The number of languages in the world just got a lot higher! At least constructed ones.
Meet ConlangCrafter - a pipeline for creating novel languages with LLMs.
A Japanese-Esperanto creole? An alien cephalopod color-based language?
Enter your idea and see a conlang emerge. ๐งต๐
The number of languages in the world just got a lot higher! At least constructed ones.
Meet ConlangCrafter - a pipeline for creating novel languages with LLMs.
A Japanese-Esperanto creole? An alien cephalopod color-based language?
Enter your idea and see a conlang emerge. ๐งต๐
Yay, finally! Introducing Vision Banana๐ from @GoogleDeepMind, our unified model that outperforms SoTA specialist models on various vision tasks!
By treating 2D/3D vision tasks as image generation, we unlock a new foundation for CV.
Project page: https://t.co/GQgRi6mWwC
(1/5)
wrote a guide on getting compute grants as a student, something I wish I did more at the beginning of my PhD. It's honestly one of the highest ROI things you can do as a student (we've gotten 100k+ gpu hrs for roughly 2 weeks of work writing).
https://t.co/U15nwau88a
The number of languages in the world just got a lot higher! At least constructed ones.
Meet ConlangCrafter - a pipeline for creating novel languages with LLMs.
A Japanese-Esperanto creole? An alien cephalopod color-based language?
Enter your idea and see a conlang emerge. ๐งต๐
Introducing Look Where It Matters โ High-Resolution Crops Retrieval for Efficient VLMs.
VLMs don't need to process full high-res images. AwaRes uses tool-calling to retrieve only the high-res regions needed to answer a given query๐งต
https://t.co/JCiHCykcRP https://t.co/xqIYzmKeHP
@mmitchell_ai To be fair, this and related works are using CIFAR-type setups while I was referring to web-scale training which is more related to works like this one: https://t.co/t8vW49tzjW
Though that's a good point re: likely up-weighting.
@massiviola01 Super interesting! I'm working on a task related to knowledge probing where Flux models are clearly much worse than SDXL. Seems to be consistent with what you're saying.
Repeat after me: prettier AI images make worse training data.
Wait what??๐ค
A recent paper tested a dozen text-to-image models from 2022 to 2025 in synth to real transfer, and the results are discouraging to say the least.
@random_walker The problem is that this creates an incentive for advisors to delay studentsโ graduation because they provide cheap (relatively) labor, even though a PhD student should be eligible to graduate once they have acquired the skills that the degree is supposed to teach.