Then and Now: AI Predictions that Were Hype

1. “AGI is just a couple years away”

Then (2022–2024)
Leaders like Sam Altman and Dario Amodei suggested that artificial general intelligence could arrive in the near term — often framed in interviews and talks between late 2022 and early 2024 as “a few years,” not decades.

Now (2026)
We have systems that are impressive pattern matchers, copilots, and reasoning approximators. But they still:

• hallucinate confidently
• struggle with long-term planning
• break outside narrow domains

AGI didn’t arrive. What we got instead is powerful narrow intelligence with wide applicability.

Reality check
The slope was real. The extrapolation wasn’t.

2. “AI will replace most knowledge workers quickly”

Then (early–mid 2023)
Across 2023 — especially after the release of ChatGPT (Nov 2022) and GPT-4 (March 2023) — executives and reports warned that lawyers, programmers, marketers, and analysts could be heavily displaced within a short window (often within 1–5 years).

Now (2026)
What actually happened:

• Demand for engineers increased (to integrate AI)
• Knowledge work became augmented, not replaced
• Output expectations went up, not headcount down

Companies didn’t remove humans. They raised the bar for what one human is expected to do.

Reality check
AI didn’t eliminate jobs. It compressed them.

3. “Prompting will replace programming”

Then (mid–late 2023)
During the “prompt engineering” wave of 2023, there was widespread belief that natural language would become the primary interface — that “you won’t need to code anymore.”

Now (2026)
Prompting matters. But:

• Complex systems still require architecture
• Debugging still requires technical depth
• Reliability still requires engineering discipline

What emerged is not the end of programming, but a shift toward:

• specifying intent
• while still needing systems thinking to make it real

Reality check
Prompting didn’t replace programming. It expanded the surface area of responsibility.

4. “Models will become reliable enough to trust by default”

Then (2023–early 2024)
There was an implicit assumption — especially during rapid scaling from GPT-3.5 (2022) to GPT-4 and beyond — that increasing model size and alignment work through 2023–2024 would naturally lead to reliable systems.

Now (2026)
We got:

• better outputs
• smoother UX
• fewer obvious errors

But reliability remains conditional:

• depends on prompt design
• varies by domain
• degrades over long chains of reasoning

In high-stakes environments, humans are still the control layer.

Reality check
Performance improved. Trust did not scale linearly with it.

5. “AI will create massive, immediate economic transformation”

Then (2023)
Following the explosion of generative AI in 2023, forecasts pointed to rapid GDP impact, sweeping disruption, and near-term productivity gains across industries.

Now (2026)
The impact is real — but uneven:

• Early adopters benefit disproportionately
• Many companies struggle to integrate AI meaningfully
• Productivity gains are localized, not systemic

The biggest gains show up where:

• workflows are redesigned
• incentives are aligned
• systems are rebuilt around AI

Not where AI is simply “added.”

Reality check
AI didn’t transform the economy overnight.
It rewarded the few who changed how they operate.

The Pattern Behind the Misses

These weren’t random errors. They share a structure:

• Linear thinking applied to non-linear systems
• Capability mistaken for deployability
• Demos confused with production reality
• Technology change assumed to bypass human systems

The models improved faster than expected.
The systems around them did not.

The Real Hype Index

If you had to score these predictions:

• Direction: mostly correct
• Magnitude: somewhat inflated
• Timing: consistently early

That’s the AI Hype Index:
Right about where. Wrong about when. Blind to everything in between.

What to Do With This

The opportunity isn’t in believing or dismissing AI.
It’s in understanding the gap between:

• what the technology can do
• and what organizations can actually absorb

That gap is where most of the value — and most of the failure — lives.

The companies that win aren’t the ones that chase the latest model.
They’re the ones that:

• redesign workflows
• reduce variance
• and treat AI as part of a system, not a feature

The next wave of predictions will sound just as confident.
This time, you know where to look:
not at the demo,
but at the system that has to carry it.