You can read the rest of this guide without understanding how a transformer works. You will get more out of it if you do. This chapter exists for readers…

You can read the rest of this guide without understanding how a transformer works. You will get more out of it if you do. This chapter exists for readers who want enough technical understanding to read AI claims with the right kind of skepticism, without needing to learn linear algebra. The chapter does not contain mathematics. It does contain the conceptual moves underneath the math, which is what actually matters for the rest of the book. The argument of this chapter is straightforward. AI is not magic, and treating it as magic, whether enthusiastically or anxiously, is the source of most confused public discussion. AI systems are statistical pattern-matching machines built through specific engineering decisions, trained on specific data with specific properties, optimized against specific objectives, and deployed with specific limits. Understanding the engineering is not the same as understanding the magic, because there is no magic. The engineering is what's there. You will leave the chapter with: a working model of what training actually is, what a transformer actually does, why scale matters and what it produces, what "fine-tuning" and "alignment" actually mean in practice, what the system can't do that it appears to be doing, and the conceptual grounding to read every subsequent chapter, especially environmental (Chapter 8), copyright (Chapter 11), and ethics (Chapter 15), with the right structural understanding of what's actually happening inside the machine. This is the longest of the foundations chapters. It earns its length. The rest of the guide is shorter and more focused because this chapter does the technical work once, well, and the later chapters can refer back rather than re-explain. ---

What "learning" actually means

Start with the most important conceptual move: when we say a machine "learns," we are using a word borrowed from human experience to describe something that has almost nothing in common with what humans do when they learn.

A human child learns the word "dog" through embodied experience: seeing dogs, hearing the word spoken in proximity to dogs, feeling fur, being knocked down by enthusiastic ones, gradually building a concept of dog that integrates visual, tactile, auditory, social, emotional, and narrative information. The child can later recognize a dog they've never seen before, a dog in a photograph, a cartoon dog, a dog described in a story they're hearing for the first time, and a dog they're imagining. The concept is robust across contexts because it was built from many kinds of input simultaneously.

When a machine "learns" the concept of dog, none of that happens.

What happens instead: the machine is shown many images, each labelled "dog" or "not dog" by humans. Each image is represented as a grid of numbers, typically three numbers per pixel describing red, green, and blue intensity. A modest image is a few million numbers. The machine has a mathematical function, a "model," with millions or billions of internal parameters (numbers) that can be adjusted. The function takes the image numbers as input and outputs a probability estimate: this image is X% dog. At first the parameters are random and the outputs are useless. The machine is then shown an image, makes a prediction, is told whether the prediction was right, and adjusts every parameter slightly in the direction that would have produced a more correct answer. This is repeated billions of times across millions of images.

After enough repetition, the function's parameters have settled into a configuration where the outputs are correct most of the time on the training images and, crucially, usually correct on new images the system has never seen. The machine has "learned" to recognize dogs in the only sense that applies: it produces correct dog/not-dog classifications at high rates.

What it has not learned: anything about dogs as living creatures, anything about the human relationships that give dogs their cultural meaning, anything about the experience of being a dog, anything about why someone might want to recognize a dog in the first place. It has learned a statistical map of pixel patterns that correlate with the human-supplied label "dog." The map is functionally useful. It is not the same thing as what the child built.

This distinction matters because it explains both the spectacular success and the strange failures of current AI. The machine is doing pattern-matching at a scale and speed humans cannot match. That's why it can read X-rays, find tumours, sort documents, recognize faces, translate languages, and write essays. The machine has no concept underneath the pattern-matching. That's why the same image classifier will confidently identify a school bus as an ostrich if a few specific pixels are altered (an "adversarial example"). The human pattern is so robust that we'd notice the manipulation immediately. The machine pattern is statistical, surface-level, and fragile in ways we don't intuit.

The technical name for what the machine is doing is statistical learning or machine learning. The mathematical sub-field is called optimization: adjusting parameters to minimize a measurement of error. The specific architecture used for most current AI is called a neural network, originally inspired (loosely) by the structure of biological neurons but now operating at a scale and in ways that have little practical relation to biology. Neural networks with many internal layers (typically dozens to hundreds) are called deep neural networks, and the field of building them is called deep learning.

You do not need to remember the vocabulary. You do need to remember the move: machine learning is pattern-matching at industrial scale, optimized against measurable objectives, with no underlying concepts. Everything else in this chapter follows from that.

---

The data is the model

Here is the second conceptual move, and it is the load-bearing one for everything the guide argues about copyright, environment, ethics, and sovereignty: what an AI model "knows" is a compressed statistical reflection of its training data. Change the training data, and you change the model. Change the data substantially, and you get a different model entirely. The model is, in a meaningful sense, the data, rearranged into a different form.

That is meant technically, not as a figure of speech.

When OpenAI trained GPT-4, they collected an enormous corpus of text (the public internet, large portions of digitized books, code repositories, and other written material) and ran it through their training process. The output is a set of model weights: hundreds of billions of numbers representing the statistical patterns the training process extracted from the input data. The numbers don't contain the original text. They contain something more abstract: the patterns of how words and concepts have appeared together in the training corpus.

When you ask GPT-4 a question, the model uses those patterns to generate a response. The response is novel; it isn't a copy of any specific document in the training data. But it is shaped by every document in the training data, because the patterns the model is following are extracted from those documents. The model is, statistically, the training corpus. Asking what an LLM "knows" is asking what was in its training data, encoded into the model's parameters by the training process, retrievable through the model's response generation.

🧌 GOBLIN CHECK — "Our model learned from publicly available data" is a sentence doing a lot of unpaid labour. The model didn't learn from the data the way you learned from your grade-eight science teacher. The model is the data, compressed and rearranged into a trillion dials. Ask what's in the pile, and notice how quickly the subject gets changed.

This is why the training-data question sits at the centre rather than the edge of everything that follows. Three direct implications:

For copyright (Chapter 11). If the model is the data, then training the model is an act that uses the data. Whether that use constitutes copyright infringement under Canadian law is contested (the academic argument from Carys Craig and Michael Geist is that it should be permitted under a text-and-data-mining exception; the creator-side argument from the Writers' Union of Canada and ACTRA is that it currently constitutes infringement). But the technical fact that the model is shaped by the data is not contested. The Canadian newspapers' lawsuit against OpenAI rests on this fact.

For environment (Chapter 8). If the model is the data, then training requires processing all of that data through the model's parameters, which is what makes training so energy-intensive. GPT-4's reported US$100 million in compute costs is, in essence, the cost of running a few trillion tokens of text through hundreds of billions of mathematical operations, repeated billions of times. The size of the training data, the size of the model, and the duration of training are the three drivers of training environmental cost. Inference is much smaller per query but happens many more times.

For sovereignty (Chapter 9). If the model is the data, then a model trained on Canadian data is, in a meaningful sense, a model that contains compressed Canadian information. If that data includes Indigenous content, the model contains compressed Indigenous information. The OCAP/NISR/CARE Indigenous data sovereignty frameworks introduced in Chapter 1 and deepened in Chapter 9 apply to the data in training, and therefore, by extension, to the model that results.

This is why the Stanford Foundation Model Transparency Index, in every edition since 2023, has identified training data as the single most opaque area of foundation model disclosure. The companies producing the models do not generally disclose what data they were trained on. When the Index scored 13 major foundation-model developers in December 2025, the data subdomain was the consistently lowest-scoring subdomain across nearly every developer. It is the subdomain where developers are most likely to face legal exposure (copyright lawsuits), regulatory exposure (privacy regulation), and ethical exposure (training-data biases reproducing in model outputs). The opacity is structural and consistent across the industry.

For our purposes in this chapter: when you read a claim about what an AI system can or can't do, ask what data was in its training set. Most of the time the answer is not publicly available. The non-availability is itself meaningful.

---

The transformer — what's actually inside ChatGPT, Claude, Gemini, Cohere

The architecture underneath every current major large language model is called a transformer, introduced in a 2017 paper from Google Brain called "Attention Is All You Need." One of the paper's eight co-authors was Aidan Gomez, who went on to found Toronto's Cohere in 2019, making Canadian technical contribution to the current AI moment more substantial than the "Canada is behind on AI" framing acknowledges.

What does a transformer actually do?

Start with a simpler version of the problem: predict the next word. Given the input "The capital of Canada is", a system trained to predict the next word would output a probability distribution over all possible next words. "Ottawa" would have the highest probability. "Toronto" might have non-trivial probability. "Banana" would have near-zero probability. The system isn't looking up the answer; it's generating a probability distribution based on patterns from training data, and the pattern of "the capital of Canada is X" appears in training data with X almost always being "Ottawa."

A transformer is this prediction problem, scaled up enormously in three specific ways.

First, attention. When predicting the next word, the transformer doesn't just look at the most recent word. It looks at every previous word in the input simultaneously, and uses a mechanism called "attention" to weight which previous words matter most for the current prediction. The attention mechanism, the title insight of the 2017 paper, is what lets transformers handle long-range dependencies in text: pronouns that refer back to nouns mentioned paragraphs earlier, logical connections that span sentences, the structure of arguments that build across multiple paragraphs. Attention is computationally expensive, which is why transformer training requires the specialized GPU hardware we discussed in Chapter 4 (forthcoming) and why the chips that NVIDIA ships are configured for attention-style matrix multiplications.

Second, layers. A modern large language model has dozens to hundreds of transformer layers stacked on top of each other. Each layer transforms the input representation slightly. Early layers handle local syntactic patterns: what word follows what. Middle layers handle longer-range patterns like sentence structure and paragraph logic. Late layers handle the most abstract patterns, such as argumentative structure, tone, register, and factual content. No single layer "knows" anything, but the stack of layers together implements something that functions like comprehension at scale. Why this works as well as it does is not well understood, even by the researchers building the systems. It is one of the open questions in the field.

Third, scale. A transformer with a thousand parameters does nothing useful. A transformer with a million parameters can do simple text generation tasks. A transformer with a billion parameters can write coherent paragraphs and answer simple factual questions. A transformer with a hundred billion parameters can do most of what current AI systems do. GPT-4 is estimated, by outside analysts since OpenAI doesn't say, to be in the trillion-parameter range. Competitors are assumed comparable, and none of them says either. Cohere's Command series operates in the tens-of-billions range. The scale produces qualitative shifts in capability that smaller models cannot match, a phenomenon called emergence in the AI research literature, though the term is contested (some researchers argue it's an artifact of measurement choices rather than a genuine threshold effect).

GOBLIN FACTS — scale has a number behind it. GPT-3, the 2020 model that kicked off the current wave, ran on 175 billion parameters across 96 layers (Brown et al., 2020). In this field "bigger" isn't a vibe. It's a count you can look up, and the count is most of the trick.

What scale costs: more parameters require more training data, more compute, more electricity, more cooling, more chips, more capital. The economics of foundation models is fundamentally an economics of scale. This is why only a small number of companies can train frontier models, why those companies have substantial capital backing, and why the "national champion" framing for any country's AI strategy ultimately depends on whether the country can support training at frontier scale. Canada's Cohere is real, but its scale is meaningfully smaller than OpenAI, Anthropic, Google DeepMind, or Meta. This is part of what Chapter 7's discussion of Canadian AI power dynamics is doing.

---

Training, fine-tuning, alignment — three distinct stages

Public discussion routinely collapses three distinct stages of making an AI system, and the collapsing produces confused arguments. The chapter separates them now so the rest of the guide can refer to them precisely.

Pre-training. The very expensive, energy-intensive stage where a model is built from scratch by processing massive quantities of text. The output is a "base model" or "foundation model": a system that can generate plausible-sounding text but has not been shaped to be helpful, honest, or safe. A raw pre-trained model would respond to "How do I build a bomb?" by generating plausible-sounding bomb-making instructions, because its training corpus contains such instructions in technical literature, fiction, news articles, and forum discussions, and the model's only objective is to predict plausible next-text. Pre-training establishes what the model knows. It does not establish how the model behaves.

Fine-tuning. A second, much cheaper, more targeted stage where the base model is trained on specific examples to shape its behaviour for particular tasks. Most commercial AI products (ChatGPT, Claude, Gemini, Cohere's enterprise products) are pre-trained foundation models that have been extensively fine-tuned. The fine-tuning teaches the model to follow instructions, to format responses helpfully, to refuse harmful requests, to handle particular professional domains (coding, medicine, legal research), and to maintain consistent tone and persona.

Alignment. A specific kind of fine-tuning aimed at making the model's behaviour match human values and intentions. The dominant technique is called Reinforcement Learning from Human Feedback (RLHF): humans rate model outputs as better or worse, and the model is trained to produce outputs humans rate as better. This is what makes ChatGPT polite, makes Claude refuse to write malware, makes Gemini cautious about election information, and makes Cohere's enterprise products stay on-task. Alignment is also where most of the contested ethical and political choices are made: what counts as "better," whose values define it, what topics the model will discuss and what it will refuse to discuss, what voices and perspectives are reflected in the training feedback.

Three operational implications:

A foundation model is not the same as a deployed product. When OpenAI releases ChatGPT, OpenAI is releasing a fine-tuned, aligned version of a base model that exists separately. The base model is more capable in some ways (it has access to more raw knowledge) and less capable in others (it doesn't follow instructions reliably). When researchers like Geoffrey Hinton, Yoshua Bengio, or the broader AI safety community talk about "frontier model risks," they are often talking about the base models, not the deployed consumer products. The distinction matters when reading any specific claim about AI capability.

Alignment is not solved. The current state of RLHF and related techniques is that they reliably make models more helpful and more polite. They less reliably make models more honest. They make models that refuse certain categories of harmful requests but can often be circumvented through "jailbreaking": adversarial prompting that gets the model to produce content its alignment training was supposed to prevent. The arms race between alignment techniques and jailbreaking techniques is ongoing.

Alignment encodes political and cultural choices. When OpenAI decides what ChatGPT will and won't say about contested topics (abortion, vaccines, climate change, specific political figures), those decisions are made by humans at OpenAI, reflecting OpenAI's specific institutional view of what AI should and shouldn't say. The same is true for Anthropic, Google, Meta, and Cohere. The political and cultural commitments of the companies producing the models are embedded into the models' behaviour through alignment. This is not a conspiracy theory; the alignment teams at these companies discuss their choices publicly, in technical papers and policy publications. But the choices are choices, and they are made by specific people at specific companies. Whose values define alignment is one of the questions the field is still working through.

ALIGNMENT — whose values got trained in? When a model answers a touchy question one way and refuses another, that behaviour was shaped by specific people at a specific company deciding what counts as a good answer. Ask whose judgment is baked into the refusal, and where you'd even go to contest it. "The AI decided" is a way of making the deciders disappear.

---

What these systems can't do — and won't be able to do soon

Honest engagement with AI requires understanding both what current systems can do well and what they cannot do, regardless of how the marketing frames them. This section is the chapter's corrective to the most common overclaims.

Current large language models cannot reliably perform genuine reasoning. They can generate text that describes reasoning, often quite well, but the underlying mechanism is still pattern-matching against training data. When an LLM produces a step-by-step solution to a math problem, it is generating the most plausible next text given the problem, not actually doing the math. This is why state-of-the-art models still produce confident errors on problems that are simple variations of problems they were trained on. "Chain-of-thought prompting" and the newer "reasoning models" (OpenAI o1, DeepSeek-R1, Claude with extended thinking) improve the reliability of reasoning-like behaviour, but the improvement is incremental: these models reason more reliably than their predecessors, not robustly across novel domains.

Current AI does not have access to facts in the way a database does. When ChatGPT tells you the capital of Canada is Ottawa, it is generating the most statistically likely completion of "the capital of Canada is X" based on training data. When the training data is rich, current, and consistent, the answer is correct. When the training data is sparse, outdated, or contradictory, the answer is confidently wrong: the phenomenon called "hallucination." The deeper issue is that the model has no way to tell which of its outputs are reliable and which aren't, because it doesn't have a separate fact-checker. Anything that looks like factual lookup is, mechanically, the same pattern-matching that produces creative writing.

Nor can current AI do tasks that genuinely require interaction with the physical world. No amount of text training produces a model that can fix a sink, perform surgery, or harvest wheat. The current generation of "embodied AI" (robotics combined with foundation models) is still in early research, and physical-world tasks remain far harder than purely-text tasks because the physical world produces continuous, novel, embodied data that AI systems cannot pre-train against in the way they pre-train against the internet's text corpus.

A subtler limit: the systems cannot meaningfully predict their own failures. When an LLM produces a wrong answer, it usually does so with the same tone of confidence as when it produces a right answer. Asking the model "are you sure?" sometimes elicits a correction and sometimes elicits a more confident restatement of the wrong answer. The model has no reliable internal signal for "I don't know" because its training objective rewards producing plausible-sounding text, not producing accurate calibration of uncertainty.

Current AI cannot replace human judgment in cases that genuinely require human judgment. Medical diagnosis, legal analysis, employment decisions, criminal sentencing: these are domains where AI systems can produce useful analysis as decision support, but where the human judgment involved is not just information processing. It is also accountability, context, the application of values, and the moral weight of consequences. Whether AI should be used in such decisions is contested (see Chapter 15). Whether AI can be used in such decisions in a way that genuinely substitutes for human judgment is also contested, and current evidence suggests not.

The reverse claim, that AI is "just" autocomplete and therefore doesn't deserve serious attention, is also wrong. Current systems demonstrably perform many tasks that previously required human expertise. The medical diagnostic accuracy is real, and so are the translation quality, the code generation, and the creative writing. But the systems are doing all of this through pattern-matching against training data, with no underlying conceptual understanding, with no reliable way to know when they're wrong, and with no genuine reasoning capacity. Both of the simple stories, "AI is magic" and "AI is nothing," are wrong. What's there is powerful pattern-matching with substantial economic, social, and political implications, and reading those implications clearly means keeping the real capability and the real limits in view at the same time.

---

A note on artificial general intelligence

You will encounter the term Artificial General Intelligence (AGI) in any extended AI conversation. The term refers to a hypothesized future AI system that would match or exceed human performance across most cognitive tasks. Whether AGI is achievable, when it might be achieved, and what it would mean if achieved are among the most contested questions in the field.

Chapter 1 introduced the foundational AGI definition from Legg and Hutter (2007): "Intelligence measures an agent's general ability to achieve goals in a wide range of environments." The chapter also introduced the Lewis/Whaanga/Yolgörmez 2024 critique that this definition draws, through a chain of citations, on the 1994 "Mainstream Science on Intelligence" document and the IQ-research tradition that produced The Bell Curve, a foundational issue that the AGI conversation does not generally engage with.

For Chapter 2's purposes, the technical observations:

AGI is not what current systems are. ChatGPT, Claude, Gemini, and Cohere's products are not AGI by any serious definition. They are specialized pattern-matchers that work very well on certain tasks. The capabilities they demonstrate are narrower than the marketing implies and broader than the dismissive accounts allow.

The disagreement about AGI timelines is genuine and substantial. Some senior AI researchers (Geoffrey Hinton, Yoshua Bengio, Demis Hassabis) believe AGI is plausibly achievable within decades and have public positions about the safety implications. Others (Yann LeCun, Andrew Ng, Sasha Luccioni) believe current architectures are fundamentally limited and that AGI would require breakthroughs not currently visible. The disagreement is empirical rather than ideological, turning on judgments about whether the limits of current AI systems are temporary engineering problems or structural features of the approach.

The political implication is the same regardless of the technical answer. Whether AGI is decades away or centuries away, AI systems are already being deployed in ways that affect Canadian workers, citizens, and communities. The policy questions the guide covers, from environment and sovereignty to IP, labour, ethics, and governance, are not contingent on AGI arriving. They are about the AI that exists now and is being deployed now. The AGI conversation is a distinct conversation, important in its own right, but separable from the immediate Canadian policy questions.

Whether Canada should commit to not pursuing the most advanced frontier capabilities, the position associated with Geoffrey Hinton, who left Google in 2023 partly to speak more freely about existential AI risks, and Yoshua Bengio, who has become a public voice on catastrophic-risk concerns, is a live question in Canadian AI politics, and the AI for All announcement does not engage it. (An earlier draft of this guide claimed the question appeared as a cross-party thread in the CBC-leaked strategy draft; the CBC article does not support that, so the claim is withdrawn pending a receipt.) The guide returns to the frontier-risk question in Chapter 20.

---

When prediction starts acting — AI agents

Everything to this point has described a system that produces text. You give it a prompt; it returns a paragraph, an image, a block of code. It predicts, and you decide what to do with the prediction. The most important shift in AI since late 2024 is the move from systems that produce output to systems that take action, and it is the shift that most tests the "it's just autocomplete" framing, so it's worth understanding before the policy chapters lean on it.

An AI agent is a language model wired into a loop. Instead of answering once and stopping, the model is handed a goal, a set of tools it can call (a web browser, a code interpreter, your email, a payments API, a calendar), and permission to act, observe the result, and act again until the goal is met or it gives up. "Book me a flight to Toronto under $400 and put it on the calendar" stops being a sentence the model writes back to you and becomes a sequence of things it does: search flights, compare prices, open the booking site, fill the form, charge the card, create the event. The engine underneath has not changed. It is still next-token prediction over patterns in training data.

EXAMPLE — your phone's autocomplete. When your keyboard suggests "running late" after you type "I'm," it isn't reading your mind. It has seen those words run together a million times and it's betting the pattern holds. A large language model is that same bet, scaled to most of the internet: still a bet, still no one in there. What changed is that the predictions are now wired to levers in the world.

That is the whole conceptual move, and it is smaller than the marketing suggests. An agent is not a new kind of mind. It is the same pattern-matcher from Section One, run in a loop, with its hands on tools. The chapter's thesis holds: fluency is the costume, pattern-matching is the creature wearing it. But a creature that can only talk and a creature that can act on your bank account are different propositions even when the machinery is identical, and most of the new risk lives in that difference, not in any new intelligence.

🧌 GOBLIN CHECK — "Agentic AI" is the same goblin you already met, now holding your wallet. Before you let one act for you, ask the unglamorous question the brochure skips: when it makes a mistake — and it will — does the mistake land on a draft you can delete, or on your bank account? The demo always shows the booking that worked. Ask to see the one that didn't.

Through 2025 and into 2026, "agentic AI" became the industry's central pitch: agents that file your expenses, triage your inbox, write and deploy code, run customer service, execute trades. The capability is real and improving. So is the failure mode. An agent that confidently books the wrong flight, emails the wrong client, deletes the wrong file, or runs the wrong transaction does so at machine speed and machine scale, and the same fluency that makes it persuasive makes its mistakes hard to catch in the moment. The hallucination problem from Section Five does not disappear when you hand the system tools. It gets a budget and a keyboard.

The deeper change is to the question of accountability. When a model writes you a wrong answer, you are the last line of defence: you read it, you catch it, you decide. When an agent acts, that human checkpoint is exactly what's being removed on purpose, because removing it is the point of automation. That is not mainly a technical question; it is a liability question, and Canadian law is not yet built for it. If an AI agent acting for a Canadian business overcharges a customer, agrees to something it shouldn't, or discloses someone's personal information, who is responsible: the person who deployed it, the company that built the model, the vendor who wrapped it into a product, or nobody? PIPEDA, Quebec's Law 25, provincial consumer-protection law, and the common law of agency were written for human and corporate actors, not for software that improvises. AI for All (Chapter 5) does not provide an answer, and the post-AIDA governance vacuum (Chapter 17) means no Canadian statute is squarely on point. That gap is the part worth watching.

So when someone sells you an agent, the portable questions are the ones the rest of this guide trains. What can it actually do without a human approving the step, and what does it merely suggest? What tools does it hold, and what is the blast radius if it's wrong? Is there a logged, reviewable record of what it did? And when it fails, who is accountable, in writing, before you switch it on? "The agent handled it" is not an answer to "who is responsible." It is a way of making the question disappear.

Agents are where the abstract "AI safety" conversation becomes concrete and Canadian: not a distant superintelligence, but a piece of software with your credit card, your inbox, and no clear line of legal responsibility behind it. Chapter 16 returns to what agents mean for work; Chapter 17 and Chapter 20 return to the governance gap they expose.

---

What to hold onto

CHAPTER RECAP — you now have: - A working model of machine learning as industrial-scale pattern matching against training data, optimized against measurable objectives, with no underlying concepts — and the conceptual basis for reading both alarmist and dismissive AI accounts skeptically. - The technical fact that the model is the data, with the three direct implications: training data is at the centre of the copyright question, the size and content of training data drives the environmental cost of training, and Indigenous data sovereignty frameworks apply to training data by definition. - A working understanding of transformers: attention, layers, scale, and the open scientific question of why scale produces the qualitative shifts in capability it produces. - The three stages of building a deployed AI system — pre-training, fine-tuning, alignment — and the operational implications: a deployed product is not the same as a foundation model, alignment is not solved, alignment encodes political and cultural choices made by specific companies. - A clear-eyed account of what current AI can and cannot do, holding both the genuine capabilities and the genuine limits in view simultaneously rather than collapsing to either the magic or the nothing reading. - The agent turn — language models wired into loops that act through tools, why that is still pattern-matching rather than a new intelligence, and why "it can act on your behalf" opens a Canadian accountability gap that no current statute squarely fills. - A working orientation to the AGI conversation: AGI is not what current systems are, the timelines are genuinely contested among senior researchers, the political implications of AI as it currently exists are not contingent on the AGI question, and the Hinton/Bengio frontier-risk question is one of the live questions in Canadian AI politics (a claimed leaked-draft thread on it remains unverified, and the guide builds nothing on it).

The next chapter (Chapter 3) takes the technical foundation this chapter provides and goes deep on training data specifically: where it comes from, what's in it, the difference between "open" and "freely scraped," and the consequences of training data being the model. The Stanford FMTI training-data-opacity finding lands properly in Chapter 3, the Indigenous data sovereignty material from Chapter 1 gets its conceptual setup deepened in Chapter 3 before Chapter 9 takes it to its sovereignty application, and the Whaanga "AI as new colonizer" framing from the Indigenous Protocol and AI Position Paper gets its full treatment.

You can now read any technical claim about AI, whether corporate, government, academic, or advocacy, with enough working understanding to ask the right questions: What was the training data? What were the alignment choices? What are the genuine capabilities? What are the genuine limits? What is the deployment context? Until those questions get answered, any AI claim is incomplete, regardless of the direction it points.

---

Bias label for this chapter: educational synthesis with explicit naming of where technical claims connect to political stakes elsewhere in the manual. Author lean: pattern-matching framing of AI capability rather than emergent-intelligence framing; respect for the genuine technical work involved while skeptical of corporate framing that obscures limitations; sympathetic to the view that alignment is genuinely contested terrain rather than a solved problem; explicit that the AGI question is separable from immediate Canadian AI policy questions while important in its own right. Where contested technical claims appear (emergence, AGI timelines, alignment status), the contest is named and positions identified rather than collapsed to a single account.

Primary sources cited or relied on in this chapter: Vaswani et al., "Attention Is All You Need" (NeurIPS 2017); Stanford Center for Research on Foundation Models, 2025 Foundation Model Transparency Index; Lewis, Whaanga & Yolgörmez, "Abundant intelligences: placing AI within Indigenous knowledge frameworks," AI & Society 40(1):2141–2157, 2024; CBC leaked draft of AI for All strategy (Lopez Steven & McKenna, June 1, 2026); Canadian newspapers v. OpenAI (Ontario Superior Court CV-24-00732231-00CL); Goodfellow, Bengio & Courville, Deep Learning (2016) for foundational technical exposition; Christiano et al. RLHF foundational work. Detailed citations in the Sources appendix.

---