About that $1b... IBM says Watson Health assets fetched $230m in pre-tax profits IBM has confirmed it made a few hundred million on the $1 billion sale of the healthcare data and analytics assets housed by its Watson Health unit. When the tech was flogged off to private equity firm Francisco Partners in January, it was said to be valued at a cool $1 billion. In an SEC 10-Q filing on Monday, Big Blue …
Reading that statement felt like playing tic-tac-toe with linguistics, so, "Yes"?
Although, re-reading the statement, I can see why the word salad was used - He was admitting fault. Rough translation: "We got into something we don't know jack about. This normally doesn't stop us, we can have consultants bail us out; but in medical, it seems even our consultants can't bullshit our way out of 'practicing medicine without a license' charges."
. Does Watson generate their executive statements? ------ No. Watson is capable of creating coherent sentences.
. Are words selected entirely at random ? ----- No way to tell. Would need a larger sample to test for randomness.
. Could we use Watson to read the statements? ----- Probably. Unless Watson refused to be associated with the effort.
. Would that be a new vertical, horizontal, or diagonal? ----- None of the above? An entanglement perhaps.
I cannot help but wonder what the inability of IBM to make Watson Health profitable or to offload it to an actual healthcare entity represents. Are we seeing a concrete example that AI/ML cannot produce the results the hype machine promised? Can we expect to see more announcements of the shuttering of other various startups and corporate AI efforts in the near future?
> Are we seeing a concrete example that AI/ML cannot produce the results the hype machine promised
I suspect that it's the same issue as per self-driving cars, in that it's a mix of legal and technical concerns.
Or to put it another way: the legal costs involved with any technical solution which isn't 99.999% perfect are simply going to be too high!
There are many promising approaches in computational oncology. Working in that field, all I can say is that Watson was never one of them and was viewed by many as an over-hyped white elephant. Health AI is chock full of conpeople in MBA jackets these days.
I'm skeptical of the whole AI/ML thing. It seems to be a buzzword which over promises and under delivers. Take translation, which has long been subject to so-called AI. Google's Translate service has the buzzwords: "Translation AI", "AutoML technology", and "industry-leading accuracy".
What do we get if we translate "hit me with your rhythm stick" to French? "Frappe-moi avec ta baguette rythmique". Why baguette can indeed be used to mean stick, baton would be far better here (IMHO). We can actually see this if we translate just "rhythm stick". We get "bâton de rythme".
Taking this one step further to see if I am wrong about batons or baguettes, a visit to amazon.fr searching for "baguette rythmique" versus "bâton de rythme", we see far more correct results for the latter, with occasional correct results for the former (amongst a lot of wrong results).
So in just once sentence we can see that AI is not quite as good as a low grade human whose French is very poor. These sub-par translations occur despite Google having enormous and innumerable corpora for source material.
Write the letters AI onto a product and make it topical. It's improving every day.
Funnily enough I've been saying this from the inside (hence anonymous) for about two years now. Current AI/ML approaches seem to be really good at solving about 90% of a problem by sheer "brute force" i.e. build a big enough computer and a big enough training set.
Unfortunately the real world is very messy, so that last 10% is as hard as the first 90%, so you have to double the resources to get to 99% effective. And double again to get to 99.9%... extrapolation left as an exercise for the reader.
Fundamentally, ML-derived AIs seem to be very bad at dealing with the unfamiliar, and I suspect that we are hitting a period of very steeply diminishing returns, and we may even have another AI Winter until and unless the industry recognizes that ML needs to be augmented with knowledge engines that can reason generally about their world even when specific circumstances are unfamiliar so that instead of getting a self-driving car that can be fooled just by changing a 30MPH sign to 80MPH, you get one that says to itself "the sign says 80MPH but everything else about this street says residential area, so I'm not going to believe it".
This post has been deleted by its author
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
