AI - Have you ever seen it properly work in the real world?
Discussion
My latest encounter with AI in the real world... Having recently changed jobs and moved from Lotus Notes to Microsoft Outlook, I just had a chasing phone call from a manager asking why I hadn't completed a task that had been emailed out...
What task? I asked. What email?
This is where we discovered that by default, Microsoft decide to bless your working day with something called a "Focused Inbox" and spare you whatever it thinks you don't need to see by putting it in an "Other" inbox.
I'm sure everyone can guess where the helpful AI assistant had decided to file my manager's email, and I discovered a bunch of other relevant messages I knew nothing about in there as well! It's now turned off and I've just got one inbox thankfully, but it did get me thinking about how rarely I've seen AI actually work in the real world. Yes, great for crunching vast amounts of data in a controlled environment, but does it ever actually work as soon as it needs to try and analyse real human needs or responses?
What task? I asked. What email?
This is where we discovered that by default, Microsoft decide to bless your working day with something called a "Focused Inbox" and spare you whatever it thinks you don't need to see by putting it in an "Other" inbox.
I'm sure everyone can guess where the helpful AI assistant had decided to file my manager's email, and I discovered a bunch of other relevant messages I knew nothing about in there as well! It's now turned off and I've just got one inbox thankfully, but it did get me thinking about how rarely I've seen AI actually work in the real world. Yes, great for crunching vast amounts of data in a controlled environment, but does it ever actually work as soon as it needs to try and analyse real human needs or responses?
There's a a lot of AI going where you actually don't realise it. I'm sure you'd notice it if there was absolutely none at all.
A simple one for me is the way that AI is helping to sort some of the mountains of data that we create continuously like say pictures - AI automatically can identify objects and organise pictures for us for example.
A simple one for me is the way that AI is helping to sort some of the mountains of data that we create continuously like say pictures - AI automatically can identify objects and organise pictures for us for example.
witko999 said:
I really dislike the fact that something doing exactly what it's programmed to do and nothing more, is these days referred to as AI.
Agreed. "AI" is nothing more than a complex script or algorithm that todays abundance of computing power has made seem fast and impressive. Nothing like true AI at all.
king long list of nested 'if' and 'else' statements 
I have worked with a number of the potential things that could be categorised as AI enablers in the unstructured information environment for some time.
Simplistic tools such as Case Based Reasoning 30 years ago, which was an early way of capturing expert knowledge to be deployed to enable lots of less experienced people to benefit from the knowledge of an expert. Not decision tree based, but analysis of problem and solution text and then qualifying questions from the cases to narrow the candidate set. Every response caused the remaining cases to be linguistically analysed to narrow them, rinse and repeat the text analysis and ask next most relevant question until a small list of potential resolutions were presented. Use cases were customer service analysts receiving calls with a technical element. Decision trees were crap as they didn’t present the knowledge and experience to talk a user through the issue over the phone without a script.
Combinations of tools such as, linguistic analysis, Bayes, LRC and similar linguistic and mathematical algorithms have been very useful in some of the spook work, health, logistics, call centre (like cross selling as you make a customer happy by fixing the problem - good time to sell etc), diagnostics, trend spotting (manufacturing and retail of drinks for specific geographies - classic example, the ready to drink bottles ensuring that your brand of spirit gets into the Monica rather than Tesco’s brand - but also about who, what and where, Ibiza clubs or Devon pubs), expertise analysis and social grouping (without having to join a specific group, just the relationships between content consumed, authored, searches run, and people interacting with content etc). All of these are internal solutions and not public facing, which is probably the difference as few people will invest the money and effort required into providing a quality public facing point solution. Text to speech has also been useful from video and spoken content as a part of that unstructured set. Sometimes automatic language translation is required, but introduces inaccuracies.
All of the above have been good systems using quite rudimentary tools. Today’s tools are “marketed” as better, but have their issues. Challenges today (as always) are training the more advanced AI systems is prevention of bias from the content set you use to train it. The algorithms take some serious understanding as to why it made a certain decision on training content and how to tune that (in the tool or training set) without stopping it making good decisions in other cases. User behavioural analytics is also now useful in certain applications such as IT security etc and frequently AI based.
Overall, in solutions where a specific business problem exists and a specific enabler to analyse massive amounts of unstructured content is required, I have seen them work well. Areas where I consistently see them not working so well is in public facing solutions, such as chat bots and similar dross where the input the tool gets is more restricted. I used the O2 chat bot last night and for my specific issue it was worse than useless, just a waste of my time.
TLDR: Have built and seen lots which work well in advising business. I have yet to see a decent public facing solution. The Microsoft paper clip used Bayes if I remember correctly, and we all know how useful that was!!!
Simplistic tools such as Case Based Reasoning 30 years ago, which was an early way of capturing expert knowledge to be deployed to enable lots of less experienced people to benefit from the knowledge of an expert. Not decision tree based, but analysis of problem and solution text and then qualifying questions from the cases to narrow the candidate set. Every response caused the remaining cases to be linguistically analysed to narrow them, rinse and repeat the text analysis and ask next most relevant question until a small list of potential resolutions were presented. Use cases were customer service analysts receiving calls with a technical element. Decision trees were crap as they didn’t present the knowledge and experience to talk a user through the issue over the phone without a script.
Combinations of tools such as, linguistic analysis, Bayes, LRC and similar linguistic and mathematical algorithms have been very useful in some of the spook work, health, logistics, call centre (like cross selling as you make a customer happy by fixing the problem - good time to sell etc), diagnostics, trend spotting (manufacturing and retail of drinks for specific geographies - classic example, the ready to drink bottles ensuring that your brand of spirit gets into the Monica rather than Tesco’s brand - but also about who, what and where, Ibiza clubs or Devon pubs), expertise analysis and social grouping (without having to join a specific group, just the relationships between content consumed, authored, searches run, and people interacting with content etc). All of these are internal solutions and not public facing, which is probably the difference as few people will invest the money and effort required into providing a quality public facing point solution. Text to speech has also been useful from video and spoken content as a part of that unstructured set. Sometimes automatic language translation is required, but introduces inaccuracies.
All of the above have been good systems using quite rudimentary tools. Today’s tools are “marketed” as better, but have their issues. Challenges today (as always) are training the more advanced AI systems is prevention of bias from the content set you use to train it. The algorithms take some serious understanding as to why it made a certain decision on training content and how to tune that (in the tool or training set) without stopping it making good decisions in other cases. User behavioural analytics is also now useful in certain applications such as IT security etc and frequently AI based.
Overall, in solutions where a specific business problem exists and a specific enabler to analyse massive amounts of unstructured content is required, I have seen them work well. Areas where I consistently see them not working so well is in public facing solutions, such as chat bots and similar dross where the input the tool gets is more restricted. I used the O2 chat bot last night and for my specific issue it was worse than useless, just a waste of my time.
TLDR: Have built and seen lots which work well in advising business. I have yet to see a decent public facing solution. The Microsoft paper clip used Bayes if I remember correctly, and we all know how useful that was!!!
I think a lot of the 'big data' analytical stuff is pretty damn good- whether it falls strictly into the AI bracket or not I'm not entirely sure but it's certainly along the same lines as the focused inbox example.
As much as people complain about targeted advertisement in browsers, on facebook, etc - at least that stuff is somewhat relevant to me. I've got no issue really in being advertised stuff that I might potentially spend money on. If the interworld was completely anonymous then I'd only be spammed with adverts for stuff which has zero relevance - so I think that is a "properly working" example of such analytics.
In the cyber security world there are some pretty powerful examples too, being able to identify malware et al. without any predefined context based purely on behaviours and probability.
A huge factor of AI is that it's silent and trundles along in the background, so as another poster put it - the best implementations of it you're probably completely unaware of!
As much as people complain about targeted advertisement in browsers, on facebook, etc - at least that stuff is somewhat relevant to me. I've got no issue really in being advertised stuff that I might potentially spend money on. If the interworld was completely anonymous then I'd only be spammed with adverts for stuff which has zero relevance - so I think that is a "properly working" example of such analytics.
In the cyber security world there are some pretty powerful examples too, being able to identify malware et al. without any predefined context based purely on behaviours and probability.
A huge factor of AI is that it's silent and trundles along in the background, so as another poster put it - the best implementations of it you're probably completely unaware of!
Dogwatch said:
I'm not a medic but I understood there have been some successes with tasks such as scanning slides for cancer cells.
Oh absolutely! I'm definitely aware of the successful applications in that sort of field, and in any sort of relatively structured environment I can see it working extremely well.Thinking back to my original Outlook example, however, it's trying to decide what's important to me not on a pretty well defined - albeit potentially vast - set of views of what cancer looks like, but on what email content is going to be important to me personally. I can see how it could work after I've been using the tool for a while so it has had a chance to analyse what I open, what I delete without reading and what I maybe leave open long enough to read in preview then apply that learning to new inbound messages, but in trying to perform that task from minute one of me using the software, it has got it so badly wrong that I've switched the function off and will probably never turn it back on again!
I can't help thinking that it would've been better to just give me a message as part of the installation process telling me that in a month from now, it's going to present me with its view of focused vs other based on my history to date, at which point I might find that it's actually accurate and keep using it?
The autonomous drive option on my new V60 - where the car actively steers around bends whilst on cruise control - properly freaks me out. You still need to have your hands on the wheel but it'll do all the heavy lifting and I'm amazed at how accurate it is. Wouldn't trust it on B-roads obviously, but for motorways it's perfect.
Kermit power said:
Oh absolutely! I'm definitely aware of the successful applications in that sort of field, and in any sort of relatively structured environment I can see it working extremely well.
Thinking back to my original Outlook example, however, it's trying to decide what's important to me not on a pretty well defined - albeit potentially vast - set of views of what cancer looks like, but on what email content is going to be important to me personally. I can see how it could work after I've been using the tool for a while so it has had a chance to analyse what I open, what I delete without reading and what I maybe leave open long enough to read in preview then apply that learning to new inbound messages, but in trying to perform that task from minute one of me using the software, it has got it so badly wrong that I've switched the function off and will probably never turn it back on again!
I can't help thinking that it would've been better to just give me a message as part of the installation process telling me that in a month from now, it's going to present me with its view of focused vs other based on my history to date, at which point I might find that it's actually accurate and keep using it?
I'll admit in my experience (supplying and consulting on the Microsoft '365 stack) most customers switch off focused inbox. The concept is sound though if you leave it long enough, as long as you train yourself to open up the "other mail" part occasionally you can train it up.Thinking back to my original Outlook example, however, it's trying to decide what's important to me not on a pretty well defined - albeit potentially vast - set of views of what cancer looks like, but on what email content is going to be important to me personally. I can see how it could work after I've been using the tool for a while so it has had a chance to analyse what I open, what I delete without reading and what I maybe leave open long enough to read in preview then apply that learning to new inbound messages, but in trying to perform that task from minute one of me using the software, it has got it so badly wrong that I've switched the function off and will probably never turn it back on again!
I can't help thinking that it would've been better to just give me a message as part of the installation process telling me that in a month from now, it's going to present me with its view of focused vs other based on my history to date, at which point I might find that it's actually accurate and keep using it?
Agree a better implementation would be to kick it in after a few weeks/months but then that would just confuse a different demographic instead.
There's also a chance that your organisation e-mail security filters/scanners are not setup correctly too given the example of your boss' message ending up in the "other" mail section. It may have been inadvertently stamped as a 'bulk mail' message which would contribute to it's poor filing.
Ransoman said:
witko999 said:
I really dislike the fact that something doing exactly what it's programmed to do and nothing more, is these days referred to as AI.
Agreed. "AI" is nothing more than a complex script or algorithm that todays abundance of computing power has made seem fast and impressive. Nothing like true AI at all.Adaptive algorithms do lots of cool stuff all the time.
Dogwatch said:
I'm not a medic but I understood there have been some successes with tasks such as scanning slides for cancer cells.
I think that's due to the 'sensors' in the technology versus the human. They may also set the computer to find positives at a lower threshold than a human. Also, a computer program isn't going to rush like some outsourced diagnostics technician in a far-off land whose ass is getting ridden by their supervisor....But in terms of real world examples I've seen - yeah Ebay has been great telling me about the batteries I could buy, which I'd already purchased a week ago.... Or that Google helpfully tells me the reviews of the restaurant/cafe I'd looked up previously, when I'm already sitting in the place.
Edited by rodericb on Friday 24th April 14:45
Kermit power said:
My latest encounter with AI in the real world... Having recently changed jobs and moved from Lotus Notes to Microsoft Outlook, I just had a chasing phone call from a manager asking why I hadn't completed a task that had been emailed out...
What task? I asked. What email?
This is where we discovered that by default, Microsoft decide to bless your working day with something called a "Focused Inbox" and spare you whatever it thinks you don't need to see by putting it in an "Other" inbox.
I'm sure everyone can guess where the helpful AI assistant had decided to file my manager's email, and I discovered a bunch of other relevant messages I knew nothing about in there as well! It's now turned off and I've just got one inbox thankfully, but it did get me thinking about how rarely I've seen AI actually work in the real world. Yes, great for crunching vast amounts of data in a controlled environment, but does it ever actually work as soon as it needs to try and analyse real human needs or responses?
That Outlook feature, like so much MS software is just tosh and bad design IMO.What task? I asked. What email?
This is where we discovered that by default, Microsoft decide to bless your working day with something called a "Focused Inbox" and spare you whatever it thinks you don't need to see by putting it in an "Other" inbox.
I'm sure everyone can guess where the helpful AI assistant had decided to file my manager's email, and I discovered a bunch of other relevant messages I knew nothing about in there as well! It's now turned off and I've just got one inbox thankfully, but it did get me thinking about how rarely I've seen AI actually work in the real world. Yes, great for crunching vast amounts of data in a controlled environment, but does it ever actually work as soon as it needs to try and analyse real human needs or responses?
As for AI - it is a buzz word and current trend. But it is nothing new and has been about as long as computers have pretty much. You will have been using it for decades most likely. And most of the time it works perfectly well.
srappy said:
A simple one for me is the way that AI is helping to sort some of the mountains of data that we create continuously like say pictures - AI automatically can identify objects and organise pictures for us for example.
lol, or more accurately it is just a computer following a computer program based on inputs & calculations.Gassing Station | The Lounge | Top of Page | What's New | My Stuff