Regarding the first quote: in any case, I think it's probably a good idea to encode this implicit knowledge somewhere. A bit like code is the best kind of specification.
Code can be studied and improved gradually. Since it guarantees repeatable results, you can try tweaking parameters and measuring the outcome, to achieve even better result. It's a bit like taking the highest-skilled workers, merging their knowledge, and extending their lifespan indefinitely so that they can continuously improve.
The only disadvantage I can see is that you're skewing the playing field in favor of those who have access to the most data, and by reducing the workforce size, make it difficult for someone else to obtain the same data.
At the end, a skilled worker can leave a company and make its own, or join a rival one. An algorithm can't do that, further entrenching a select few companies'positions, thus locking knowledge away from humankind. Companies probably contribute to the loss of information over time (see lost knowledge: Roman cement, massive bronze castings, etc), as free-flowing information is against their interests, in most cases.
> Companies probably contribute to the loss of information over time
I agree with what else you said, but I see much of the opposite process in corporations actually. I have worked in a couple of very large corporations and what I have seen is less of information hiding (and "not invented here") and more of cooperation through partnerships, open source software and standardisation of processes.
Of course, it isn't all bad or good, and I was just writing down my train of thought. Corporations also contribute to maintaining that information: by grouping together individuals skilled in the same domain, they can foster the exchange of ideas, and by training new employees, they perpetuate those skills. Patents are another mechanism that forces them to contribute back to society.
Once you start training algorithms instead of people, though, you attach that knowledge to the company, and it may not leave. In my experience, open source software and standards are the exception rather than the norm.
Fair point, I didn't think that deeply about trade secrets. I guess it could be even more hidden with various forms of machine learning, even the engineers themselves possibly couldn't explain how or why it worked when they left the building.
While I mostly agree with your comment, I would like to point out that a company may have a lifespan much greater than that of a skilled worksman (which may or may not have transferred his "specific" knowledge to any co-worker and/or apprentice). A company, simply because of the stream of new people that will have to be trained, need to have that knowledge passed down at some point, and having it laid down in writing or any non perishable medium is the safest way to do so.
I'd like to add a little context around the quote [0], because I do think it makes another issue very apparent:
> Building a dam requires knowledge and skill developed through years of experience. Obayashi's automated system is expected to be a game-changer in dam construction, as well as in other applications.
> "By transferring expert techniques to machines, we're able to analyze what was once implicit knowledge," said Akira Naito, head of Obayashi's dam technology unit.
> Every process for constructing the 334-meter-wide dam will involve some form of automation. That includes the initial work of establishing the foundation, and pouring concrete to form the body.
Implicit knowledge here refers to dam builders' workmanship and experience. It is an empirically constructed knowledge which is "stored" in the worksman's mind as instincts, concepts, know-how... While these may "be analyzed" and re-used for automation purposes, I have very strong doubts whether such implicit knowledge may even be truly understood and translated (be it in code or any other media) if there is such a precise goal. Specifically because the knowledge is implicit and applies situationally, one may never be able to grasp it all without some kind of very complex knowledge transfer set-up that could cover a broad range of situations, allowing all of the patterns to emerge and be identified.
Enough knowledge may be gathered that a company could successfully apply it to dam building by machines. But as automation replaces dam builders, chances are that the portion of the extracted knowledge which is not necessary to the company will join the expertise which did not transfer from the workmans' minds to the company to lay forever in the land of lost knowledge (...until it is found again through experience).
The rest will most likely be preserved "forever" through more or less obscure patents and "hard", explicit knowledge in the company's hands.
Code can be studied and improved gradually. Since it guarantees repeatable results, you can try tweaking parameters and measuring the outcome, to achieve even better result. It's a bit like taking the highest-skilled workers, merging their knowledge, and extending their lifespan indefinitely so that they can continuously improve.
The only disadvantage I can see is that you're skewing the playing field in favor of those who have access to the most data, and by reducing the workforce size, make it difficult for someone else to obtain the same data.
At the end, a skilled worker can leave a company and make its own, or join a rival one. An algorithm can't do that, further entrenching a select few companies'positions, thus locking knowledge away from humankind. Companies probably contribute to the loss of information over time (see lost knowledge: Roman cement, massive bronze castings, etc), as free-flowing information is against their interests, in most cases.