This guy's metaphor completely and totally breaks down when he begins comparing today's space agencies to China and Portugal of the 15th century.
Apollo and the shuttle weren't caravels and galleons. To extend the aquatic metaphor here, Apollo was our most sophisticated floating log at the time, and the Shuttle was our first real attempt at tying some logs together so we could ferry supplies back and forth to a little sandbar (the ISS) we created 15 feet off shore. If we ever try for Mars, we'll probably be sending our first dug-out canoe.
China and Portugal knew how to create decent and practical sailing ships. We barely know how to get into space, much less have anywhere near the technology available to exploit it properly.
And most importantly: THERE IS NO ONE AND NOTHING OUT THERE TO TRADE WITH. No developed resources and easily-available goods we can just pluck out of the metaphoric ground and take back. Everything up there has to go through a very complicated and expensive process of development before it becomes useful and profitable.
The better metaphor for the current state of space development may be if Portugal was the only human settlement in all the world, and its only watercraft were floating logs. The China vs Portugal argument might work in two hundred or more years, when we actually have the space-going equivalent of caravels and galleons, but for right now just finding a decent oar we can paddle with is quite an accomplishment.
If Windows is the only platform (apart from the Xbox) where the dongle works, there's also not much of an option if you want to intercept valid traffic relatively easily.
It's a really honest description of the attempts to move to neural networks.
At least the way it's written it feels like no 'data scientists' were involved, it was all done by data engineers (software developer rather than statistical/modelling knowledge)... Which is depressing, if even Airbnb are biased to hiring only good developers (rather than a mix)
Most companies don't need to hire data scientists.
They need to hire engineers who can take an algorithm implemented well by somebody else, and apply it to a business domain.
That's the future of deep learning, machine learning, and all other linear/logistic regression style technologies.
The mathematicians are going to have to wait for the age of quantum annealing to feel valuable again: reducing code into a function that works on a quantum setup actually needs those skills that developers struggle with.
Everything else though, outside of pure research and in the vast majority of companies, is already well on the road to commoditisation.
This is pure comedy. Want to waste a bunch of cloud resources operationalizing a garbage model? Have engineers do it.
The most successful workflow strategy I’ve seen in practice is where the deep learning researcher is also the person operationalizing the model. The same person who is grokking the latest paper in arxiv is also studying correlations in product data to perform feature engineering and also writing Dockerfiles to make the work reproducible and optimizing containers for production deployment, latency, failure tolerance, and evaluating performance in the specific context of the business application and creating well crafted software components with adequate testing along the way.
The commodity part is the cloud engineering, kubernetes pod setup, load testing tools, and general software engineering. Machine learning engineers are typically great at these things and they are easy to learn.
Meanwhile, learning about the nuance of hyperparameter tuning, how to evaluate overfitting, model complexity tradeoffs, when to use which kind of statistical modeling tool, how to improve models based on observing error cases, and a host of other statistical modeling concerns are wildly not commoditizable at this point of history. Knowing how to copy paste some Keras tutorials will not help you.
80% of real business problems are going to be solved by figuring out how to get XGBoost working with it. And you're done.
Research is valuable, and there are some problems where doing real thinking is useful, but pareto principle is at play here: a lot of people just aren't going to need to do that, for the same reason most developers don't need to know the difference between a merge sort and a quick sort: sorting was commoditised into most programming languages decades ago. Same deal, different tech with machine learning.
This is not a bad thing, and there will be a bumpy road, and there will always be a market for experts to help with the edge cases, but most firms will drop them like hot bricks within a decade.
"They need to hire engineers who can take an algorithm implemented well by somebody else, and apply it to a business domain."
I dont think so. While Engineers are needed to operationalize the models. For Statistical Learning part, understanding of data preprocessing is an important step which requires knowledge in statistics.
I've got bad news for you, but the current state of rabid hiring means that this was very likely written by people that airbnb calls "data scientists" or "machine learning engineers", I'm sure half of these people have a phd in some arbitrarily "quantitative" field.
The state of all large companies that I've seen that are not FAANG is that they are rushing to build teams of "data scientists" that slap together keras models and "ship" them, meaning the outputs are stuffed in a db only to be consumed by other keras models.
My favorite gem from the original paper, which shows the sad, sad state of deep learning in industry is this line:
>Out of sheer habit, we started our first model by
initializing all weights and embeddings to zero, only to discover that is the worst way to start training a neural network.
I can't imagine anyone who has even a mild understanding of how neural networks are implemented and trained making this mistake.
Yes, the scrolls have been published in their entirety in the series Discoveries in the Judaean Desert numbering some 40 volumes. These are the official "first editions". There are numerous other collections of a non-critical nature as well as translations. A simple search should yield the relevant results.
Unlikely. There are many orders of magnitude more ways to be well adapted to one environment than to be adapted to two or more environments. So if an organism is well adapted to high radiation it’s probably terrible at many, many other things that are irrelevant where it is now. There are plenty of extremophiles apart from these radiophiles. You don’t see microorganisms adapted to very high levels of acid, pressure, heat, cold or anything else suddenly exploding into new ecological niches regularly.
Consider the analogy to multiply resistant to antibiotics bacteria. They’re far more common in hospitals than elsewhere because hospitals are swimming in antibiotics. Once they leave the hospital environment they’re paying a cost to be antibiotic resistant with far lower benefits so they either lose the capacity or are outcompeted by other organisms in the same niche.
The explosion into a new niche does happen occasionally, when a new capacity is extremely useful or when an adaptation is discovered that is far more efficient but has similar costs or better than the alternative. I’m not familiar with any examples in microbiology but invasive species exist and so do we so we have existence proofs in macroorganisms.
In general organisms create bad consequences when they're too adapted to Earth and reproduce excessivly in some way. If they die out because they're adapted to space instead of Earth that won't be a problem for us, though it will be for the organism.
This was my first thought as well. Although that adaptation wouldn't be much of an advantage anymore once the organism got back down here. Or would it...
It's much easier to do reverse engineering on programs compiled with old compilers, nowdays compiler are really good at optimizing shit, which means making the assembly code more complex, using new instructions etc...
Reminds me of Coccinelle - a language for writing semantic patches for C (like eg: add close_foo(fooid) in every function that has open_foo for each fooid etc...) , but I see Rascal can do some more.
I remember someone saying something along the lines of “Facebook will try and copy and crush us, but not this time- by the time they work out what’s happening, it’ll be too late”.
Facebook seem to have noticed what’s happening and are moving like the wind to try and crush us.
Are they too late or are we too late? Or neither?
Apollo and the shuttle weren't caravels and galleons. To extend the aquatic metaphor here, Apollo was our most sophisticated floating log at the time, and the Shuttle was our first real attempt at tying some logs together so we could ferry supplies back and forth to a little sandbar (the ISS) we created 15 feet off shore. If we ever try for Mars, we'll probably be sending our first dug-out canoe.
China and Portugal knew how to create decent and practical sailing ships. We barely know how to get into space, much less have anywhere near the technology available to exploit it properly.
And most importantly: THERE IS NO ONE AND NOTHING OUT THERE TO TRADE WITH. No developed resources and easily-available goods we can just pluck out of the metaphoric ground and take back. Everything up there has to go through a very complicated and expensive process of development before it becomes useful and profitable.
The better metaphor for the current state of space development may be if Portugal was the only human settlement in all the world, and its only watercraft were floating logs. The China vs Portugal argument might work in two hundred or more years, when we actually have the space-going equivalent of caravels and galleons, but for right now just finding a decent oar we can paddle with is quite an accomplishment.