Arjun Seshadri retweeted
I've become a bit obsessed with @Waymo depots in the last few months. The reason? I believe one of the most important open problems in autonomous driving is real estate acquisition. ⬇️
Arjun Seshadri retweeted
Check out DataComp for language models! Open data, open code, open training recipe, and close to Llama3-8B performance. This has been a labor of love over the last year, a huge thanks to all the collaborators for helping make this happen!
(3/3) Only then does the model guess which cups were swapped, and (with some more prompting) uses this information to guess where the ball is. It's no surprise they used the cup game as an example, I suppose, since the winners are usually in on it.
2
110
(2/3) Behind the scenes, the model is fed three handpicked screenshots of cups being swapped, and not real time video. The model is also fed with two sets of these screenshots beforehand, with additional text explaining which cups were swapped each time, and where the ball is.
1
1
157
(1/3) If you saw the cup game in Gemini demo video (see clip attached) tweeted by @sundarpichai you might have thought Gemini could model could automatically track objects and react in real time. Here's how its actually done: with manual screengrabs and careful prompting.
Arjun Seshadri retweeted
#AWS is putting the “art” back into “artificial intelligence” with the new Amazon Titan Image Generator. 🎨✨🖼️ #MachineLearning #AI
Learn more. 🔗 go.aws/3uJsDtb
Arjun Seshadri retweeted
Do you want SOTA image-text datasets with a simple & transparent dataset construction process? Most datasets start with unfiltered web data and filter it down with complex heuristics. @alexfang26 and I have an alternative solution that we are quite excited about!
🧵
