When is AI not AI?
Following: AI isn’t coming… it’s here!, let’s see how AI crept into IT for a while now overtly or covertly… Then please see a follow-on: Stop AI scraping your Internet data.
Update: see at bottom a discussion on a similar topic
The banner picture shows how insanely simple so-called generative AI (Wikipedia) is — and I’m no artist! — and let’s not re-tread what’s been said on opportunities & perils of AI… it’s too late! It’s been with us for a while now, we’re only being told by the media as GAFA (Wikipedia) jumped into the fray.
Like SUVs been around forever, but the Ford Explorer legitimized the market: AI has only just been legitimized by large players and by internet data and software capabilities; its previous incarnation in the 80s was Expert Systems (Wikipedia).
Google Search and Google Translate been around since 1995 and 2006, and contrary to some beliefs, they’re not mega online encyclopaediae or dictionaries. They’re in fact meta-dictionaries in the sense that they store data about what’s been searched or translated, in order to re-use it and improve searches and translations over time. While their avowed intent was to better rank pages to monetize advertising revenue, what happened behind the scenes was none other than machine learning (Wikipedia) and large langue models (Wikipedia) even if it wasn’t called that then. And those are the hallmarks of AI (Wikipedia).
It was very evident in Translate I used a lot in Hungarian and Latin. But Hungarian was so bad at the beginning that I used SZTAKI (online) a Hungarian academic translator... When you took to the option to contribute your own, you quickly realize Google only re-uses previous translations! In other words it improves over time with use, and that’s machine learning by any other name.
Amazon
In their Mechanical Turk (Wikipedia) starting 2005, you could volunteer or hire your time to help process outstanding questions posted online. What happens behind the scenes again, is that data about those searches and queries are stored and re-used as metadata to improve the on-line processing. That is none other than machine learning and eventually large language modelling as Amazon Web Services store and re-use the results.
Virtual Keyboards
I had Samsung Galaxy A3 (2017) and A12 smartphones for roughly 3 yr. successive intervals: Not only did I use Swype (Wikipedia) to speed up typing on the touchtone face of the phone, but I also had French and Hungarian keyboard settings; virtual keyboards are easily re-mapped and I can switch via a simple setting depending on what language I used. Predicitive typing improved over time so much so, that after couple years I needn’t change language settings: predicitive typing guessed at the language and did the job on a single setting! And if you think of Swype, there’s some clever filtering going on to ignore partially covered letters. In other words that’s none other than low-brow machine learning or adaptive execution of clever software quietly operating in the background. The downside also a confirmation was that a new phone meant re-training the new mobile…
YouTube / Music
The last but most startling examples are YouTube and YouTube Music starting in 2005 and 2015, respectively. Same as Google, the avowed intent is to capture eyeballs on screen and to keep people watching to thus help monetize the platform through advertising. My experience is that over the years, the suggestions on the right-hand side of the screen have been gradually improving. Again considering the volumes of metadata tags — so much so that it had to change its hits-counting algorithm late 2014 when Psy’s Gangman Style exceeded its capacity — this is none other than large language modelling and machine learning.
YouTube Music is the best case scenario for me: its suggestions are not only a way to keep me there LOL but also turned out to be a fabulous tool to keep up with artists, learn about similar genres and follow music I couldn’t have even going to gigs and talking to friends — I used to help run the Calgary Folk Music Festival 30–40 years ago, but moved around too much since then to ever plug into the local scene… even to the local Cambridge Folk Festival just up the road! — so much so that I can have viewing chats on-line with friends, where we simply follow together its suggestions posted as a function of our listening... YouTube Music is training itself to our taste!
Update 1: Pitch&Mix Meetup
This group asked if “AI should be used to create marketing materials?” The kick-off was that materials had been scraped (see follow-on post) and put together forever it seems, and AI only accelerates or facilitates it… as long as we admit to it transparently, and use it to improve our own research rather than replace it!
An interesting anecdote emerged there: everyone knows how IBM’s Deep Blue supercomputer AI beat chess champion Kasparov in 1996 & 1997 rematch (Wikipedia)… but no-one knows that the following day, Kasparov and Deep Blue together beat Deep Blue every time! It also emerged that we could not find anything on that on-line, so thorough was IBM’s supression of that event.
Let me relate a personal experience: My daughter’s year-end report was about Hungarian folklore, and she asked me to translate it into Hungarian to add couleur locale. I said, “No, it wouldn’t be your work”. I hit on this idea, however: I asked her to put it thru Google Translate, and I would do minor edits; that way, she did it, and I only assisted. Bingo! She was a happy camper… A simplistic case of hidden AI as discussed above, augmenting not replacing a task, and collaboratively to boot.
Footnote: Interestingly, this Cambridge Union debate occurred not 12 hrs. after the meetup: This House Believes Artificial Intelligence Is An Existential Threat
Update 2: My non-profit colleague
I have a Community Interest Company (see here) started as Cottenham.info to address isolation in rural East Anglia under the pandemic and climate change. It has now been revived post-pandemic as Cambridgeshire.ai to leverage AI for social prescribing (see here).
As my colleague and I were working on a prospectus to engage local community health services, we used both Bing AI (here) and Google Bard (here) to bird-dog NHS England services and information. I struck us that while prompt engineering was indeed important to properly pose the questions (see blog posts here and here), we were in fact the prompt engineers! Like the Deep Blue story above, it’s humans + AI that work rather than AI per se.
So to echo Mark Twain, “the reports of [human intelligence’s] death [vs. AI] are greatly exaggerated”.
