"AI for All" Is Getting On Our Nerves

We live in the glorious land of maple syrup and hostile wildlife, and we are therefore obliged to pay attention to the burps and grumbles of Ottawa more often than not. A recent expulsion of noise has caught our attention, and unfortunately, we have some concerns.

Our Bureaucrat-in-Chief, Mark Carney, made a statement on AI and officially launched a much-celebrated “AI for All” strategy that the more technically-inclined and cynical among us may describe as rather quaint. It reads, to us, like a after-school special in which a bubbly and non-threatening host encourages all the good little boys and girls to sing the alphabet, but with infinitely more corporate buzzwords.

Dear readers, we spent quite a few years in corporate office work during our career, and corporate buzzwords are and always have been so much wasteful drivel. Let us cut through to the meat and bones of this strategy and try to puzzle out what the government actually intends to do.

The statement at the link above has a large preamble that we frankly do not intend to engage with because we have not the brain cells to spare for that particular task, so we shall skip directly to the second part: that of the government’s action plan. Our comments are below.

“To build trust we will[…] [strengthen] protections for Canadians’ personal information[…] and [introduce] an online safety regime to better protect social media and chatbot users”

As per our previous article discussing this bill, we suspect the government’s efforts in this area may not be effective. But we will give them partial credit for the effort.

[We will] strengthen multinational partnerships with trusted allies

Oh yes? Which allies? Wait, wait, there’s a list here farther down the page.

Canada’s new government has signed agreements and joint statements with Australia, the European Union, Finland, Germany, India, Norway, Qatar, Saudi Arabia, Spain, Sweden, the United Arab Emirates, and the United Kingdom.

Interesting that the two biggest players in AI, the USA and China, are not on this list.

[We will] Establish a National AI Literacy Initiative[…] As part of this effort, AI literacy will reach 1 million entry-level post-secondary students

What, exactly, is this going to look like? It mentions “free AI learning kits” but what does that mean? Leaflets? ChatGPT accounts? Is there something required apart from the ability to type text into a box and read the answer? We are entirely annoyed by this on its face, because it suggests that students somehow need to be taught how to use these systems that every major tech firm is trying to force into every piece of software under their control, as if they are not being constantly bombarded by it on a daily basis already. We also find it horrifying that the government thinks that this is a good initiative considering a recent study that using LLMs causes cognitive decline, something that appears to be borne out by anecdotal stories we have observed on Reddit.

[We will] provide access to trusted AI agents for every post-secondary student

How, exactly? Again: free ChatGPT accounts? Or perhaps will the government sink millions of dollars into training their own? Who will provide these? Did the universities and colleges agree to this? Can they refuse if they wish? To date, there is much bluster about how incredible and powerful AI agents will be, and scant little evidence of what they actually are.

[We will] provide up to 90,000 AI-related jobs and work placement opportunities for young Canadians[…] help small and medium-sized businesses adopt AI to support workers, raise productivity, and drive breakthroughs in priority sectors such as health, energy, transportation, agriculture, manufacturing, robotics, and government services.

The bold is ours.

The government of Canada, who should frankly know better, have gotten confused. Please understand that there are two different things that are frequently conflated when they should not be, and we (with our background in technology) can explain.

What is colloquially known as “artificial intelligence” today are all Large Language Models. These systems underpin the chatbot services like ChatGPT and its ilk, and coding tools like Cursor. An LLM is simply an advanced neural network, which is a machine learning algorithm; such networks have been developed and used in various technical applications for decades. (Do not let the name fool you. These networks are loosely based on the concept of neurons and synapses in the brain but do not actually function like a brain, if only because no one truly understands how the brain functions.)

Neural networks were interesting to us when we learned of them as part of data analysis some ten-ish years ago. These systems are marvelous tools, truly powerful in their own right. The classic example is one of a neural network trained on thousands of images of cancer cells, effectively “teaching” or loading visual information about what is a cancer cell into the system, which can then be used to identify such cells in new images supplied from patients. Machine learning is fascinating, dear readers, and has incredible potential when applied to problems which do not have a single set solution.

These systems are probabilistic. They will always produce the most likely output matching a given input, based on the training data. They are, at heart, dumb machines doing mathematics. When having the most likely answer is good enough for a given problem, then machine learning shines as long as the training data is high quality and specific to the task. Unfortunately, these most excellent machine learning systems are now being squashed into the same box as LLMs because they both come from the same technological origin.

A Large Language Model is a neural network trained on enough samples of human-generated text (i.e. the Internet at large) that it can take in natural language as an input and respond with natural language as an output. The output, in this case, looks like something written by a human, and we humans have a most terrible and troubling flaw when faced with something that resembles a human: we tend to begin thinking of it and treating it as a human. For LLMs, it is a case of “if not human, then why human-shaped?” writ large, and from there the term “artificial intelligence” took hold to describe them. At a glance, it does appear as if these systems are intelligent, just like in popular sci-fi.

To be clear: these systems are not intelligent. They have been built to be very good at simulating human conversation, and nothing more.

Now, the issue here (and more specifically that of the “AI for All” bill) is that conflating modern LLMs with other machine learning systems is unwise at best. They may have similar underlying technology but they are very different; we would argue that almost every time anyone says that AI has done some useful thing, what they actually mean is that a properly trained neural network has done some useful thing, and inevitably it turns out that an LLM was not involved.

The industries listed here by the Canadian government are specifically those for whom machine learning has been used for quite some time; by and large, most are organizations with a sufficient amount of data who have both a need for nonlinear data analysis and the resources to do it. But small and medium sized business, we assume, are meant to “adopt AI” by using LLMs in some way, because these businesses will absolutely not be building and training their own neural networks. And here we become unreasonably annoyed because how they are meant to use LLMs is not explained; we would even argue that, say, a restaurant has little need for such a system and certainly shouldn’t be spending their money on one.

Let’s move on.

[We will] launch the first AI Missions Program, with a flagship health mission to accelerate the adoption of AI in diagnostics

They mean neural networks trained on high quality data.

[We will] provide training and upskilling opportunities for workers from mid-career professionals to frontline workers so they can adapt to AI-enabled workplaces

Why, why would we want this when stories abound of workers whose jobs and mental health are being utterly degraded by the LLMs forced on them? What even is an AI-enabled workplace and why would anyone want to work there?

To reinforce sovereignty, we will[…] build the foundations of sovereign Canadian AI: compute, cloud, connectivity, data, and talent[…] Build a world-leading public AI supercomputer

No, no you will not.

Creating an equivalent to ChatGPT, based in Canada, on Canadian hardware, using Canadian data, is a fantasy. Canada does not have the billions it would take accomplish that, nor the political mandate. We suspect that what they mean is that they will… build a scant few data centers which will inevitably contain a relatively small number of Nvidia GPUs, to be rented out to Canadian firms who intend to train minor models for their own purposes and have a hard requirement to keep the data in Canada due to privacy laws.

Support globally competitive Canadian champions by improving access to growth capital, using government procurement as a strategic anchor customer, and helping Canadian AI companies access compute, commercialisation resources, and intellectual property protections.

This has so many buzzwords that we struggle to understand what it actually means.

Expand Canada’s AI talent base through investments in -

Listen, we are not prepared to read any more of this idiocy. It is tiring. The “AI for All” strategy was informed by national consultations and an AI Strategy Task Force that is composed entirely of academics, business people whose livelihoods depend on all this being successful, and lay people who do not understand the technology. The consultations, which ran for a month last year, used leading questions that assumed AI is a positive development and made no distinction between LLMs and regular, boring, useful neural networks. To say that we are unimpressed is an understatement.

Our hope, if we have any at this point, is that nothing much gets done with this in spite of government enthusiasm; or at least that much of the government’s efforts are geared towards, again, regular boring neural networks with actual utility. LLMs in their current form are a fundamentally broken technology driven by the egregious waste of precious natural resources, built on hype, overvalued stock, and the fever dreams of rich white men who have no concept of the work done by their employees.