Turtles all the way Down

What have I learned from working inside the AI black box with Aman & Henry?

I've been working with generative artificial intelligence with students in my computer technology program since 2018 when we were fortunate to get a new grade 9 whose dad was on the team that brought IBM Watson to Jeopardy. That got us connected to IBM cloud and building AI chatbots five years before the "AI revolution" everyone has been caught out by.

That wasn't our first point of contact with AI though. I'd been keeping an eye on AI dev as far back as 2015 because we launched our gamedev course then and getting handle on building intelligent responses to player actions in our games immediately became our biggest challenge. Thanks to Gord and IBM we were able to get our juniors familiar with AI prior to asking them to take on significant software engineering challenges with it in the senior grades.

I presented on AI use in the classroom at the ECOO conference pre-COVID in fall of 2019. Gord from IBM even came all the way down to Niagara Falls to offer world class suppport. The room was all but empty:

This is how many Ontario educators (already interested in edtech because this is ECOO!) you get in an introduction to gnerative AI in 2019 (yes, it was four in an otherwise empty room). Ahead of our time (again)? Four years later it's an emergency and suddenly there are education AI experts everywhere. I wonder where they were in 2019.

If you ever wonder why education always seems two steps behind emerging technologies that will have profound impacts on classrooms, here's a fine example. Except you won't even see four people sitting in an empty room in 2024 because all edtech conferences like ECOO focused on teacher technology integration have evaporated in Ontario.

***

OK, so I've been banging my head against pedagogically driven AI engagement in education for almost a decade only to see it swamp an oblvious education system anyway, so what's happening now? I'm ressearching the leading edge of this technology to see if we can't still rescue a pedagogically meaningful approach to it.

In the summer Katina Papulkas from Dell Canada put out a call for educators interested in action research on AI use in learning. I've been talking to Aman Sahota and Henry Fu from Factors Education over the past year looking for an excuse to work on something like this, so I pitched this idea: De-blackboxing AI technology and using it to understand how it works.

Our plan is to use the Factors AI engine that Henry himself has built and Aman administrates to build custom data libraries that will support an AI agent that will interact with students and encourage them to ask questions to better understand how generative AI works. As mentioned before on Dusty World, GenAI isn't intelligent and it's important that people realize what it is and how it works to demystify it and then apply it effectively. Getting misdirected by the marketing driven AI hype isn't helpful.

So far we've built modules that describe the history and development of AI, how AI works and the future of AI. In the process of doing this I've come across all sorts of public facing research material that breaks down generative AI for you (like Deep Learning from MIT Press), but it's technically dense and not accessible to the casual reader.

During the last week of August Factors had a meeting with interested educators through UofT OISE (their AI system came out of the OISE edtech accelerator). I demonstrated in the presentation how the AI engine might be used to break down a complex article for easier consumption through agent interaction. The example was WIRED's story on how Google employees developed the transformers that moved generative AI from a curiosity to real world useful in the late teens. I picked this one because it explains some of what happens in the 'blackbox' that AI is often hidden in.

With some well crafted prompting and then conversational interaction, students can get clear, specific answers to technical details that might have eluded them in the long form article. The reading support side of GenAI hasn't been fully explored yet (though WIRED did a recent interesting piece on cloning famous authors to become AI reading buddies as you tackle the classics which is in the ballpark).

What have I learned from working in the engine room (BTW, that image at the top is Adobe Firefly's AI image generator) building an AI data library and then tuning it? AI isn't automatic at all. It demands knowledgable people providing focus and context to aim it in the right direction and maximize productive responses with users. An interesting example of this was finding documents that provided relevant data on the subjects we wanted the AI to respond to. When I couldn't find specific ones Henry suggested using Perplexity, an AI research tool that coalates online sources and then gives you concise summaries along with a bibliography of credible sources.

I thought I was being perverse asking Factors to design an AI that expalins AI using AI, but Henry's always a step ahead. His suggstion is to use an AI to build a library of information to feed the AI engine that then uses AI to interact with the user... about AI. It's turtles all the way down!

A.I. Isn't What You Think It Is

 I've been in a series of presentations over the past couple of months where organizations are getting frantic about catching the 'AI Wave'. This urgent need to feel like they aren't missing out on a fad is understandable, but like so many emerging technologies, getting 'into it' won't be effective if you ignore the foundations its built on, and the foundations of AI and the technology itself are... problematic.

You can't have 'generative' AI without massive data sets to train it on. This data is scraped from the internet and then fed into systems that can eventually give users "a statistically likely configuration of words" that look like an answer. That's right, the brilliant answer you just got on a generative AI platform isn't really an answer, it's a cloud computer cluster giving you its best guess based on crowdsourced data. None of that stops people from thinking it's intelligent (it isn't), and being in a panic about missing out.

Putting the fact that AI isn't nearly as smart as the marketing portrays it aside for a moment, large data and the cloud infrastructure that stores and delivers it are a house of cards teetering on the edge of collapse. You can't have AI without climbing to the stop of this wobbly infrastructure. How precarious is it? Data growth worldwide is in an explosive phase of growth (partially driven by the AI fad). Our overloaded storage infrastructure is under pressure because AI uses it much more aggressively that simply storing information. AI demands fast data retrieval and constant interaction making the rise in AI particularly problematic for our stressed storage systems.

We're facing data storage shortages in the next couple of years because of our belief that the cloud is an infinite resource. It isn't, it's an artfully hidden technological sleight of hand. The irony is that our digital storage infrastructure limitations will also end up limiting our current crop of AIs as well.

The staggering environmental costs that underlie our myth of an infinite digital cloud haven't  been mentioned yet, but like many of our other ecological marketing myths (electric vehicles), pushing the messy business of how it works out of sight of the consumer is a great way to market a green future while doing the opposite. Data centres in the US consume over 2% of all electricity in the country. There are benefits to scaling large data centres, but the trend into the foreseeable future is that the cloud will continue consuming more energy out here in the real world. That we're increasingly throwing limited resources at building AI guessing machines tells you something about our priorities.

One of the first posts on Dusty World was about dancing in this datasphere twelve years ago. Back then I'd found a quote by Google CEO Eric Schmidt talking about the coming information revolution:

I'd make a distinction between information and data. One is useful, the other is raw binary numbers and storing the majority of it is a complete waste of time and resources. Sussing out information from data is an ongoing challenge. That doesn't change the fact that the amount of data being generated back then wouldn't even register on the graph below, which looks like a runaway growth curve - you can make good money from all that data.

So, we live in a world that is well into an aggressive phase of digital growth, though very few people understand how any of that works. Even as we compile more content than we have in the entirety of human history to feed the attention economy, we also decide to play a sleight of hand game with machine learning on massive datasets just to see if it'll work.

From an educational perspective, AI is in the wild now and ignoring it will only get you and your students in trouble. If we're going to make functional use of this progenitor of true artificial intelligence, we need to teach the media literacy around it so that people understand what it is, how it works and how best to use it to amplify rather than replace their humanity.

I've seen a lot of people panicking about AI taking their jobs away, but if your work output is a statistically likely configuration of words, then you're not applying much of your vaunted human intellect to the task at hand and probably should be replaced by one of these meh AIs. But if you're one of those humans who actually thinks, even this stunted AI can be a powerful ally. In a fight for intellectual supremacy who would you think would win?

1. just machines
2. just people
3. an empowered hybrid of the two

The move here during this awkward adolescence of artificial intelligence where we're faking it until we make it is to leverage the tool to best effect. If effective use of AI speeds up our ability to gain actionable information in the chaos of data that surrounds us, then we can more quickly move on to the next real steps in technology evolution.

The other day I described AI as we currently define it as a hack to keep classical computing ahead of the data tsunami we're living in. At the time I was surprised by how I described it, but classical computing is reaching the limit of what it can do. For the past few years we've been finding speed in parallel processing such as adding computing cores to CPUs rather than making faster ones. We've also been finding efficiencies in how we manage data such as creating more organized memory caches to better feed our processors. Ultimately, I feel like generative AI in 2023 is another one of these patches. It's a way to make our overwhelming data cloud more functional to us.

This is from a presentation I've been giving that attempts to bring people into a better understanding of the hype. AI (even this meh one) will replace you if you let it, so don't!

Digital technologies aren't going to go anywhere, but they are a 'low resolution' way to compute. There is also the problem of reducing the complexity of reality to ones and zeroes. Mathematical concepts can help us understand relationships, but they will always be inherently reductive; they're never the thing itself but a simplified abstraction of it. Digital reduces the world to ones and zeros, but there is a better way.

When we run out of nanometres like we have with electronics, the next step is a big one, but it's one we're working on globally as a species right now. In the next decade we're gong to figure out how to use the building blocks of nature itself to compute at speeds that classical computers can't imagine. What will this do for our data clogged world? One of my hopes is that it will process much of that data into usable information, information that we can then use to solve this mess we've gotten ourselves into. When you have an answer you no longer need to keep the information leading to it.

I'll weather the current AI hype storm, but if you ask me what I'm really excited about it's artificial intelligence realized on a fault tolerant quantum computer. The future beyond that moves in directions I can't begin to guess, and that is exciting. imminent and absolutely necessary if we're going to prevent a global collapse of human civilization. Some people might get panicky about that, but they're the same ones who think a cloud based statistical guessing machine will replace them.