Taste graph infographic

Cool infographic about our core technology at Hunch, what we call the “taste graph.” My favorite stat is that we have >25,000 API clients making >400,000 calls per day (many clients seem to be devs building not-yet-release apps). Upcoming infographic will focus just on API usage and growth. And yes, crazy as it seems, we have Assembly and C code, which was necessary to optimize core inner loops which are extremely computationally intensive.

(click for full size)

http://technbiz.blogspot.com paramendra

“…..
>25,000 API clients making >400,000 calls per day ….” Wow, man, looks like Hunch is about to explode.

http://www.cdixon.org chris dixon

Thanks! Things have been going well. We have a lot of improvements coming soon to API and also website and a new mobile app that I’m very excited about. It’s a hard tech and has taken us a long time to get it right but I think we are almost there.
http://urgentessays.co.uk/ custom essays

i wanted to say exactly the same!!!!!!!!! %))))))))))))))))))

http://technbiz.blogspot.com paramendra

Hunch Is About To Explodehttp://goo.gl/fb/70jnf
http://twitter.com/andyidsinga andyidsinga

fun.
have to admit though, i got lost on the juxtaposition of the little milky way bubble and the big hunch bubble

http://www.cdixon.org chris dixon

Yeah, a little infographic creative license exercised there

http://www.victusspiritus.com/ Mark Essel

“In a survival situation, do you think you’d be able to eat human flesh?”

only at Hunch

Great infographic, piped it down my tumblr tube.

Forgot you guys leverage the one uber computer approach with GODZILLA memory to identify relationships between people and taste metrics. Assembly was mentioned because standard linear algebra libraries have issues with massive matrices, it’s far out that your team ran into that limitation.

After reading about those huge matrices I wondered which are really the best THAY’s. I did a study a few years ago on a hyperspectral collection and determined the 5-6 dominant bands out of thousands yielded 95+% of detection performance. Exhaustive combinations is one way to do it, dominant eigenvalues is another. Is the challenge classification for Hunch?

Also, I had a great conversation with Charlie Crystle who visited Manhattan last Friday. One of the ideas I had for Hunch was as an optimal core power user audience for startups. I come to you with a startup that focuses on say a handful of features/tastes, and you generate the 100 or 1000 people who would absolutely love the idea.
http://lmframework.com David Semeria

I can understand C, but assembler, seriously? Even avionics systems can get by with compiled binaries….
Wouldn’t like to be responsible for maintaining that code….

http://www.cdixon.org chris dixon

It’s not a lot of assembly. Inner math loops.

http://filmicgames.com John Hable

Nice! It’s cool to see some good old-fashioned optimization of large sparse matrices. Have you considered going the CUDA route, or are your matrices too large to fit?

As a side note, it seems like college kids are being trained to be almost anti-optimization. On stackoverflow.com and similar sites you see people ask “How would I optimize this?” and someone always answers “Why would you want to do that? The compiler is smarter than you!”
Very annoying.

Anonymous

I’ve tried CUDA’s cublas libraries for multiplication but unfortunately the inner dimensions are too small to get a speedup. It ends up taking more time to chunk and transfer the data over the PCI express bus to the GPU and back than it is to perform the calculations on the CPU. We are looking forward to Intel’s Sandy Bridge architecture though, which puts the GPU on the same die as the CPU.

I suppose kids these days are mostly right about letting the compiler do it’s job, but particularly with matrix operations, the most efficient way to unroll your loops and order your operations depends on the dimensions of the matrices, the size of your cache lines, the number of floating point registers you have and which streaming instruction sets you can use. Compilers aren’t quite omnipotent enough to take all of that into account yet and choose the most efficient route. Actually I don’t even think CPU’s report the sizes of all the cache lines to the BIOS/OS.

http://filmicgames.com John Hable

Yeah, the main misconception that most people have is vectorization. Most people seem to think that your compiler can autovectorize your code for you, but in reality it can’t. Even when you are compiling an OpnenCL kernel, which is far more constrained, the Intel and AMD kernels aren’t usually vectorized. For the PS3, we would usually see about an 8x improvement by vectorizing our code and then another 2x improvement when you software pipeline the assembly. Makes a big difference Yes, you typically get an 8x improvement by converting to 4x-wide instructions.

Sandy Bridge is awesome, although I wouldn’t bank on the chip’s GPU to do much. It’s about as powerful as a very cheap discrete card, and it looks like you won’t be able to code with OpenCL on it. Although getting AVX instructions is awesome for me at least…8x wide vector operations and several very useful instructions (Select, Shuffle, and Half/Float conversions in particular).

Anonymous

Yeah AVX will be a big win for us as well. There is also the possibility of big gains from in-socket FPGA’s if the GPU route doesn’t pan out.

http://profiles.google.com/zynest 翊张

Dear Sir,I am a social enterprise Entrepreneurs (E-commerce) come from China, Current, In addition to introduction of several partners on the Website, There are a number of Beijing College Students willing to join. I hope you can pay more attention – http://www.aladdinsring.com
Sincerely, Zhangyi
Pingback: Daily Digest for May 17th – 4% Inspiration
http://researchpaperwriter.net/ research paper writers

very interesting) not even only for children… pictures can be important. this one carries los of information and can tell itbeter than using words!
BuyGiftsItems

I wouldn’t bank on the chip’s GPU to do much. It’s about as powerful
as a very cheap discrete card, and it looks like you won’t be able to
code with OpenCL on it.
Buy Gifts Items
Pingback: Chris Dixon: Hunch, Taste Graphs & the Link Between Lettuce & Politics
Pingback: Klean Kanteen | BICYCLE
Pingback: Life After | AUDIO BOOK
Pingback: Ebay adquiere Hunch para lograr un mejor servicio de recomendaciones | Seo Peru
Pingback: Ebay adquiere Hunch para lograr un mejor servicio de recomendaciones - La Isla Buscada
Pingback: Ebay adquiere Hunch para lograr un mejor servicio de recomendaciones | EmaCorp News
Pingback: Ebay adquiere Hunch para lograr un mejor servicio de recomendaciones | ideasweb.info | Noticias, Software y novedades. Las mejores aplicaciones web, con los trucos más útiles y toda la información en nuestro blog.

Chris Dixon