the chrome browser spends a decent amount of time on other steps such as parsing HTML. I wonder how much time could be saved by not rendering pages into pixels.
The gap between sensationalism and pragmatism is measurement. The graph shows an accelerated growth of CO2 emission, but lacks exact profiling data to deduce reasonable actions.
I agree with most, but referrals are one of the most important indicators. "A hire As, B hire Cs". A known issue, however, is the workforce diverse given the heavy use of referral methods.
This might not be true given just a couple of hours staying in the office and talking to people. Management issues / technical debt are not visible until you work for the company.
It really depends the type of growth. Organic growth is driven by the quality of product therefore scaling issue comes later. But most venture-backed startup services such as O2O need quickly dominate market by throwing cash to get users so scaling is an issue from day one.
I don't view irratiality as a positive quality for entrepreneurs. From outsiders, they may seem crazy and stubborn. This is because those outsiders cannot understand the underlying assumptions or don't envision the world in the way the entrepreneur does. The entrepreneur should in fact always be rational, and make sure every decision is made based on solid ground.
The high dimensionality is merely for the convenience of optimization, so that we get smoother manifold on which a good approximation solution can be found. In that regard, it is interesting to think about of space of models that we humans can interpret, and the possibility of "distill" deep neural nets into models in that space.
I think one difference is that for many languages like Python, the http endpoints are hosted in one process. That's less horizontally scalable than lambda services.
I am very surprised that it takes Apple years before they finally decide to release Siri API (supposedly). The benefit of speech assistant for apps seems huge. And would it been released earlier, Apple can collect much more speech data as well, which may help build next-gen Siri.
One way is to think about BI as insurance. Then the emergency cases are averaged across the entire population. Joe's twelfth car break-down can be classified as "insurance fraud".
They claimed to be "the only public toolkit that can scale beyond single machine". Well, mxnet (https://github.com/dmlc/mxnet) can scale across multiple CPUs and GPUs.
In a world so abundant of information, i think user attention is indeed a significant resource every major player should fight for. I guess the distribution of attention follows power law, that a few entrances take up most of the mobile use cases. For example, I use Uber, Wechat, GMaps much more often than the other apps. Search bar definitely is one of the most critical entrances.
it is true that most recent success of deep neural network are in the regime where n, d are large. And we surely shouldn't fantasize general AI solved in this way. However, the very appealing aspect of deep neural network is end-to-end training: for image recognition, we can map from raw-pixels to output. This is very different from other ML techniques. In some sense, deep neural networks learn "algorithms", not just "models". This formulation can be richer especially when given lots of data.
it seems like a general phenomenon that our scientific understanding decreasing with the complexity of the system. in classic physics, we could have very clean equations capturing the dynamics of a system. this is not true any more in chemistry/biology/social science. modern deep learning pipeline, compared to its shallow counterparts, gains more complexity, so it's not surprising at all to me that we are not able to understand it well at the moment.
I think automated theorem proving replaces the "search function" of a mathematician. But there is another part, which seems to be more important, is the creative insights. For example, inferring and conjecturing a theorem from few "data points". Machines at the moment do not have such capability yet.
I guess the point is flow of talent from outside Europe. For example, there are significantly more good engineers from Indian, Europe and East Asia moving to Silicon Valley.
Julia certainly does have several features worth being excited about: calling C without wrappers, macros, good JIT performance. But I think what really matters is the ecosystem. With python, there is tons of libraries such as Numpy and Scipy for scientific computing, django for web backend, etc. It is unclear whether the community will embrace Julia fully within the next few years.
I think there are two major reasons why CNN might not be a good model. The first is that CNNs expect translation invariance, which is pretty common in images. But language sentences do not have this structure. Another reason is that in NLP outputs usually have varying length, this is why RNNs, LSTMs are so popular these days.