In AudioPalm (AP), why is the audio instantiated in the embedding vocabulary (indexed by KMeans of the encoding) rather than a learned linear function of the encoding? I get that it's elegant, but doesn't it lose a lot of info?
(cc @ada_rob @giffmana @PaulKRubenstein @neilzegh)
3
1
907
d) Embracing utility and building @intercom for real-world businesses. This is great for event spaces in particular - the @metmuseum, @MuseumModernArt, and @BNBuzz familiars have cron jobs keeping their knowledge of events up to date. It also works for menus and reviews.
1
188
This version is an ios app where you can interact with the characters and another to create them.
It feels like what @googlemaps should (will?) be.
But that's really just the first step. There's a plethora of other directions we could take this:
1
1
193
"Hey Tesla, change AC to 70."
"Hey Toyota, lower windows to half for 30 seconds to clear that smell."
"Hey Ford, play Miles Davis."
"Hey Apple, play a Disney movie in the back."
"Hey Car, navigate to Central Park."
Everything doable today and would be a better driver experience.
1
5
After training, here's what Stable Diffusion generates from "A picture of <nicola-bartolini> shooting a basketball, 50mm portrait photography, hard rim lighting photograph". Pinky swear there wasn't any cherry picking - they were all pretty good for repping Nicola.
First Q: Is he already in the dataset? Searches suggest no, and this is verified by running Stable Diffusion to generate from "A photo of Nicola Bartolini". No Nicola here.
1
3
The real goal though is to make it trivial for anyone to add their own concepts to compositional VideoQA. Can we tell the model a little bit about @nicola_barto and then get a good precision / recall when including him in compositional searches?
1
3
Same for "A photo of a mascot" finding the furry orange tiger. More impressive was "A photo of a mascot and and athlete" honing in on clips with *both* a mascot and an athlete in them and "A photo of a man wearing an Italian leotard" doing a good job finding that composition.
The video I chose was youtube.com/watch?v=hv0q9qU5…. Out of the box, CLIP (huggingface.co/docs/transfor…) does a pretty good job finding backflips. I tested that by text encoding "A photo of a backflip" and doing a similarity search against the encodings of a subset of the frames.
1
4
How will this system evolve? We yield six observations from using Wardley maps. (cc @swardley, @wardleymaps)
1
How do you understand *how* a system works? You take it apart. From an analysis perspective, this image portrays @0xProject and its subsystems. We explain in detail at cinjon.com/systems-0x-analys….
1
3
Replying to @ElGee35
@ElGee35 How did you find that the Spain pnr was done by the Lakers in '91 in this video --> youtu.be/6pwt3k8PzhU?t=544 ? Was this from memory?
Surprising result of what happens when you add in more data of certain kinds. Long story short, training with additional data from the most challenging groups (e.g. biker) produced large gains across the board.
1
2
The main result is that this actually works really well for one step interventions! The groups that it finds as challenging (illustrative table below) correlate well with those that are actually difficult for the model. There is a host of analyses in the paper and a ...
1
2


























