The Debut: Harvard, November 2024
First in a series about my AI lecture tour through Ivy League and international universities
The Setup
November 2024. I’m at Harvard to give a talk about AI and blockchain for Andromeda Protocol. On paper, I’m there to explain how AI complements blockchain technology - smart contract optimization, predictive analytics, fraud detection.
But I have a secret.
What Everyone Else Was Doing
At this point, the AI world was obsessed with OpenAI’s custom GPTs. They’d launched a few months earlier and everyone thought they were the silver bullet - finally, a way to give ChatGPT your own context, your own documents, your own knowledge base.
Except they weren’t.
The RAG system in custom GPTs was brutally limited. Any document you uploaded got truncated at 10,000 characters. You could only add a handful of documents. And depending on what you put in there, it would hallucinate - badly.
Some clever people experimented with self-indexing workarounds: “Hit A if you want to do X, hit B for Y.” It helped, but it wasn’t the revolution everyone had hoped for.
Engineers I talked to considered AI tools “unreliable.” They’d tried them, gotten burned, moved on.
What I Was Doing
While the industry was trying to make custom GPTs work, I was doing something different.
Google’s Gemini Pro models had 2 million token context windows. Two million. So I thought: what if I just… stuffed everything in there?
I took all of Andromeda’s documentation. Relevant parts of our entire codebase. Technical specs. Architecture decisions. Everything.
Then I developed a technique:
- Feed a chunk of context
- Ask the model to index itself - create a table of contents, explain what it now knows
- Feed more context
- Ask it to index and explain again
- Rinse and repeat
By the end, I had a conversation with nearly a million tokens of primed context. The model understood our entire system. And critically - it was never wrong.
I ran needle-in-haystack tests. Benchmarked it against edge cases. Threw curveballs at it. It held up.
This was chain-of-thought prompting before we called it that. I was forcing the model to reason through and organize information before generating anything. The same principle that OpenAI would later bake into o1 and call “reasoning.”
The Lecture
So there I am at Harvard, showing students tools and pieces of AI history, trying to give them a sense of what was coming down the pike.
But I couldn’t explain WHY I was so confident. That would reveal the technique I thought would be Andromeda’s competitive advantage.
I was cramming the room with visual demonstrations, trying to paint a picture of the future, while knowing I had evidence they couldn’t see. The models everyone dismissed as “unreliable” were actually incredibly powerful - if you knew how to prime them.
It was frustrating. I wanted to shout from the rooftops what I’d discovered. But I kept it to myself.
The Irony
Andromeda was not destined for greatness. It was destined to be vaporware.
So I kept this technique secret for a company that didn’t make it, while the industry eventually caught up anyway. Reasoning models are now the most significant development in LLMs. They’ve built the “think before you respond” pattern directly into the architecture.
I was right about what was coming. I just couldn’t tell anyone why.
What I Actually Talked About
My portion covered:
- My background (Kurzweil’s “How to Create a Mind” talk in 2012 that got me into AI)
- The progression of AI from 2012-2024 (AlexNet → GPT → ChatGPT)
- The current AI engineering stack (Cursor, AI Studio, Claude, Copilot)
- How we were using AI at Andromeda (including our internal tool Androptimus)
Leif from Andromeda covered the blockchain specifics. I did the AI deep-dive.
The students were engaged, but skeptical. In late 2024, most engineers still considered AI tools interesting but unreliable. They hadn’t seen what was possible with proper context priming.
They would, soon enough.
Next: Cornell Tech - where o1 dropped four days before my lecture and I got to connect the dots in real-time.
Date: November 19, 2024 Venue: Harvard University Co-presenter: Leif (Andromeda Protocol) Topic: “Transforming Blockchain with AI: Andromeda’s Next Frontier”