We interviewed GPT-3 in 2022 before ChatGPT. Here’s what it got right (and so very wrong)

“I believe that the all-time high of Bitcoin was around $20,000.” — GPT-3, April 2022

“Bitcoin peaked at $68,990 in November 2021 due to a confluence of pandemic-era stimulus, institutional momentum, and speculative fervor. The misquote from GPT-3 highlights the limitations of pre-trained, non-updated models.” — GPT-4o, May 2025

Three years ago, I interviewed GPT-3 several months before the launch of ChatGPT. I had closed beta access to OpenAI’s research dashboard and was fascinated by what this technology could do.

I spent a week recovering from Covid, ‘chatting’ late into the night with the AI about everything from Bitcoin to consciousness. At the time, the AI responded in the first person and claimed to be able to move around the world, even claiming to be in the room with me.

In the article, I treated it like an alien guest, curious, eerily articulate, but fundamentally limited. It was a historical moment: I was interrogating a machine that claimed to understand blockchain, investment logic, and even its own purpose. Looking back, the transcript reads as both charming and flawed, a document shaped by a young AI’s eloquence and a journalist’s skepticism.

Today, the game has changed.

What GPT-3 got right

Much of the framing in 2022 still holds. GPT-3’s prediction that blockchain and AI could intersect in sectors like health data and voting remains prescient. The model described AI as a pattern-spotter, capable of parsing massive datasets, precisely the role many LLMs now play in analyzing on-chain activity or monitoring financial anomalies.

Its explanation of how LLMs generate text, predicting the next most likely token, still serves as the go-to metaphor in 2025.

And perhaps most crucially, its answers raised the big question: Could AI eventually replace human crypto journalists?

What aged poorly

A lot of the piece now feels quaint.

• Hallucinations: While GPT-3 often fabricated facts (e.g., mixing up Dota 2’s OpenAI Five with its own architecture), GPT-4 and GPT-4o now hallucinate 40–60% less often thanks to retrieval-augmented generation (RAG), improved prompting, better fine-tuning, and real-time fact-checking tools.
• Knowledge Gaps: GPT-3 didn’t know Bitcoin had hit $60k+ or that NFTs, Solana, and L2s had exploded post-pandemic. Today, GPT-4o can live-browse SEC filings, read Etherscan, and integrate on-chain data with milliseconds of latency.
• Anthropomorphism: I treated GPT-3 like a sentient being. That’s no longer acceptable practice, but it felt ‘natural’ at the time, given the leap forward I was experiencing. OpenAI and most researchers have clarified that models are not conscious, no matter how lifelike their prose.
• Journalistic Benchmarks: GPT-3 couldn’t cite sources. GPT-4o now hyperlinks directly, timestamps answers, and can generate reference lists in any citation style.

What Changed in AI Since 2022

• ChatGPT (Nov 2022): 100M users in two months, AI went mainstream.
• GPT-4 (Mar 2023): Introduced multimodal inputs (images + text), better reasoning, and 90th percentile exam scores.
• Tooling Boom (2023–24): Advanced Data Analysis, DALL-E 3, voice synthesis, and plugin ecosystems expanded use cases.
• GPT-4o (May 2025): Real-time voice and vision; the interface now feels like talking to a savvy intern with an eidetic memory and superhuman knowledge base.

Blockchain x AI

In 2022, AI helping analyze blockchain was a theory. In 2025:

• Multiple platforms allow metrics to be queried in natural language.
• AI agents create dApps and trade tokens and post them on Crypto Twitter.
• AI-driven wallets suggest staking strategies based on LLM analysis of validator histories.

AI is no longer just an observer.

Voice is the New Keyboard

In April 2025, I conducted a voice interview with GPT-4o on my phone. I asked about tokenomics for a hypothetical carbon-credit DAO. It replied in <300 ms, with emotional nuance and multilingual support. That shift, from typing prompts to talking through ideas, is redefining editorial workflows.

Hallucination vs. Retrieval: A Quick Demo

I ran the same test I did in 2022:

Prompt: “Did Tesla buy more Bitcoin in April 2022?”

• GPT-4o (no retrieval): “Tesla increased its Bitcoin position slightly during that period.”
• GPT-4o (with retrieval): “No, Tesla did not buy additional Bitcoin in April 2022. In fact, the company sold 75% of its holdings in Q2 2022, according to a July 20, 2022, earnings report.”

Verdict: The difference is night and day.

What GPT-4o would say to GPT-3 today

I prompted GPT-4o to write a response to GPT-3. This is the response:

“You weren’t wrong, just early. Your ideas about voting systems and healthcare data still have traction. But the world—and the tech—moved on. You sparked a conversation that your successors are now refining. Thank you for getting the first draft down.”

In 2022, GPT-3 was a curiosity. In 2025, GPT-4o is a tool, a co-author, and a daily collaborator. But it is not a replacement. Human journalists still bring ethics, skepticism, and instinct, things no LLM can authentically simulate.

If this article reads smoother, faster, and smarter, it’s because I had help. Parsing through the historical article, prompting the latest model, and synthesizing the differences were all supported by GPT-4o and GPT-o3.

But the judgment, the narrative, the questions, and the final published content? Those are still human.

And that’s how it should be.

Source