If Machine Learning is the question, open source is the answer. Right? Machine learning (ML) and artificial intelligence (AI) are extraordinarily hard to pull off in the real world, so of course the solution must be open source. From Google’s TensorFlow to Microsoft’s Cognitive Toolkit, the world is awash in open source ML/AI code... none of which seems to be solving the gaping void between AI hype …
It was part of my M.Sc. at Kingston Polytechnic in 1988. Even then it wasn't new. By the 1990s there were simple systems to apply ID3 to your data. One of my students built a system to diagnose congenital dislocation of the hip in two days.
It is doing it at a large scale, and with more neural network layers, that is new.
Machine Learning most certainly is nothing new, the phrase originally coined even earlier by an engineer at IBM. The problem I have with it is the hype surrounding it's links to AI - despite recent large scale developments and neural networks it's still nothing more than clever algorithms and code just doing what it's been programmed to do. Yes, neural network systems do this slightly differently but they are still essentially just a large collection of mathematical equations! True deep AI is still a huge unknown - the holy grail of creating something that can think for itself (i.e. sentience) is still a complete pipe dream. I'm not even sure the day man can play god will arrive - you are, after all, using the brain the recreate itself without a biological reproductive process. The brain is very good at inventing tools to ease repetitive work but I'm not sure it's capable of analyzing itself. We're better off looking for the self-diagnostics gene and it's associated activation button!
So lots of the clueless rediscovering stuff they could have been taught in a decent CS degree and implementing it in the clumsiest way imaginable on AWS or Google.
Which suites Amazon and Google just fine.
Competence is not required.
A bigger use of machine resources is a benefit from their PoV.
As usual Mr Asay is off the mark. If a company really needs ML then it's very likely to be considering running the ML on Google's infrastructure because scaling ML gets very expensive very quickly and you don't do it all the time. Once you've got the training done then you can run combined setups.
Also, try and reduce the confusion by not continually mentioning ML and AI as if they're interchangeable. ML is pattern matching and can only be a part of an AI system.
Don't like replying to myself but I realise I failed to reply to Asay's jibe about open source and innovation. It can certainly be argued that open source can discourage feature innovation because it effectively precludes commercial exploitation by being free. If your only business is selling software you develop then going open source will kill your business.
But, as has been pointed out many times, selling software is not the only way to make money. I see an increasing number of companies happy to use open source stuff and pay for support, improvements, etc. This isn't the case for everything but you see it with some of the larger projects such as Postgres. Google is not planing to sell TensorFlow, it plans to sell computing time to run TensorFlow. Going open source reduces the barrier to entry by making it free to use locally so that people can get familiar with it.
Then there are the many open source projects that have driven innovation in the industry. We probably wouldn't have the internet we do without BSD unix and that was development without a market. Sams goes of course for the programming languages we all use. The real value of open source is the possibility of peer review, which is increasingly important for security reasons.
Now there's a Gartner 'analyst' who will say whatever you want him to say for a price.
He was IBM's go to guy for their Big Data press. I caught one of his marketing tour keynotes...
Got a lot of things wrong but kept to the script touting IBM....
Posted Anon for the obvious reason...
Which seems to be a bit of a benchmark for all other areas of software development.
AFAIK historically in AI the interface has only been used to populate fairly small internal databases.
No ones seems to have tried to use the tool (for example "The Programmers Apprentice" project at Stamford) to maintain itself. IE use the algorithms you've supposedly coded into the system to process another copy of itself into a better version.
And of course if you could do it once, what would happen if you had V 2.0 process a copy of itself?
Drop the AI and you'll see projects are using ML in static code analysis to pick up typos and errors and some of them will even fix things for you.
An awful lot of coding could be automated and chances are that it increasingly will be.
From my understanding and nonzero experience, every machine learning solution us domain/applications cation specific so far. That is the very lim Ted state of the art. Yes you could likely build a system that could gradually improve itself. But that is what it would be good for.
Long ago I argued that a good compiler should gradually learn to be a better compiler. AFAIK that has not happened yet. But all of these possibilities domlay before us.
I've told several people that the next generation of computer "programmers" will be more like teachers, helping baby AI to learn how to solve the problem(s). This is radically different from classic imperative or functional programming but still requires the special ability to understand machine processing from the ground up.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
