> Mostly because 3P support has not been a engineering priority.
Got it: assuming you're at Google, in eng. parlance, it's okay if it's not Prioritized™ but then product/marketing/whoever shouldn't be publishing posts around the premise it's running 60 fps multimodal experiences on device.
They're very, very, lucky that ratio of people vaguely interested in this, to people follow through on using it, is high, so comments like mine end up at -1.
Man this is a funny situation. Ty for sharing, more or less confirms my understanding. Couldn't quite believe it when I was in Google, or out of Google. This should be a big scandal afaict. What is going on???
(n.b. to readers, if you click through, the Google Pixel Tensor API is coming soon. So why in the world has Google been selling Tensor chips in Pixel as some big AI play since...idk, at least 2019?)
Yes, you can use first-party models on the Pixel NPUs or you're stuck with NNAPI which is self-admittedly deprecated by Google and doesn't work all that well.
On third party model workloads, this is what you will get:
Google is clearly not serious on Pixels in practice, and the GPU performance is also behind by quite a lot compared to flagships, which really doesn't help. CPUs are also behind by quite a lot too...
Thanks for the pointer. This is an impressive job - reducing a grammar correcting model to as much as 20MB. Theoretically this could even be shipped to browsers and if we are able to wrap it in an extension that works everywhere, this could seriously compete with Grammarly.
I could understand why Google wouldn't open-source this tech, but the blog pretty much covers how to build one. I'm surprised there isn't any open source project that took this direction to bring a privacy-focused grammar checker.
> Informally, the Gibbs phenomenon reflects the difficulty inherent in approximating a discontinuous function by a finite series of continuous sine and cosine waves. It is important to put emphasis on the word finite because even though every partial sum of the Fourier series overshoots the function it is approximating, the limit of the partial sums does not. The value of x where the maximum overshoot is achieved moves closer and closer to the discontinuity as the number of terms summed increases so, again informally, once the overshoot has passed by a particular x, convergence at that value of x is possible.
> There is no contradiction in the overshoot converging to a non-zero amount, but the limit of the partial sums having no overshoot, because the location of that overshoot moves. We have pointwise convergence, but not uniform convergence. For a piecewise C1 function the Fourier series converges to the function at every point except at the jump discontinuities. At the jump discontinuities themselves the limit will converge to the average of the values of the function on either side of the jump. This is a consequence of the Dirichlet theorem.[11]
Meta nitpick: the Gibbs phenomenon is a statement about any arbitrary finite decomposition. The limit of the wave form as the number of terms goes to infinity is in fact exactly a square wave.
Folks push it because it's Indian and understood by most. Folks resist and protest because it's not their native language and somehow worse than having English everywhere.
> Folks push it because it's Indian and understood by most.
I don't ... dispute that? You're continuing with the non sequiturs and it's really hard to tell what point you're trying to make.
("most" is also "bit more than half the population", which isn't that compelling)
> Folks resist and protest because it's not their native language and somehow worse than having English everywhere.
You have not understood the resistance. People don't think it's "somehow worse than having English everywhere". There is no analogous push for English everywhere that people are not opposing.
I used to believe this. As a masters student who is going to graduate soon, I have to put in months of preparation for the interviews. Nobody believes in your ability to write code even after having close to 100+ commits to a popular open source project. People expect you to come to the correct solution after 5 minutes of them asking the question.
IME the interview you have depends a lot on org size and how close the interviewer is to the money. A technical founder will tend to interview with an eye towards product and marketing and sitting down in a real environment to do sample code. As you go down the line and interview at bigger orgs with more departments, it becomes more specialized, the culture overwhelms the business, and you get more of the "idealized CS graduate" syndrome.
Meh, I never prepared for any interviews and I'm gainfully employed. Granted, I don't apply to companies who are known to give whiteboard challenges I have to prep for.
