UX question of the validity of using the same keys you use to make irreversible financial transactions to sign your tweets.

Post reply on HN