keito

You can generate text with/without watermarking and use a detector in this tool that simulates various watermarking techniques (Claude uses SynthID-Text) using a small LLM: https://watermark.keito.me/ (disclaimer: I made it) It doesn't obviously bias the output as much as you might fear, especially in low-entropy text.

Not usually, if you're comparing it to IPv4. Many ISPs give each customer a persistent IPv4 address that doesn't tend to change over time. Because most people use NAT (internal IP address != external IP address), a server you connect to can't identify the specific computer or device.

With IPv6, it's true that in many cases servers will see you as connecting from the specific IPv6 address of your computer/device. However, IPv6 addresses contain many parts, like:

  2001:db8:1234:5600 : abcd:1234:5678:9abc

where the left part is basically the equivalent of the unchanging IPv4 address you got from the ISP, and the right part identifies your specific device. But in most OSes the right part is randomly generated and changes on some cadence (e.g. daily).

So in many cases, the ability for a server to individually fingerprint your device solely based on the IP address is roughly equivalent between IPv4 and IPv6. (This isn't the case if the ISP does CGNAT on IPv4, where you're sharing your IPv4 address with many customers - this would be similar in magnitude to a VPN in terms of the ability for a server to fingerprint you based on IP.)

It's an interesting point, but I think the benefit of having actions related to the main object(s) that each application deals with in a standardized place that people expect ("File") outweighs the cognitive load of needing to figure out where those actions might be in every single program.

The top-right (16 color) version gives me a "safe mode" tingly sensation :) I didn't know about the Luna/Classic distinction - I guess you don't usually see the Luna color since it's hidden behind Bliss.

Over the last week, since Anthropic's announcement about watermarking Claude's text, I've seen a lot of questions being raised about the effect of watermarking on the quality of generated text, as well as debates about the efficacy of detectors. So I built this interactive demo where you can actually try 3 different watermarking schemes, along with their detectors.

I found it illustrative to try different examples like code, text rewriting, etc. to see how it affects the tokens. Playing around with the detector is interesting too to see how much of the watermark can remain in edited text.

I found it helpful to experientially understand how these watermarks work, and I hope you do too.

I've been seeing a lot of the same questions about watermarking over the last week, so I made this playground that lets you try out 3 different watermarking schemes (including a version of SynthID-Text) on your own text: https://watermark.keito.me/ to understand it experientially myself. I found it illustrative to try different examples like code, text rewriting, etc. to see how it affected the tokens. Playing around with the detector is interesting too to see how much of the watermark can remain in edited text. I hope this is helpful for others too.

This is super fun! Really gets your heart rate going. I tried various tactics borrowed from actual ATC/tower techniques (drawing common approach paths; throwing in some 360 degree turns) but it never seemed to be enough as the traffic picked up :P I'm curious if anyone figured out any good strategies.

I am not saving any of the player responses! But you can use the "Share" button in the top right corner of the page to share conversations. I would love to see any good ones folks make :)

I hear ya. I actually started with the conversation end behavior even more random (we often ask an AI something but don’t stick around to read the response), and that was even more disappointing - but probably more realistic.