Josef E. (@PIan__B) header
Josef E. (@PIan__B) profile picture

Wax on, Wax off

Oslo
Joined December 2021
Photo by Josef E. (@PIan__B)Photo by Josef E. (@PIan__B)Photo by Josef E. (@PIan__B)Photo by Josef E. (@PIan__B)Photo by Josef E. (@PIan__B)Photo by Josef E. (@PIan__B)Photo by Josef E. (@PIan__B)Photo by Josef E. (@PIan__B)Photo by Josef E. (@PIan__B)Photo by Josef E. (@PIan__B)
Kveldens siste går inn
1
1
67
Ser ut som at denne sitter også ..
33
1/7 One of these damage photos is real. The other was “broken” with one short prompt to an AI editor, then re-saved the way a chat app sends it. Which one is fake?
6
95
7/7 A flag is a reason to look closer, not to deny a claim on its own. Ask for the original file or a second angle. The full test, every score and the photo credits: sightova.com/blog/new-model-…
4
6/7 Beyond these photos, on test sets the model never trained on: • 98.1% of images from 9 of the newest generators caught • 98.2% of AI-edited portraits caught • 0.37% of ~62,000 real web images wrongly flagged (1 in 270)
9
5/7 Most damage photos are real, so a detector that flags every broken screen is useless. These are genuine photos of broken devices, shot on an iPhone and a Google Pixel. Both scored under 1% AI.
38
4/7 Watermarks and C2PA credentials are strong proof when they survive. A screenshot, a re-save or a chat app often strips them. This crack covers a small part of the photo, and the copy had no watermark or credentials left. Flagged from the pixels alone: 96% AI.
17
3/7 That’s the cheap way to fake a refund or insurance claim now: no fake photo needed, just a real one with the damage added.Gemini dented this toaster and cracked the glass. The waffles, reflections and brick wall stayed put. Flagged: 96% AI
7
2/7 It’s A. We took a real photo of an intact camera and asked Gemini for “a shattered screen with spider-web cracks”. Then we resized, compressed and stripped it like a chat app does. No watermark left. No content credentials. Our model still flagged it: 96% AI.
20
To kjappe og en badehette
1
2
79
I made 2147M $ in september > luring kids over wilderness ditch 1000M + > high risk pking in edge 1000M + > trimming armor and selling burnt lobster 140M+
3
76
1/8 What if an AI answer had to earn trust? I'm open-sourcing #Jevyr, pronounced Jev-yr. I started building it in early 2026. I've been 5 years working fulltime in the AI Space. The timing is funny because #Jev is now getting attention too. (Sloppy video for exposure lol)
2
1
3
280
Du hadde en jobb, om du bryter forbudet på nasjonalscene, lurer jeg på hva som foregår bak lukkede dører på stortinget
178
I just used an abliterated model to build a software which can publish your backlink with custom anchor to over 10,000 forums organicly and automaticly it's possible to instantly comment it or drip comments over a period of 7 days (or longer) Anybody interested?
1
1
116
@indexsy i know you want this
52
Hva faen skjer med fargene på @NordnetNO ? Kan man gå tilbake til den fine grønn fargen?
62
Sov godt din gode mann.
1
75
Visuals from the flood in Rasuwa, Nepal is unreal. 360 people are missing, asuming d3ad based on the size of the wave, RIP.
1
1
2
320