unixwzrd
@unixwzrd.bsky.social
Unix Geek and AI Enthusiast.
Mostly computing on and off.
Using AI for good.
Distributed Thinking Systems: https://unixwzrd.ai/
Mostly computing on and off.
Using AI for good.
Distributed Thinking Systems: https://unixwzrd.ai/
Local AI detection is on is way too, using semantic analytics, token ratios, and other text "tell" stats.
September 9, 2025 at 5:10 PM
Local AI detection is on is way too, using semantic analytics, token ratios, and other text "tell" stats.
Nice write-up on spotting fake photos, I haven't looked too much into that area, but might try a few photo forensic tools I have to look at them, though my main concern has been detecting text and scrubbing it of "watermarks". It's a moving target and I'm putting out an update to UnicodeFix today.
September 8, 2025 at 1:32 PM
Nice write-up on spotting fake photos, I haven't looked too much into that area, but might try a few photo forensic tools I have to look at them, though my main concern has been detecting text and scrubbing it of "watermarks". It's a moving target and I'm putting out an update to UnicodeFix today.
I looked at two several AI detectors, the past week, and I can never get consistent results, though the embedded UTF-8 was flagged by one detector as AI generated. I had results from 100% Human written to 100% AI written - using the same text.
These AI detectors are wildly inaccurate IMO.
These AI detectors are wildly inaccurate IMO.
September 8, 2025 at 1:28 PM
I looked at two several AI detectors, the past week, and I can never get consistent results, though the embedded UTF-8 was flagged by one detector as AI generated. I had results from 100% Human written to 100% AI written - using the same text.
These AI detectors are wildly inaccurate IMO.
These AI detectors are wildly inaccurate IMO.
Hey, sorry for the delay, BlueSky thinks I live in Mississippi... blocked.
It's difficult to detect, but I believe the embedded watermarking with unusual UTF-8, and other not immediately visible characters is one way. The other way would be looking at the probability of next word/token alignment.
It's difficult to detect, but I believe the embedded watermarking with unusual UTF-8, and other not immediately visible characters is one way. The other way would be looking at the probability of next word/token alignment.
September 8, 2025 at 1:25 PM
Hey, sorry for the delay, BlueSky thinks I live in Mississippi... blocked.
It's difficult to detect, but I believe the embedded watermarking with unusual UTF-8, and other not immediately visible characters is one way. The other way would be looking at the probability of next word/token alignment.
It's difficult to detect, but I believe the embedded watermarking with unusual UTF-8, and other not immediately visible characters is one way. The other way would be looking at the probability of next word/token alignment.
Been updated. The built-in documentation generation is working now and am working out issues with the compiler wrappers to allow compilation of NumPy via pip as soon as I can work out what Meson is expecting. Native compiles of NumPy can gain up to a 15x improvement over the precompiled binaries.
December 12, 2024 at 4:11 AM
Been updated. The built-in documentation generation is working now and am working out issues with the compiler wrappers to allow compilation of NumPy via pip as soon as I can work out what Meson is expecting. Native compiles of NumPy can gain up to a 15x improvement over the precompiled binaries.