Board index Other Hangar talk

The age of AI - or not...

Talk about (almost) anything, as long as it is no serious FlightGear talk and does not fit in the other subforums.
Forum rules
Please refrain from discussing politics.

The age of AI - or not...

Postby Thorsten » Wed Aug 29, 2018 7:03 am

Supposedly, we're moving towards a world in which google creates an AI to solve major problems of the planet and amazon knows your desires before you are even aware of it.

The everyday reality looks a tad different though.

As we had the release of part II of our 'Cliodhna' fantasy series on Sunday, I ended up trying to get some picture of access statistics from YouTube (owned by supposedly data-smart google).

The page shows both access statistics for every video clip and their sum - unfortunately the two numbers were 5 apart when I checked first (the difference then started to vary, at one point reached a whopping 20, but...). Also, on the same page I could conjure up the number of subscribers to the channel in two different places - which were also two different numbers.

How hard can it be for a multi-billion dollar company that advertizes how it can crunch 'big data' to retrieve a handful of numbers from a database, add them correctly and display the result on a webpage? That doesn't sound like rocket science to me... I mean, computers could add integer numbers correctly well before I was born...

And given that google is really unable to do such a simple task - how likely is it that the complicated stuff actually has reasonable output (in fact, Amazon never managed to offer me what I like - despite me trying to teach the algorithm, but that's a different story...)
Thorsten
 
Posts: 10156
Joined: Mon Nov 02, 2009 8:33 am

Re: The age of AI - or not...

Postby ludomotico » Wed Aug 29, 2018 10:49 am

This is a common problem in big data, distributed databases and the need of an immediate answer to the user: many not critical systems prefer to offer fast but approximate results.

Check this link for ElasticSearch: https://www.elastic.co/guide/en/elastic ... tions.html In the next section, you'll see you can even configure the precision threshold of a cardinality request to suit your needs. Many other indexed databases work in a similar way.

Probably, data providers must warn the users when the results they show are approximations. I'm sure that, if needed, data providers use a different, not immediate type of query to calculate the exact number. For example, at night for billing purposes.

Anyway, being able to process partial data and give an answer accurate and fast enough to allow making an informed decision, in my book, is a good definition for "intelligence" :)
User avatar
ludomotico
 
Posts: 1003
Joined: Tue Apr 24, 2012 1:01 pm
Version: git
OS: Debian GNU/Linux

Re: The age of AI - or not...

Postby Thorsten » Wed Aug 29, 2018 11:08 am

Note that while I have no way of verifying exactness of the answer, the glaring issue is that the answer is not even self-consistent.

So even if the interface manages to pull only some approximate numbers from somewhere, it fails to add them correctly.

I'd fail any student who does that kind of presentation, so no, that's not a sign of intelligence.
Thorsten
 
Posts: 10156
Joined: Mon Nov 02, 2009 8:33 am

Re: The age of AI - or not...

Postby ludomotico » Wed Aug 29, 2018 11:16 am

If APPLES, ORANGES and TOTAL_FRUITS are random variables, I'm nearly sure there is a probability distribution where these numbers are possible:

APPLES = 100 (+-10%)
ORANGES = 100 (+-10%)
TOTAL_FRUITS = 190 (+-10%)

Anyway, I agree the presentation is confusing.
User avatar
ludomotico
 
Posts: 1003
Joined: Tue Apr 24, 2012 1:01 pm
Version: git
OS: Debian GNU/Linux

Re: The age of AI - or not...

Postby Thorsten » Wed Aug 29, 2018 12:06 pm

Yes, but in reality TOTAL_FRUITS is not a separate random variable - it's the sum of two random variables. Only if you know nothing about the other two would you assume it has a distribution. Your example defies the way probabilities work - there is no such distribution.
Thorsten
 
Posts: 10156
Joined: Mon Nov 02, 2009 8:33 am

Re: The age of AI - or not...

Postby ludomotico » Wed Aug 29, 2018 3:08 pm

Oh, sorry, in the experiment I completely failed to explain, TOTAL_FRUITS is not the sum of two random variables, but another, separated observation defined as another request to the system:

- Big data, please, can you estimate the number of apples? Search "apples AND NOT oranges" in Google: 60M results in 0.31 seconds.
- Big data, please, can you estimate the number of oranges? Search "oranges AND NOT apples" in Google: 31.1M results in 0.30 seconds
- Big data, please, can you estimate the number of apples or oranges? And the Big data, instead of adding the previous answers, runs the estimation algorithm again. Search "(apples AND NOT oranges) OR (oranges AND NOT apples)" in Google: 53.7M results. Also, if you reverse the terms: "(oranges AND NOT apples) OR (apples AND NOT oranges)", you only get 30.3M results: less than the individual terms!

Anyway, remember I completely agree that estimating separately terms and the addition of terms, when they are presented at the same time to the user, is not the best UI decision.
User avatar
ludomotico
 
Posts: 1003
Joined: Tue Apr 24, 2012 1:01 pm
Version: git
OS: Debian GNU/Linux

Re: The age of AI - or not...

Postby Thorsten » Wed Aug 29, 2018 5:15 pm

Well, I understand what the algorithm does - but it's evidence for a really dumb algorithm.

There's no world in which the following conversation would be any evidence of intelligence:

'Please tell me the number of women in this room' - 'I estimate 34' - 'Okay, and men?' - 'I see about 20' - 'So how many people do you think there are in total' - '50!'

Intelligence is about making use of what you already have and factoring in that if anything the total number needs to be higher than all parts of the sum you considered. Not to try to start every problem from scratch applying your same dumb scheme over and over.

(Also, retrieving the number of subscribers two times and getting different values each time is... not smart).
Thorsten
 
Posts: 10156
Joined: Mon Nov 02, 2009 8:33 am

Re: The age of AI - or not...

Postby ludomotico » Wed Aug 29, 2018 8:43 pm

Maybe the system is smart, but the web designers that requested the same data twice instead of reusing the first answer are dumb...
User avatar
ludomotico
 
Posts: 1003
Joined: Tue Apr 24, 2012 1:01 pm
Version: git
OS: Debian GNU/Linux

Re: The age of AI - or not...

Postby Thorsten » Thu Aug 30, 2018 3:53 am

Evidently not either, otherwise it could have alerted the web designers of their stupid mistake.

(If I'm smart but my student makes a stupid presentation of things I tell him, I would correct that).
Thorsten
 
Posts: 10156
Joined: Mon Nov 02, 2009 8:33 am

Re: The age of AI - or not...

Postby BEL ISAAK » Thu Aug 30, 2018 7:07 am

This gives me some insight in the issues we have with our (customised Microsoft) CRM-system. We are a youth member organisation and have about 22000 members. Every member has a record in our CRM-system, but still the CRM-system isn't able to correctly show us the evolution in our member base. We often get very weird results and are very disappointed that this prevents us from correctly analysing our member evolution.

We have different age-groups of youngsters between the age of 6 and 35. The system tells us that we have e.g. 8000 6-12 year old's, 5000 13-16 year old's, 5500 17-20 year old's and 4000 21+ year old's, but then instead of 22500 members, it tells us we have 22000 members. This confuses us a lot, because we are not able to look up where the difference of 500 originates and simply don't trust our multi-hundred-thousand € system anymore. Any suggestions on how to improve this?
User avatar
BEL ISAAK
 
Posts: 238
Joined: Sat Jun 04, 2011 2:52 pm
Location: Leuven, Belgium
Callsign: OO-ISA
Version: 2018.3.2
OS: Windows 10


Return to Hangar talk

Who is online

Users browsing this forum: No registered users and 2 guests