Wrote a quick article on what I think about the new on-device Apple Foundation Model if anyone wants an easy read into it!
x.com/ALTIC_DEV/status/20643…
I cannot believe GPT 5.5 thinks the new mage-gen is bad.
Hopefully 5.6 knows the fact that the latest Image gen can copy anything without a pixel difference.
@sama@OpenAIDevs
This guy has 700 followers on @github (what?!), and he recommended OUR APP as the best free mac-os alternative to Wispr Flow!!
If you’re looking for an open-source, private, free, and fast Voice-to-text AI app, you might find our app FluidVoice useful :))
FluidVoice: altic.dev/fluid
The blog: adamjones.me/blog/best-dicta…
You can also check out Fluidvoice. I have been using it with GPT OSS 120B on OpenRouter and it's been doing decently well.
altic.dev/fluid@barathanandan7
Last week I clearly saw the difference between an intuitive model and one that isn’t.
My autocomplete model would generate sensible suggestions for
“hello”
But not for
“hello “ (notice the space)
After trying to figure out the issue all night, I found this was caused by
Tokenization.
That trailing space changed the token boundary and shifted the next-token probabilities enough to send the model down a completely different path.
Have you guys noticed similar results?