OpenAI Wrote Their AGI Plan in 2018. I'm Sharing the Full 64 Slides.
117M parameters. 8 GPUs. A thesis the entire AI community dismissed. Eight years later, they were right about everything.
Before ChatGPT existed. Before GPT-4. Before the $300 billion valuation and the Microsoft deal and the drama with the board.
OpenAI had a plan.
In 2018, someone inside the company put together an internal deck laying out exactly how they intended to build artificial general intelligence. This was never meant for public consumption. No marketing polish. No PR spin. Just a raw, honest roadmap for building something most researchers thought was decades away.
The entire strategy fit on 64 slides.
I got my hands on it. And reading it in 2026 is genuinely surreal.
What They Believed When Nobody Else Did
The deck opens with a thesis that got OpenAI labeled as naive by most of the AI research community:
“Near-term AGI should be taken as a serious possibility.”
You have to remember the context. In 2018, reinforcement learning experts were saying tasks like Dota 2 were fundamentally unsolvable by AI systems. Critics argued that deep learning had already hit a wall and the hype cycle was about to collapse. The mainstream view among serious researchers was that AGI, if it ever happened at all, was somewhere between 30 and 100 years away.
OpenAI disagreed. And they wrote down exactly why.
The deck quotes Arthur C. Clarke on how scientists repeatedly declare things impossible right before breakthroughs happen. Flight was impossible until the Wright brothers. Space travel was impossible until Gagarin. Neural networks were a dead end until they suddenly powered everything.
The pattern repeats. The goalposts keep moving. And the people who look foolish in hindsight are always the ones who assumed the current limits were permanent.
The Four Pillars
Their entire technical strategy came down to four things:
1. Do hard things in simulation
Start with environments where you can generate unlimited training data. Games. Virtual worlds. Controlled sandboxes where agents can fail billions of times without real-world consequences.
2. Transfer skills to the real world
Once capabilities emerge in simulation, figure out how to move them into physical reality. Robotics. Interfaces. Practical applications.
3. Learn world models
Build systems that understand cause and effect, that can predict what happens next, that have something resembling intuition about how reality works.
4. Safety and deployment
Develop the capabilities and the safeguards together, instead of building first and worrying about alignment later.
That was the roadmap. Everything since then has been execution.
The Numbers That Matter
Their best model when this deck was written: 117 million parameters.
The training setup: 8 GPUs running for one month.
Context window: 512 tokens.
Compare that to today. Models with hundreds of billions of parameters. Training runs across tens of thousands of GPUs. Context windows stretching past 128,000 tokens. The infrastructure scaled by orders of magnitude.
But here’s what got me: the strategic direction barely changed. The four pillars from 2018 map almost perfectly onto what OpenAI actually built. GPT-4. DALL-E. The reasoning models. The agent systems. You can trace a direct line from the slides to the products.
They knew where they were going. They just had to figure out how to get there.
Why This Deck Is Worth Your Time
Reading this in 2026 is like finding Amazon’s original business plan after they became worth $2 trillion. You see what the founders saw before anyone else took them seriously. You understand why they made the bets they made. And you realize that the people who dismissed them were pattern-matching against the past instead of thinking about the future.
If you’re building in AI, this deck shows you how to think about long-term technical roadmaps when the research community thinks you’re wrong.
If you’re investing in AI, this is a masterclass in identifying conviction before consensus catches up.
If you’re just trying to understand where all of this is going, this deck puts the current moment in context better than anything else I’ve read.
The 64 slides are available to download below. I’ve also added notes on what to look for in each section.
The Full Deck (Downloadable PDF)
Pay attention to what they got right. Pay attention to what they underestimated. Both are worth studying.
The deck is downloadable as a PDF so you can save it, annotate it, and reference it later:
Keep reading with a 7-day free trial
Subscribe to The AI Corner to keep reading this post and get 7 days of free access to the full post archives.