TL;DR: Narrow AI, a system built for one defined task, is the only kind of artificial intelligence in use today, while general AI and superintelligence stay hypothetical and only two of the four functionality classes have members.
- Every deployed system, from voice assistants to chess engines to chatbots, handles one scoped task and fails at its edge.
- A review of 95 AGI forecasts found them spaced 15 to 25 years ahead of the year they were made, a pattern repeating since 1950.
- The functionality axis holds two live classes, reactive machines and limited memory, and two empty ones: theory of mind and self-aware AI.
- AlphaGo learned from roughly 30 million recorded human moves, so it files under limited memory, a correction most taxonomy charts still miss.
The honest label for every AI product you opened this week is narrow AI. Voice assistants, ranking feeds, chess engines, autocorrect, chatbots: each was built around one scoped job, and each breaks past that job’s edge. Every few weeks a headline announces that general intelligence, or superintelligence, has arrived; the verified record says nothing of the kind has shipped. Watch the full episode on YouTube, or read on: the same boards, every date and number tied to its source.
The episode this walkthrough accompanies: sixteen boards in just over eight minutes, every claim tied to a linked source.
The opening board: the label over every product you have used.
Narrow AI is the only kind that shipped
Narrow AI, the class also called weak AI, covers systems designed for a single defined task: they perform it, and they fail outside it. Look at your own devices. Siri and Alexa take voice commands. A chess engine plays chess and nothing else. Face matching sorts photos, recommenders rank your next watch, dictation and autocorrect handle your typing. Every system in production today sits inside this class, chatbots included.
Ask a chess engine to draft an email and it has nothing to offer. That blankness outside the task is what the next two capability levels promise to remove. The standard reference draws the line exactly there, and nothing deployed crosses it.
| Capability level | What it would mean | Status |
|---|---|---|
| Narrow AI | One defined task, performed well, failure beyond the task | Everything deployed today |
| General AI (AGI) | Human-level performance across essentially all mental tasks, skills carried between domains | Hypothetical; never built |
| Superintelligence | Performance beyond the most capable humans across essentially every domain | Debated; no schedule |
Two axes, one common mixup
Capability is one axis, running from narrow AI up through general and super. Functionality is a second, independent axis that sorts systems by how they work rather than how much they can do. The four functionality classes come from the computer scientist Arend Hintze, who laid them out in 2016: the two that exist, reactive machines and limited memory, and the two that do not, theory of mind and self-aware AI. His article is the one most explainer charts borrow without naming.
The mixup that feeds most hype collapses the two axes into a single ladder: a chatbot that will discuss any topic gets called one step below AGI, when what it can do and how it does it are separate questions.
AGI: a definition with a sixty-year miss record
General AI, or AGI, names a hypothetical system that would match or beat human performance across essentially the full range of mental tasks, carry what it learns in one domain into another, and take on problems it has never met. The reference article is careful about the status: nothing built matches the definition. The name is younger than the idea, coined by Mark Gubrud in 1997 and brought into wide use by Shane Legg and Ben Goertzel around 2002, the era the AGI subfield was founded.
The prediction record reinforces that footnote: 95 forecasts reviewed by Stuart Armstrong and Kaj Sotala, spanning 1950 to 2012, sit 15 to 25 years ahead of the year each was made, a bias running from Herbert Simon’s 1965 estimate through forecasts as late as 2012.
Herbert Simon wrote in 1965 that “machines will be capable, within twenty years, of doing any work a man can do.” Marvin Minsky expected the problem substantially solved within a generation, writing in 1967. Stuart Armstrong and Kaj Sotala reviewed 95 forecasts made between 1950 and 2012 and found a strong bias toward putting human-level AI 15 to 25 years ahead of the year each forecast was made. Expert polls in 2012 and 2013 put the median arrival between 2040 and 2050, the mean at 2081, and 16.5 percent answered never. In 2024, Geoffrey Hinton gave a range of five to twenty years; in 2023, Demis Hassabis said within a decade. Hubert Dreyfus and Roger Penrose have spent careers arguing the goal is unreachable. The spacing between a forecast and its promised arrival has never closed.What is the track record of AGI predictions?
Board 5: the AGI definition, two bullets and the young label.
Superintelligence: an argument, with dates attached
Superintelligence clears a higher bar than AGI. Nick Bostrom’s formulation asks for a system whose performance “greatly exceeds the cognitive performance of humans in virtually all domains of interest.” Can such a system exist? The question remains an argument, and only one side carries dates.
When researchers were surveyed in 2022, the median year given for high-level machine intelligence was 2061. The leaders of OpenAI gave it less than ten years, speaking in 2023. David Chalmers holds that it is feasible, since the brain is a mechanical system and can in principle be emulated and amplified. Russell and Norvig, writing in their textbook, note that improvement in a particular area “tends to follow an S curve,” a steep climb that flattens into a plateau.
Strong AI is the one phrase in this neighborhood that carries two meanings. The philosopher John Searle coined it in 1980, in his Chinese room argument, to mean a machine that genuinely has a mind, and he has argued against that possibility ever since. In engineering conversation the same two words mean nothing more exotic than AGI-level capability, the level at which a system handles any intellectual task a person can. The two readings are both in circulation, and headlines swap them freely, which is how a capability claim turns into a consciousness claim. When a headline uses strong AI, decide which sense is on the page before you share it.What does the phrase strong AI mean?
Hintze’s four classes: two exist, two do not
On the functionality axis the question changes from how much a system can do to how it works. Hintze’s four classes answer it, and the scoreboard is lopsided: the lower two hold every deployed system, the upper two hold nothing.
| Functionality class | How it works | Members |
|---|---|---|
| Reactive machines | No memory, no learning; identical input always returns the same preset answer | Deep Blue; no famous modern member |
| Limited memory | Learns from stored past data, keeps no long-term memory of its own | Self-driving perception, chatbots, LLMs |
| Theory of mind | Would model what others believe, intend, and feel | None built |
| Self-aware AI | Would have a mind of its own | None built; treated as far off |
Reactive machines are the floor. In May 1997, Deep Blue beat Garry Kasparov by brute force, weighing roughly 200 million chess positions every second against an evaluation function tuned with grandmaster Joel Benjamin. It learned nothing while playing; its developers were permitted to adjust it between games, which is humans doing the adapting. The match record carries the details, including the 3.5 to 2.5 score.
Then the correction most taxonomy charts get wrong: AlphaGo does not belong in the reactive row. AlphaGo’s 12-layer networks started by imitating human experts, learning from a database of around 30 million recorded moves, and were then sharpened through reinforcement learning as the program played long series of games against other copies of itself, until, in the AlphaGo record’s words, it was “free to learn the game for itself.” A system that learns sits one class up, in limited memory, and the reactive class is left without a famous modern member.
Limited memory is where the narrow AI you use every day sits. A self-driving car reads the road from its recent sensor history. A chatbot holds its last few turns. An LLM works inside a finite context window, and everything outside that window is gone. The LLM article spells out what a bare model cannot do: operate in a changing environment, remember its own earlier turns, or plan ahead.
The upper two classes are empty: nothing built models another agent’s beliefs, and nothing built has a mind of its own. What gets marketed as theory-of-mind AI is affective computing, which Rosalind Picard launched with a 1995 paper and a 1997 book; the deep dive below covers why its emotion labels stop short of modeling minds.
The misfiling comes from chart inheritance: AlphaGo plays a board game, Deep Blue played a board game, so the two get glued into one row. The training record says otherwise. AlphaGo’s 12-layer convolutional networks began by imitating human experts, learning from a database of roughly 30 million recorded moves, and were then set against other copies of themselves in long runs of self-play, with reinforcement learning doing the sharpening. That is learning from stored past data, which is the working definition of the limited-memory class. The AlphaGo record adds a detail that widens the gap further: its algorithms are seen as potentially more general-purpose than Deep Blue’s, which was built for one board and one job.Why is AlphaGo filed under limited memory?
It is marketed as one, and the two are different things. Affective computing, the field Rosalind Picard opened in 1995 with a paper and followed with a 1997 book, classifies patterns taken from faces, voice, and body signals. Its output is a label: stressed, engaged, happy. A label of that kind is pattern classification, and the field’s own reference is careful to say the labels may or may not match what the person feels. Theory of mind is the modeling of what another agent believes, intends, and wants, and no built system does it: the theory-of-mind literature notes that AI “largely lacks the capacity to comprehend and predict human behavior influenced by social and emotional factors.” On self-awareness, Russell and Norvig set the question aside as beyond what the field can currently take on. The distance between reading a face and modeling a mind is the entire distance between the second and third functionality classes, and products cross it only in the marketing copy.Is emotion recognition a step toward theory of mind?
Where chatbots sit on both axes
Whether today’s chatbots already count as proto-AGI is the loudest open argument in the field, and the positions are easy to state. Microsoft researchers, writing in 2023, said GPT-4 “could reasonably be viewed as an early (yet still incomplete) version” of an AGI system. A framework DeepMind published the same year sorts chatbot-class models as “instances of emerging AGI.” The stochastic-parrots camp answers that these systems merely remix and recombine existing writing.
The systems are easier to place than the argument suggests. On capability, an LLM is narrow AI: one job, language in and language out. On functionality, it is limited memory: a finite context window, no recall of past behavior, no planning of future actions. Both placements can hold while the argument continues, because capability and functionality answer different questions.
The vocabulary the rest of the series runs on
The episode closes with a glossary because the series leans on it, and the nesting matters most. Artificial intelligence is the umbrella. Machine learning sits inside it: systems that learn from data instead of following hand-written rules. Deep learning sits inside machine learning: neural networks stacked in layers, three to thousands, with useful features learned from the data rather than a designer’s checklist. The glossary holds all of these.
Since 2017, LLMs have been transformer-based, with a billion-plus parameters, running from the 2018 arrivals, BERT and GPT-1, to ChatGPT’s public release in late 2022. Generative AI produces new content from learned patterns, and transfer learning carries a model trained on one task over to a neighboring one.
Every mention of training in this series falls into one of four paradigms. Supervised learning uses labeled examples, an input with the right answer attached. Unsupervised learning takes unlabeled data and looks for structure in it. Semi-supervised learning mixes a small labeled set with a large unlabeled one. Reinforcement learning puts a system inside an environment and rewards it for the cumulative reward it earns there. A single product can use several paradigms; the LLM training pipeline pairs a supervised stage with a reinforcement stage.How do the four learning paradigms differ?
The glossary distinguishes four artifacts that casual coverage blurs together. An algorithm is an unambiguous specification for solving a problem. Training data is the set of examples the algorithm learns from. A model is the learned representation that comes out of training. An agent is an entity that perceives its environment through sensors and acts on it toward a goal. The pipeline the series draws is a straight line through them: an algorithm plus data, run through training, yields a model, and a model wrapped in sensors and actuators becomes an agent.What separates an algorithm, a model, and an agent?
Two failure modes explain a fair share of machine-learning headlines. Overfitting is when a model memorizes its training set and then fails on data it has never seen. Underfitting is the opposite failure, a model too simple to capture the trend even in the data it was trained on. A useful discipline when you read a performance claim: ask which set the number came from, the training set or a held-out set. The gap between the two is where overfitting hides.What do overfitting and underfitting mean?
Four checks before you repeat a claim
The gotchas board carries three traps and a standing date; they compress into four checks for any AI headline.
- A taxonomy chart files AlphaGo as reactive: the chart was copied from an older tree. AlphaGo trained on about 30 million human moves, which is learning, which files under limited memory.
- A product claims theory-of-mind AI: read what it labels. Emotion recognition is pattern classification, and its labels are provisional by the field’s own description.
- A headline uses strong AI: decide which sense is on the page, Searle’s conscious machine or an engineer’s shorthand for capability.
- A forecast names an AGI date: measure the distance. The standing guess since 1950 has been arrival fifteen to twenty-five years away, a spacing that has held for over seventy-five years.
Board 13: three traps and the standing date, the four checks in board form.
Board 16: the recap, four checks and the sticky note.
Keep the two axes separate and the field reads cleanly. On capability, narrow AI is all that runs today, AGI is a definition whose arrival dates have been missed for sixty years, and superintelligence carries a 2061 survey median and no schedule. On functionality, reactive machines and limited memory have members; theory of mind and self-aware AI do not.
Continue learning
- What Artificial Intelligence Actually Is: the series opener, the working definition of AI and the pipeline from data to model.
- Machine Learning: The Honest First Lesson: the training loop, and the two ways systems learn from data.
- AI History: Booms, Winters, and Revivals: the two booms and two winters behind every forecast you will read.
Disclosure: this article was drafted with AI from the script of my video lesson, then checked against its sources before publishing.
This lesson first appeared on iqraa.tech.
More whiteboard lessons on the Whiteboard Architect YouTube channel.
Top comments (1)
Hmm lol - most humans are also "really good" at just one thing - being a language virtuoso, and a chess master, and an ace programmer at the same time is very very rare ...
However, given the fact that an LLM can be trained on more than one topic (knowledge area), isn't "general intelligence" (AGI ?) just a matter of scaling up the capability of the software/hardware (more GPUs, more memory) and then training the thing on every area of expertise you can throw at it? And then it's not 'narrow' anymore ...
Which sounds pretty mundane and not all that exciting, if that's all there is to 'AGI' ...
I have a different take - a fundamental difference between AI and "biological intelligence" (human or animal brains) is that AI is solely based on 'symbolic' reasoning (which translates neatly to tokens), while humans/animals have what I'd call 'intuitive' intelligence as well - you could call it "digital" (symbolic, 'tokens') versus "analog" intelligence.
As a result, even the simplest animal is much better at sensing/navigating a physical environment, with its intuitive/analog (non-symbolic) intelligence ...
Put differently, AI tends to be really good at "difficult" tasks (math, chess, programming), but very bad at "easy" things, like guiding a robot to traverse your living room, while avoiding to stumble over every piece of furniture.
Modeling the physical/analog world (3 D vision with depth, motion, etc) through 'symbolic'/digital logic (and hence with LLMs) is possible, but pretty hard, and not that efficient ...
Which is the reason why (as it stands) AI is more a threat to "white collar" jobs than to "blue collar" jobs :-)
But incorporating 'intuitive' intelligence would require totally different concepts - and you can even wonder if we'd want it, because AI with that kind of intelligence sounds a bit too much similar to ourselves, and too much equipped to replace us!