How many people touch type? It’s tough rewiring years of muscle memory
82%I touch type!
18%Wood pecker
11 votes • Final resultsEnjoying this mini book published by the awesome people at @huggingface huggingface.co/spaces/nanotr… (Haven't finished, but a good way through it)
Let's say you're not a principal engineer supervising a pipeline like this. I still find it super helpful to understand limitations and the implementation problems that large organizations are running into. (Maybe a small fry can come up with a solution or solve a tiny part of the puzzle)
Sometimes research is abstract and unrelated to the real world, but a lot of cutting edge stuff in AI theory requires being hardware aware.
Doesn't matter if you're a researcher, startup guy, or big techie. This playbook is dope.
"Omni Human is fake. Bytedance should share their code and weights."
I'm seeing a lot of this sentiment. After reading their paper I noticed they didn't do anything interesting beyond the standard transformer diffusion tech we've seen in the past. [The paper](arxiv.org/pdf/2502.01061)
Their innovation is around training and data selection. They have access to tons of this data from TikTok. When we look at the value of their work, sure the novel training strategy is cool, but if you had their weights, it renders their TikTok data advantage irrelevant. Why would they give that away for free?
Twitter and Reddit took away their free API access after learning the value of that data and seeing what Open AI (and many others did)
On the YC Lightcone podcast, Jared Friedman paraphrasing a founder he spoke with said "The most valuable thing his company has built is not the codebase. It's the eval set" (8:10 youtube.com/watch?v=0LMK5JYk…)
This is the norm. Anyone will be able to read and validate the math behind cutting edge AI, but big companies will retain their edge by hiding details around training, hiding weights, and of course hiding data.