AMUI Journal AI & product development10 min read · Interactive
After the AI hype
If the AI story is exaggerated, why does the software keep getting more useful?
Richard Campbell’s 2026 keynote becomes much easier to understand once you stop treating “AI” as one thing. There is the capability of the software. And there is the story we tell about what that capability means.
An AMUI interpretation of Richard Campbell’s NDC Copenhagen talk. The curves, percentages, and simulations are illustrative teaching devices, not measured capability scores or financial forecasts.
Explorable essay based on Campbell’s NDC Copenhagen keynote
Move only the expectation.The software stays exactly the same.
moretime →
The gap is large enough that disappointment becomes almost inevitable.
A bubble and a breakthrough can happen at the same time. That sounds contradictory only if we assume the market story and the underlying technology must rise and fall together.
Campbell’s talk is, on the surface, a history of “AI”: ELIZA, deep learning, ImageNet, language models, data centers, coding agents, AlphaFold. But the more useful thread is about misidentification. Humans keep mistaking one thing for another.
We mistake fluent language for a mind. We mistake an impressive benchmark for a general intelligence. We mistake investment for evidence. And then, just as carelessly, we can make the opposite error: once the story becomes ridiculous, we assume the technology underneath it must also be ridiculous.
So what are we actually looking at when software looks intelligent?
words back at you
capability
story in our head
Very little machinery can produce a surprisingly large feeling of being understood.
1960s
A program barely has to understand you for you to feel understood.
Joseph Weizenbaum’s ELIZA used simple rules to turn a person’s words back into questions. The mechanism was primitive. The reaction was not. People could become absorbed in the exchange because language itself is a powerful social cue.
That is the first trap: the human mind supplies missing intelligence. We are not passive observers of the interface. We finish the illusion.
Later called the “ELIZA effect”: attributing more understanding to a program than its mechanism warrants.
patterns from data
capability
story in our head
Then something important changes: the mechanism begins doing work that was genuinely hard for software.
2012
Then the software really did get better.
The ImageNet breakthrough was not a better illusion of conversation. Deep neural networks, large labelled datasets and GPU compute sharply improved machine vision. Tasks that had been stubborn became practical.
This matters because a skeptical reaction can now become just as lazy as a credulous one. “It is only software” does not mean “it is unimpressive software.”
predict what comes next
capability
story in our head
Language brings the old human vulnerability back, now attached to a much more capable machine.
2020s
Language models combine both curves.
They are genuinely capable pattern machines. They can transform text, write useful code, summarize, classify, translate and operate inside larger workflows. But they also speak through the interface that humans are most likely to anthropomorphize: language.
So capability rises and the old ELIZA mistake gets supercharged. A useful system can be described as a coworker, a companion, an oracle, or an approaching species. Those are different claims.
software inside a story
capability
story in our head
A real capability becomes the seed of a claim whose scale is no longer constrained by the mechanism.
The hype cycle
Money enters through the gap.
Campbell’s sharpest economic point is simple: hype is easiest when the buyer cannot independently judge the mechanism. A story like “all jobs disappear in two years” can attract more attention than a narrower claim like “this cuts the cost of one workflow by 35%.”
The first is grander. The second is testable. Markets, alas, have occasionally shown a certain fondness for grander.
Why a scripted reply can feel personal.
Choose one of the example sentences below, then notice how the reply sounds attentive. This ELIZA-inspired demonstration has just three fixed replies. It does not interpret your situation; the conversational wording invites you to supply that understanding.
Choose an example below to see its scripted reply.
Select a sentence to try it. Each button returns a fixed reply; this demonstration uses no AI model.
Modern language models are far more capable than this demonstration. They learn statistical structure from training data and can perform useful tasks. The distinction still matters: a reply that feels understanding and a result that survives verification are different kinds of evidence.
A smaller mental model
Before “intelligence,” notice the prediction.
A language model repeatedly estimates what token should come next. Real systems are vastly more complex than this toy, but the toy keeps one important fact visible: fluent output can emerge from prediction without requiring a little person inside the machine.
The cat sat on the …
Pick the continuation you expect the model to favor.
The same distinction applies to investment: a useful technology can attract expectations far beyond what it has delivered. Adjust the two sliders to explore that gap.
A conceptual simulation
Make the story outrun delivery.
The numbers here are illustrative, not market data. They show a mechanism: expectations can grow faster than practical value, especially when infrastructure spending and copycat investment amplify the story.
early promiseexpectation meets reality
The larger the gap, the harsher the correction can be. The useful curve does not have to return to zero.
Campbell reaches for the dot-com boom because it preserves both halves of the idea. Investors funded absurd internet businesses. Many collapsed. The internet did not.
The correction removed companies, valuations and fantasies. It did not un-invent browsers, networks, online distribution or the economics that later made enormous businesses possible.
So if today’s AI investment boom corrects, the important question is not whether every current company survives. It is which capabilities remain useful when they no longer get to borrow credibility from the word “AI.”
The part worth keeping
Useful does not mean human-like.
Campbell’s most persuasive positive example is AlphaFold. Protein-structure prediction is valuable precisely because it does not need to imitate a person. It attacks a hard scientific problem and can be judged against external reality.
Coding agents belong in the same conversation when they are embedded in workflows with tests, pull requests and human review. The evidence is not that the software sounds competent. The evidence is that the work survives verification.
more externally verifiable usefulness →more human-like appearance →ELIZAAlphaFoldcoding agent + testsgrand AI claim
The safer direction is toward work whose success can be checked outside the conversation.
Tap each system. Two dimensions stay separate: how human-like it feels, and how easily useful work can be checked outside the conversation.
ELIZA feels social, but there is little externally verifiable work underneath the conversation.
This is also why Campbell keeps pushing against anthropomorphic language. The vocabulary is not cosmetic. If you call a system a friend, an employee, a mind or an autonomous agent, you silently import expectations about memory, intention, responsibility and judgment.
Campbell goes further than this essay on chatbot harms, regulation, data-center investment and AGI claims. Those topics are real parts of the talk, but they are not all needed to understand its central mechanism.
The bubble can burst. The capability can stay.
The opening question looked like a choice: either the AI story is exaggerated, or the technology is useful. Campbell’s history makes that choice unnecessary.
A market correction can reset expectations without erasing useful capabilities. What remains worth using depends on the work the software can reliably do.
The durable habit is simple: describe the mechanism, test the outcome, and let the story come last.
Putting it into practice
What this means for your next product.
Start with a specific workflow, a measurable result, and a way to catch failure. Compare an AI feature with a simpler alternative. A convincing demo is a reason to test further; reliable results are the reason to keep using it.
A hypothetical support assistant: a friendly answer about refunds is easy to admire. A useful test asks whether it applies the correct policy, checks the actual order, and escalates cases it cannot resolve. Compare those results with a searchable FAQ or a simple form. For AMUI, that is the product-design question: which approach helps the user finish the task, and how will we know?
Campbell’s historical arc from ELIZA through deep learning, LLMs, hype cycles, coding agents and AlphaFold.
His argument that anthropomorphism distorts how people judge language software.
His dot-com analogy: speculative excess can collapse while useful infrastructure and capabilities remain.
His closing emphasis on solving human problems rather than treating code, or “AI,” as the point in itself.
Teaching devices & source fidelity
The two-curve model, ELIZA mini-chat, next-word toy, hype sliders and usefulness map are teaching devices created for this essay.
The simulation numbers are illustrative. They are not measurements of AI capability, investment or market expectations.
The accessible Habr transcript is a freely restructured translation. Its translator also flags and corrects several factual slips in the spoken keynote. This essay preserves Campbell’s argument while avoiding disputed details where they are unnecessary.
