Let’s talk more directly about de-mystification of AI. Starting with some ways that you might come across AI in your daily life as an office worker: what are some AI touchpoints in today's world?
Wake up, check phone for weather: AI facial or fingerprint recognition is used to unlock your phone.
During breakfast, check Facebook on phone: AI recommender system populates your feed.
Stop by New World on way to work: AI facial recognition in store is used for crime prevention.
At work there are many touchpoints:
- When using your computer, AI network traffic behaviour analysis determines if there's a bad actor on the network
- AI text prediction in your word processor
- AI used in Copilot, ChatGPT, etc.
- Microsoft Defender heavily uses AI for anti-virus threat detection
Apply for credit line increase at bank: AI risk detection at multiple levels when dealing with bank's online website and your credit increase request.
Go home: Whilst driving on the street, AI is used in "average speed cameras" and AI is used by councils to monitor traffic flow.
At home, your rice cooker uses primitive AI to cook your rice to perfect consistency.
Go to a bar to meet a friend, take a selfie: AI adjusts color, exposure to give you a nice, pleasing image.
Back home watching Netflix, AI recommendation algorithm populates your feed with movies it thinks you'll like. if you're watchng a Netflix-made movie, AI was used to write the script to optimize what people like to see in a movie.
Watch some stuff on Youtube: AI recommendations algorithm populates your home screen / feed with content based on your Google browsing history and other signals.
Quick Zoom call with family member before bed: AI "pretty filter" smooths out wrinkles on your face.
...and likely a lot more situations.
Vanishingly thin
Some of the use cases above used to be considered magic when thery first came out and now we consider them boring. Examples are rice cookers with AI, funny real-time face filters on TikTok or Snapchat, and facial authentication for phone login. At some point in the past, each of these was wildly innovative. Do you remember when AI fuzzy logic rice cookers first arrived from Japan?
Facial recognition for login and cute filters on TikTok are still pretty clever! And I consider them good examples of AI. However, most people see these as more commodified, everyday things such that they're not considered special. As users, we get used to them quickly and they lose their shine. AI facial recognition on your phone is "how you log in." Augmented Reality AI filters on Snapchat are just "filters."
Long term utility
Right now LLMs are hot and that's what people associate with AI. From my perspective the massive shrinkage of CPU chips over the past decades has allowed for increasingly complex software to be run in smaller and smaller places. All the examples of AI in this blog post are just "software" to me. All of it clever, some of it astounding. When developers set out to write the software that performs these functions, it's rare that the developers think in terms of AI. That's more of a marketing term. Developers are more likely to think in terms of linear algebra, kv-stores, and matrix multiplication. Problems to be solved with code.
There is a lot of hype right now around AI, and there will be more to come. Before the dotcom bubble burst, there was a huge amount of scanned paper brochures that organisations proudly called "web sites." There's still some around today! The useful tech that survived the dot-com bubble such as ecommerce and advertising technology kept getting better over the years. No matter what happens with the current AI hype machine / bubble, useful tech will continue to thrive and improve no matter what.
De-mystifying AI is important so that we can understand it better. Of course, the AI vendors want you to see it as magic that only they can understand and provide to you. ;)
Update on randomness
I dug in a little bit on the randomness improvements Gemini 3 has made over ChatGPT. It turns out Gemini is applying some fairly simple filters to the query text such that when it sees "random" and "10,000 runs" and "most common" it knows to create and run a Python script instead of doing its probabilistic text generation like it usually provides. Clever. So to the user is appears Gemini is "reasoning" and providing much more random numbers than ChatGPT. The method it's using to do so is far from magic, but that doesn't matter much to the user if the answer is good. The point is, by de-mystifying the random number generation a bit, we can learn how it's done and maybe apply similar knowledge for our own use cases in the future.
Thank you to all the CDEI folks for allowing me the opportunity to post my thoughts over the past 2 weeks. I greatly appreciate it. I enjoyed reading the first two guest editors' posts and I look forward to reading what future guest editors will write! If you'd like to reach me, I welcome your email at [email protected]
Note: The preceding blog post series originally appeared on Stats NZ’s Centre for Data Ethics and Innovation hub which engaged with the larger community.