Artificial Intelligence — A Field Guide

What & why

Artificial intelligence is the project of building machines that do things we call intelligent when people do them — perceive, reason, decide, converse, create. It is one idea wearing many bodies. For most of its history AI meant hand-written rules; today it almost always means machine learning, where a system is not programmed with the answer but shown examples and left to find the pattern itself. The version that reached everyone at once — the chatbot that writes essays and code — is a narrow, recent slice of a field that is seventy years old and has died and resurrected at least twice.

Two reasons to understand it rather than just use it. First, it is quietly rewriting what "knowledge work" means, and the difference between people who grasp what these systems actually do and people who treat them as magic is about to matter a great deal. Second, the field is in one of the most contested moments in its history — reasonable, informed experts disagree about whether today's systems understand anything, whether they are a path to human-level intelligence or a spectacular dead end, and whether they are safe. A field guide's job is to give you the map firm enough to follow the argument, and honest enough to show where the map runs out.

The long view: three waves

AI has moved in waves — a burst of optimism, a wall, a "winter" of lost funding, then a new idea that breaks through.

Symbolic AI (1950s–1980s). The founding bet, made at the 1956 Dartmouth workshop that named the field: intelligence is symbol manipulation, and thought is a kind of logic that can be written down as rules. This gave us chess programs, theorem provers, and "expert systems" that encoded a specialist's know-how as thousands of if-then rules. It hit two walls: the rules never covered the messy real world (the "common sense" problem), and hand-coding knowledge didn't scale. Funding collapsed — the first AI winter.

Machine learning (1990s–2010s). The shift: stop writing the rules, learn them from data. Statistical methods, and especially neural networks — loosely brain-inspired layers of simple units whose connection strengths are tuned by exposure to examples. The enabling trick, backpropagation, was known in the 1980s but only became powerful when three things arrived together in the early 2010s: big labeled datasets, cheap parallel computation (GPUs), and a few architectural insights. The 2012 result where a deep network crushed the field on image recognition is usually marked as the moment "deep learning" won.

Foundation models (2017–now). In 2017 a new architecture, the Transformer, made it practical to train enormous models on essentially all the text on the internet. The surprise wasn't the architecture itself but what happened at scale: models trained only to predict the next word turned out to acquire — as a side effect — translation, summarization, reasoning-like behavior, and code. Nobody explicitly taught these skills. This is the era of "foundation models" and large language models (LLMs); the public met it as ChatGPT in late 2022.

How today's systems actually work

Strip the mystique and a large language model is a next-token predictor. Trained on a vast corpus, it learns the conditional probability of the next chunk of text given everything so far. Generation is just doing that repeatedly. Three ideas carry almost all the weight:

Representation by embedding. Words, images, sounds get turned into long lists of numbers (vectors) positioned so that similar things sit near each other. Meaning becomes geometry. This is why a model can answer about things it never saw verbatim — it operates on the neighborhood, not the exact string.

The Transformer and attention. The core mechanism, "attention," lets the model weigh how much every part of the input should influence every other part, all at once. This is what let training parallelize across thousands of chips and what gives the model its grip on long-range context. The founding paper is Attention Is All You Need (Vaswani et al., 2017).

Scale, and the training pipeline. Capability rose with size in a way regular enough to be called scaling laws: more data, more parameters, more compute buys predictably lower error. Modern models are then aligned in a second stage — fine-tuned on human-rated examples (RLHF, reinforcement learning from human feedback) so the raw predictor becomes a helpful, less-toxic assistant. The base model knows the internet; alignment teaches it manners.

A crucial consequence: the model has no separate store of "facts" it looks up. Everything is baked into the weights as statistical association. That is the same fact that explains both its fluency and its hallucinations — confident, well-formed statements that are simply false. It is not lying and not retrieving; it is generating what plausibly comes next, and plausible is not the same as true.

The frontier: from chatbots to agents

The 2023–2026 shift is from models that answer to systems that act. An agent wraps a language model in a loop: it can call tools (search, code execution, other software), observe results, and take further steps toward a goal. Multi-agent systems have several such loops cooperating or checking each other — the pattern our library's AutoGen paper helped popularize. This is where the practical action now is: coding assistants that run and debug their own output, research agents that read and synthesize, operators that click through real software.

It is also where the risks concentrate. A model that only talks is bounded by the person reading it; an agent that acts has a real blast radius. The security literature is racing to keep up — see the library's AgenticRed paper on adversarially testing agentic systems. Reliability, not raw capability, is the current bottleneck: a system that is brilliant 95% of the time and confidently wrong 5% of the time is hard to hand real authority.

Contested ground

This is where an honest guide earns its keep. These disputes are live among serious researchers — not settled questions dressed as debate.

Do these systems understand anything? One camp: meaning is use, and a system that wields language this fluently has, functionally, some understanding — the rest is goalpost-moving. The other: LLMs are "stochastic parrots," statistically remixing form with no grip on meaning or truth, and their fluency fools us (Searle's old Chinese Room argument, updated). The honest position is that we lack an agreed test — and that the question exposes how shaky our definition of "understanding" was to begin with.

Is scale the path to general intelligence, or a plateau? Scaling optimists point to the smooth, still-unbroken capability curve and argue that human-level AI is a matter of more compute and a few years. Skeptics argue that next-token prediction, however scaled, lacks something essential — grounding, causal models, genuine reasoning — and that current gains are the flat top of an S-curve we haven't recognized yet. Both cite real evidence; the disagreement is about extrapolation, which no data settles in advance.

How dangerous is it, and how soon? The safety spectrum runs from "existential risk from misaligned superintelligence deserves serious present attention" (the alignment-research position) through "the real harms are here and mundane — bias, labor displacement, misinformation, concentration of power" (the AI-ethics position) to "the risk talk is hype that serves incumbents by inviting regulatory moats." These are partly empirical and partly about values; they are not reconcilable by a single fact, and treating any one as obviously correct is a tell.

Whose values, whose data? Alignment aligns a model to someone. Training data was scraped from a public web that never consented, encoding its biases and its copyright. Who decides what a global system should refuse, believe, or promote is an unsolved governance problem wearing a technical costume.

Where this connects

AI is not a sealed domain — its hardest questions lean on older ones. "Does it understand?" is epistemology and philosophy of mind with new stakes. "Could it be conscious?" reaches into the study of consciousness — one of the deepest open problems in any field, and one where the honest answer to "is this system aware?" is that we cannot currently tell, because we do not have a working theory of awareness to test against. The decision theory under "how should an agent act under uncertainty?" traces back to Bayes and to game theory. A serious map of AI is a map of half of philosophy with GPUs attached.

Where to go deeper

Foundational papers (all open access — read them at the source):

Standard references (the field's canonical books — obtain via library or purchase):

  • Stuart Russell & Peter Norvig, Artificial Intelligence: A Modern Approach — the definitive survey, symbolic through modern.
  • Ian Goodfellow, Yoshua Bengio & Aaron Courville, Deep Learning — the deep-learning foundations (authors' full text is readable free at deeplearningbook.org).
  • Stuart Russell, Human Compatible — the alignment problem, argued by a founder of the field.
  • Melanie Mitchell, Artificial Intelligence: A Guide for Thinking Humans — the best skeptical-but-fair popular account.

Sources

  • Alan Turing, "Computing Machinery and Intelligence," Mind (1950) — doi:10.1093/mind/LIX.236.433
  • Vaswani et al., "Attention Is All You Need" (2017) — arXiv:1706.03762
  • Brown et al., "Language Models Are Few-Shot Learners" (2020) — arXiv:2005.14165
  • Wu et al., "AutoGen" (2026) — read on rho.md
  • "AgenticRed: Evolving Agentic Systems for Red-Teaming" (2026) — read on rho.md
  • Russell & Norvig, Artificial Intelligence: A Modern Approach (4th ed., 2020)
  • Goodfellow, Bengio & Courville, Deep Learning (2016) — full text free at deeplearningbook.org
  • Stuart Russell, Human Compatible (2019)
  • Melanie Mitchell, Artificial Intelligence: A Guide for Thinking Humans (2019)
  • Emily Bender et al., "On the Dangers of Stochastic Parrots," FAccT (2021) — doi:10.1145/3442188.3445922
  • John Searle, "Minds, Brains, and Programs," Behavioral and Brain Sciences (1980) — the Chinese Room argument — doi:10.1017/S0140525X00005756
  • Kaplan et al., "Scaling Laws for Neural Language Models" (2020) — arXiv:2001.08361

Part of the Knowledge Atlas field guides — one guide per field, sources linked, disagreements kept.

This field guide is original synthesis by Knowledge Atlas, licensed under CC BY 4.0. Sources are credited above; quoted passages remain under their own terms.

Open in Rho MD →