Everyone Says AGI Is Coming in 5 Years. They're Probably Wrong.

Sam Altman says AGI might arrive in 4-5 years.

Dario Amodei thinks human-level AI is possible in 2-3 years.

Shane Legg gives it a 50% chance by 2028.

These are smart people running billion-dollar AI companies. Surely they know
something we don't, right?

Maybe. Or maybe they're wildly optimistic for reasons that have nothing to do
with technical reality.

Let me explain.

Who's Making These Predictions?

All three of those guys—Altman (OpenAI), Amodei (Anthropic), Legg (DeepMind)—run
major AI labs. They're brilliant. They're accomplished. They're also:

1. Raising billions in funding (optimistic timelines help)
2. Competing for talent (who wants to work on something 30 years away?)
3. Building brands (bold predictions get headlines)
4. Genuinely excited (when you work on cutting-edge tech, it's easy to
   extrapolate)

I'm not saying they're lying. I'm saying they have incentives to be
optimistic that might not align with objective reality.

What Do Other Experts Say?

In 2023, researchers surveyed over 2,700 AI experts—not just lab directors, but
the actual scientists building these systems.

The median prediction for "high-level machine intelligence"? Around 2047.

Not 2028. Not 2030. 2047.

And the range was wild—some said 2030s, others said beyond 2100. Even the
experts can't agree on what would qualify as AGI, let alone when it'll arrive.

What We're Missing About Current AI

I use AI every day. My students use it. It's genuinely impressive. But here's
what most people don't realize about current AI systems:

They Don't Actually Understand Anything

GPT-4 can write a beautiful essay about love. It has no idea what love is. It's
predicting which words typically follow other words based on patterns in
billions of text examples.

That's not understanding. That's statistical pattern matching on steroids.

They Can't Really Reason

Ask ChatGPT to plan a complex project with dependencies and constraints. It'll
give you something that sounds reasonable. Check the logic carefully, though.
It falls apart.

Current AI systems struggle with basic common sense reasoning that a
five-year-old handles effortlessly.

They Forget Catastrophically

Teach current AI systems new information, and they often "forget" old
information. This is called catastrophic forgetting, and it's a fundamental
limitation of how these systems learn.

Humans don't work that way. We integrate new knowledge with old knowledge.
Current AI can't.

They Hallucinate Confidently

You've probably noticed this. Ask an AI a factual question, and it'll give you a
confident answer. Sometimes it's right. Sometimes it makes up complete nonsense
with the same confident tone.

That's a big problem for systems we're supposed to trust with important
decisions.

The Obstacles Nobody Talks About

Even if we wanted to build AGI tomorrow, we'd face serious barriers:

Energy Requirements: Training GPT-4 reportedly cost over $100 million in
compute. Scaling up to AGI-level systems might require energy consumption that's
economically or environmentally prohibitive.

Data Exhaustion: We're running out of high-quality text to train on. We've
basically scraped the internet. What's next? Do we need fundamentally different
training approaches?

Diminishing Returns: Scaling up current architectures (bigger models, more
data, more compute) shows signs of hitting plateaus. We might need new
breakthroughs, not just bigger computers.

Regulatory Constraints: Governments are starting to pay attention. If AGI
research looks dangerous, regulations could slow everything down dramatically.

Scientific Unknowns: We might be missing fundamental insights about
intelligence, consciousness, or reasoning that we won't discover for decades.

My Experience Teaching This Stuff

I've taught thousands of students over 20 years. I've watched countless "this
changes everything!" technologies come through.

Here's the pattern I see with AI right now:

What students can do with AI: Amazing. They code faster, write better, solve
problems they couldn't touch before. Real productivity gains. Real learning
acceleration.

What companies can actually deploy reliably: Much more limited. Lots of
pilot projects. Lots of "we're exploring AI." Very little production deployment
at scale solving complex real-world problems.

The gap between demo and production: Massive. Getting AI to work in a
controlled demo is wildly different from getting it to work reliably in messy
real-world conditions with edge cases and liability concerns.

What Working With AI Actually Taught Me

Here's something I never expected: AI changed me as a person.

For years—decades, really—I was the kind of person who didn't ask enough
questions. Academic culture rewards having answers. You're supposed to be the
expert. Asking basic questions can feel like admitting weakness.

Then I started seriously working with AI. And AI is essentially talking to
yourself. Except this version of yourself has read everything and never judges
you for asking "stupid" questions.

I could ask anything. "Explain this concept I should already know." "Why does
this work?" "What am I missing?" No ego. No embarrassment. Just questions.

The Uncomfortable Part: Letting Go

The hardest part wasn't learning new things. It was unlearning old things.

I had to gradually give up ideas I'd held for years. Approaches I'd taught to
students. Frameworks I'd built my expertise around. AI would explain something
differently, and I'd think "that's wrong"—until I checked and realized I was
wrong.

That required humility I didn't know I needed.

You have to hold two contradictory ideas simultaneously:

1. Ask it to teach you things you think you know. The gaps in your
   understanding are often invisible until someone (or something) explains it
   differently.

2. Remember it might be completely making things up. AI hallucinates
   confidently. I've caught it generating plausible-sounding nonsense more times
   than I can count.

The skill isn't "use AI." The skill is knowing when to trust it and when to
verify everything.

Why I Study What I Study

This personal experience is why my research now focuses on provenance in
AI-generated content and hallucination reduction.

Specifically: advanced RAG (Retrieval-Augmented Generation) using knowledge
graphs and vector databases.

Translation for non-technical people: I'm working on ways to make AI systems
that can show their work and ground their answers in verifiable sources
instead of just making things up confidently.

Because I've lived both sides:

• The incredible learning acceleration when AI gets it right
• The dangerous overconfidence when AI gets it wrong and you don't notice

If we're going to use these systems for important decisions—education, business,
healthcare—we need to solve the hallucination problem. Not perfectly. But better
than "trust me, I'm an AI."

The Irony

The same technology that taught me intellectual humility is the technology that
everyone thinks will replace human expertise in 5 years.

Maybe. But that replacement assumes AI will eventually stop hallucinating, stop
needing human verification, and stop requiring expert judgment about when to
trust its output.

I use AI every single day. It's made me more productive, more curious, and more
humble.

It hasn't made me obsolete. It's made me better at being human.

And that's why I'm skeptical of the "AGI in 5 years" crowd. They're
underestimating how hard it is to build systems we can actually trust with
important decisions.

The Moment I Knew Everything Had Changed

I've been doing this for 9 months now. Teaching students what I call "vibe
coding"—working with AI to build things you couldn't build alone.

Then I gave my class an assignment: Sign up for the Shopify Partner Program.
Build a completely custom theme. You have 7 days.

My students had various levels of programming experience. Some had never built
anything for production. None of them had built a Shopify theme before.

Every single one of them delivered a beautiful, completely custom theme in 7
days.

They used VS Code with Copilot, Claude 4.5, GPT-4. No significant problems. No
one got stuck for long. No one failed.

That's when I realized: The tech world will never be the same.

For $10-20 a month, every student now has access to what is essentially the best
programmer in the world in their pocket. Available 24/7. Never tired. Never
annoyed by basic questions.

Watching the Wind Change Direction

I've been testing this intensely—vibe coding production applications, complex
systems, SaaS platforms. I sold my own startup years ago (a MEAN stack platform
for rapid SaaS development, back before it was even called MEAN stack). I know
what production code looks like. I know what real systems require.

And I've watched the models improve, month by month.

GPT-4 to GPT-4o to GPT-4.5. Claude 3.7 to Claude 4.5. I started with ChatGPT in
a browser the week it came out. I've seen the trajectory.

I can't tell you where this is going. I don't think anyone honestly can.

But I can see which way the wind is blowing.

And I can see enormous opportunity alongside potential tragedy.

Why This Matters Beyond Tech

Here's what keeps me up at night:

We're investing trillions of dollars in AI data centers. The United States
has an official AI strategy laid out on the White House website. It's connected
to the future of our economy.

We also have $36 trillion in national debt.

Simple math: Either AI works for everyone, or we face catastrophic inflation
from money printing, or massive taxation and service cuts for the people who
need help most.

And with what AI is about to do to the job market? There are going to be a lot
more people who need help.

We have to get AI right for everyone. We don't have a choice.

If we don't, our country won't just go bankrupt. We'll all be destitute.

You may not be interested in politics, or economics, or AI.

But politics, economics, and AI have interest in you.

The Real Question Isn't "When Does AGI Arrive?"

The real questions are:

• How do we deploy AI that makes everyone more productive, not just replaces
  jobs?
• How do we train people fast enough to adapt to the changes?
• How do we build systems we can actually trust with important decisions?
• How do we ensure the productivity gains benefit workers, not just
  shareholders?

AGI in 5 years, AGI in 30 years—it almost doesn't matter.

What matters is: Are we preparing people for the transition that's already
happening?

Because when every one of my students can build production-ready code in a week
with no prior experience, the transition isn't coming. It's here.

The Realistic Timeline

Instead of AGI in 5 years, here's what I actually think happens:

2025-2030:

• Steady improvements in AI capabilities
• Better at specific tasks, still clearly narrow
• Increasingly useful tools, not autonomous agents
• Productivity gains in specific domains

2030-2040:

• More general-purpose AI systems emerge
• Still not AGI, but more flexible than today
• Integration into most industries
• New AI-related jobs outnumber jobs lost

2040-2060:

• Possible AGI prototypes if we solve fundamental problems
• More likely: very capable narrow AI that looks general in limited domains
• Major productivity transformation finally materializing

Beyond 2060:

• AGI, if it's achievable, probably emerges here
• Superintelligence (if possible) would be decades after AGI
• Or we hit fundamental limits and plateau

Is this less exciting than "AGI in 5 years"? Maybe. But it's a lot more
realistic.

Why This Matters for You

If you're a student: You have time. Learn AI skills, but also learn critical
thinking, creativity, and human skills. You're building a 40-year career, not
preparing for next year.

If you're a worker worried about being replaced: The transition will be
gradual. Yes, some jobs will change or disappear. But history suggests new
jobs emerge. Focus on skills that complement AI, not compete with it.

If you're a company trying to plan: Don't wait for AGI. Use the AI available
today to solve real problems. The companies that win will be the ones that
figure out practical AI deployment, not the ones waiting for artificial
superintelligence.

The Honest Truth

I want to see AGI in my lifetime. It would be incredible. Transformative.
Maybe the most important development in human history.

But I've been around tech long enough to know that timelines are always
optimistic and hard problems stay hard longer than you expect.

When the CEO of an AI company tells you AGI is 5 years away, remember: he's got
billions in funding to justify and talent to recruit. His incentives aren't
aligned with accuracy—they're aligned with excitement.

When 2,700 AI researchers with no companies to sell tell you it's probably 20-30
years away, maybe listen to them instead.

The Second Renaissance is happening. It's real. It's exciting.

But it's going to take a lot longer than the hype cycle suggests.

And honestly? That's probably a good thing. It gives us time to figure out how
to do this right.

---

Next week: What actually happens to jobs when AI gets better. Spoiler: It's way
more complicated than "robots take everything."