chimtim

Uber currently doesn’t maintain a large (or any) fleet either. I wonder how much the marginal gains from not paying drivers is shaved off by fleet maintenance of expensive cars.

I've been working in the SW industry for around two decades now, and I see a distinct difference since the last 5 years, which is that everyone is optimizing all their actions towards their next promotion.

Prior to this time frame and more so in the last decade, there was a lot of focus on mastery of SW skills and even towards random exploration and learning, and all of that appears to have been lost. I personally felt all this has accelerated with the increase in big tech compensation, and levels.fyi and blind information.

The challenge with a greedy approach is that a career spans over multiple decades, and a lot of careers plateau over local minima and often one may have to down level or restart careers to find their true calling and excellence and the career growth that follows. This is better done at the beginning of the career when risks are lower than at a late career stage.

Unfortunately, one in 10 times is far from good enough (and this is with good prompt engineering which after using large language models for a while, one starts to do).

I feel like the current generation of AI is bringing us close enough to something that works once in a while but requires constant human expertise ~50% of the time. The self-driving industry is in a similar situation of despair where millions have been spent in labelling and training but something fundamental is amiss in the ML models.

caching a computation result is called memoization.

For anyone not from Scandinavian countries or not used to the low sunlight and mostly cold weather, the downsides from the long-term side-effects of living in Sweden such as depression, low vitamin-D state significantly outweighs any short-term happiness.

I am glad I cancelled my subscription early this year and never bought its stocks.

This may appear to "work" but I doubt it is better than a 300K FPS camera by any means. It will not work perfectly across all scenes, it cannot interpolate uncommon movements and finally, it cannot interpolate many low-res cases like 5-20 FPS where there is actually significant missing information across frames.

Have an expectation that the work-life balance is a circle i.e. it is zero

I went through all of them and I did not see any regret minimization background. If you think there is, please point me to it and clear my misunderstanding.

is it just me or these are really poorly designed courses? I took the RL course and it first started describing papers in RL with little context (under applications of RL). The course then jumped to bandits and I was hoping it will describe the foundations but it assumed students know about the regret minimization framework. Researchers are traditionally not the best educators -- I usually pick tutorials in conferences that are done by professors as opposed to researchers despite their content.

I actually have worked with lots of customers in deploying ML. This was my perspective. Thanks for sharing your perspective at Google.

Exactly, most non-IT companies have very limited skills in this area. This is why they outsource. As an analogy, when I want to furnish my office, I look for an interior decorator who takes care of ordering furniture etc., within a budget. I don't screw around with a 3D printing API to make chairs and tables. It is too low level and not my core expertise.

And second, your comment that most folks achieve any accuracy is strange. These are not real businesses, they are mostly developers and hobbyists trying to learn. These folks sign up for kaggle and poke a few scripts and view half a class on coursera on ML. They are not real businesses and they have no money. Most of the real businesses are hiring startups or large companies that hire data scientists with domain expertise like in oil, manufacturing etc. (the IBM model). ML as an API is a disaster as a business model.

Also, AutoML is not even close to being better than humans even for a specific problem (across datasets). These click-bait titles don't fly outside of AI conferences.

This does not sound very useful. The entire market and AI consulting is for that last 5% of accuracy. Is there a reason why Google is investing so heavily in this space? Perhaps a simple model/UI is a good starting point for many customers, and many of them end up renting the classic cloud after they gain some expertise. And the rationale for all this PR is that Google does not want to miss these customers to Amazon

Then what is "Auto ML"? It sounds like just another cloud service.

So much buzz/hype with multiple articles even though the early results are only on MNIST.

these are not random, raw youtube video datasets. they are hand curated dataset in specific classes (like using mechanical turk). youtube has really diverse videos with different lighting, and real world scenarios which makes it an excellent dataset. Movie clips look great but models trained on them are useless in real world.

this is what makes this dataset poor. other datasets mentioned in the blog are based off youtube which is more realistic. movie based datasets have perfect lighting, center the subject are almost never useful (e.g. HMDB)

1) managing an office building is a lot of hassle.

2) startups are always growing or shrinking. no point putting so much significant capital unless it is an established in for a long haul at a location where talent is surplus.

3) if you indeed have that much capital, you can pretty much dictate what you want and get some tax benefits (amazon, second HQ) or work with a city build a new building suitable for your long-term needs (apple).

4) not much inventory that meets your needs in SF market because the property prices are always going up.

5) the leasing company actually has limited leverage. finding another tenant esp. the size of dropbox is not easy. you can bin pack other tenants but you lose lot of money until all new tenants move in.

All this raw compute/parallelism is great but it does not really help the algorithm output in terms of efficiency. The CPU/GPU difference is much smaller than nvidia would like you to believe especially for more complex networks which are becoming more and more common now. Of course if you want to just do convolutions (which the TPU paper claims is only 5% of their google workload), building a hardware around it may work well for specific algorithms.

Intel needs to double or quadruple their processor's AVX width, add 8 bit instructions and make it slightly more power efficient (for AVX) and it will easily beat many of these TPUs. Of course this is easier said than done but I am really surprised they haven't done it so far and instead bought Nervana and other hardware chip startups.

re-write it in C++

The demand is quite low. I think if you want a good low-level code implementation, you can easily find one on github. For example, you can easily find a good compression algorithm for your CPU budget or other application needs on github. So you can get away with mostly writing high level code and using these external libraries. Second, the demand for writing really optimized code is not as high as 10-20 years ago. Now, writing reasonably fast, readable code is OK and no one expects you to run gcc -S and count cycles unless you are working in some high frequency trading firm.

Low level programmer here. This is great but the job market for low level programmers is very small. Recently, while interviewing, I described a thread safe, concurrent queue implementation to a hiring manager and he asked me -- "so how does this relate to big data and ML"?

The linked press release shows an animal is detected -- tiger etc. It does not say tiger running or hunting, which is where the time component would have been used.

what is the "video" bit here? This is just running image recognition on a bunch of frames.

Do you plan to meet every week?

or buy an umbrella

In terms of number of neurons, human brain is around 10^11. Current image classification deep networks are around 10^7. However, its not just the number of neurons, the connections per neurons also matter. The connections per neuron in a human are in thousands while in modern networks they are in hundreds.

The human brain is also very efficient so it also makes a very good use of those neurons, and modern networks aren't. However, a human brain trains over a very very long time and also prunes itself constantly, while such efficiency techniques at scale for very large networks are being explored in deep learning only recently.

PCM -- coming soon next year since last 7 years.

It is a dataflow language and deals with tensors (multi dim arrays), so tensor + dataflow = tensorflow.