papercruncher

What has worked best for us is to narrow the category as much as possible before attempting to do unsupervised clustering.

We focus solely on sports and our classifiers(supervised) reduce the scope first to the sport and then to the specific team before we apply any sort of clustering (k-means, LDA, etc). That allows us to reduce the vocabulary to what is mostly a list of named entities for the sport/team and key words such as 'injury', 'quarterback', etc. With a significantly reduced vocabulary, even algorithms such as Hierarchical LDA work surprisingly well.

Curious: Were real-time updates something that your customers had high demand for?

I'm asking because we thought about doing real-time updates for our product but decided against it because it made topic modelling (clustering based on topics) much harder for us.

Also, what is your experience with publishers being consistent about pinging the hub with updates? Do you still need to poll feeds periodically to make sure you didn't miss something?

How about a service that you could enter all these things in the morning and it would send them throughout the day? Like a Buffer for "unit of thinking of you".

I'm only half joking

Showers: Get a gym membership. Has the extra benefit of the occasional workout to stay healthy and sane.

Along the same lines, my favorite question is "What is your favorite data structure?". Most of the time, the answer is HashTable followed by a complete lack of understanding as to how a hash table works. Occasionally someone would respond with 'Quick Sort'

Every now and then you get a candidate whose eyes lit up and you can't get them to shut up (in a good way). And that is the story about how I discovered judy arrays.

Building a model and personalizing a news stream on a per-user basis is entirely feasible today. Even with very simple ML, you can see some very good real world results.

To avoid the problem of not being able to surface new topics because the user model has converged somewhere, you can use a bandit approach(seriously, who comes up with these names?). By that we mean that you do a bit of exploration by showing news topics drawn at random from some distribution and then adapt your model to what you learned. The naive implementation is "Every random(10,20) news items suggested by the recommender, display 1 drawn at random from the entire corpus. Adapt_Model(user_clicked)"

More info: http://en.wikipedia.org/wiki/Multi-armed_bandit

The timex module in NLTK contrib does a fairly good job at that but uses regular expressions

https://github.com/nltk/nltk_contrib/blob/master/nltk_contri...

That is amazing work, I really like the clean lines between the tiles and the centered 'Get Started' button. The top bar is great too and the green/purple colors really help the words jump out of the page.

The "movement" effect on the tiles however is IMHO not a great idea. Unless I stare at the tiles long enough, I don't realize there is a fade-in/out effect, the pictures that are faded out just look bad.

I had no idea this was a tech start-up! Tried their d/t SF location twice. Both times they managed to get the order completely wrong even though I was the only customer there. The sandwich was decent though

It doesn't seem to provide any Bayesian inference methods

How are you dealing with bit rot? Are you periodically scrubbing the data to give the controller a chance to repair or are you waiting to get a URE during an array rebuild? Are you running end to end checksums against all your data to protect against bad firmware, bad ram, etc. What is your mean time to repair in case you lose a drive?

One more question: you saturated the network link with a sequential read/write, but is that how you actually store the data? If not, how long would it take you to be up and running on another CDN in case Akamai goes down in flames?

I very much like Unity. The eye-candy makes me happy and from a productivity perspective it allows me to keep my hands on the keyboard and off the mouse. It does indeed use way more resources than it should, but what else am I going to do with 4 cores?

"Coffee Klatch"

I'm struggling to find a coffee shop in San Francisco that has unlimited wifi and stays open late, where late >= midnight. Suggestions?

Cloud implies running in virtualized environment. Dedicated implies that only your bits run on that hardware which is huge for I/O.

Because there is a large number of illegal immigrants from the southern border already here. That and a few other reasons make it a much harder problem to solve than the problem of skilled immigration.

A lot of immigration proceedings depends solely on the mood of the officer.

Consider yourself lucky, I can assure you that telecommuting is a violation of the VWP terms (you are producing work while staying in the US)

For Windows you can use the DPAPI, take a look at http://msdn.microsoft.com/en-us/library/ms995355.aspx. Of course if the intruder has your credentials then he can create a process to steal the encrypted passwords, but at least he can't just read the file by mounting the FS somewhere else

Yes, the NYC subway in the summer is lovely. Extra points if the train air-conditioning is broken.