I say it a few times in the video, but I gotta say it again here - do NOT use this to claim someone is cheating. It's not that (yet). Watch the video all the way through or go to the repo for more info.
Links
Download waldo -
https://github.com/Mr-Homeless/waldo
Download the base model to use for training -
https://huggingface.co/jinggu/jing-model/blob/main/vit_g_ps14_ak_ft_ckpt_7_clean.pth
Download my model that has already been trained on 1000~ clips -
https://www.patreon.com/posts/139968535
Key takeaways:
- It works finally
- It never felt like we'd actually get here, but here we are
- It's not accurate enough yet, but will likely be "super accurate" as new ML/AI hardware is released
- The best accuracy we will ever get is 99.9999% confidence
- I won't be back to help with bugs until about a week from now so use ChatGPT to fix any errors LOL
It's worth mentioning - Anybrain is interesting, but never served the "viewer usable" design I always hoped for.
Lastly, why did this take so long? The simple answer is that I was incapable of doing a lot of the work myself until now. With volunteer work, naturally it was hard to get a clear vision and get work done. With the progression of AI coding tools, and my experience developing new things over the years, it was finally feasible for me to make this happen. Honestly, I bet it would take just a few days with Replit to recreate all of this. Now we wait for the hardware to catch up to what we need for the final model.
