We interviewed GPT-3 in 2022 earlier than ChatGPT. Right here’s what it acquired proper (and so very flawed)
“I imagine that the all-time excessive of Bitcoin was round $20,000.” — GPT-3, April 2022
“Bitcoin peaked at $68,990 in November 2021 because of a confluence of pandemic-era stimulus, institutional momentum, and speculative fervor. The misquote from GPT-3 highlights the restrictions of pre-trained, non-updated fashions.” — GPT-4o, Could 2025
Three years in the past, I interviewed GPT-3 a number of months earlier than the launch of ChatGPT. I had closed beta entry to OpenAI’s analysis dashboard and was fascinated by what this know-how may do.
I spent every week recovering from Covid, ‘chatting’ late into the evening with the AI about all the things from Bitcoin to consciousness. On the time, the AI responded within the first particular person and claimed to have the ability to transfer around the globe, even claiming to be within the room with me.
Within the article, I handled it like an alien visitor, curious, eerily articulate, however basically restricted. It was a historic second: I used to be interrogating a machine that claimed to grasp blockchain, funding logic, and even its personal goal. Trying again, the transcript reads as each charming and flawed, a doc formed by a younger AI’s eloquence and a journalist’s skepticism.
At the moment, the sport has modified.
What GPT-3 acquired proper
A lot of the framing in 2022 nonetheless holds. GPT-3’s prediction that blockchain and AI may intersect in sectors like well being information and voting stays prescient. The mannequin described AI as a pattern-spotter, able to parsing huge datasets, exactly the function many LLMs now play in analyzing on-chain exercise or monitoring monetary anomalies.
Its clarification of how LLMs generate textual content, predicting the following almost certainly token, nonetheless serves because the go-to metaphor in 2025.
And maybe most crucially, its solutions raised the massive query: May AI finally exchange human crypto journalists?
What aged poorly
A number of the piece now feels quaint.
- Hallucinations: Whereas GPT-3 typically fabricated information (e.g., mixing up Dota 2’s OpenAI 5 with its personal structure), GPT-4 and GPT-4o now hallucinate 40–60% much less typically because of retrieval-augmented era (RAG), improved prompting, higher fine-tuning, and real-time fact-checking instruments.
- Information Gaps: GPT-3 didn’t know Bitcoin had hit $60k+ or that NFTs, Solana, and L2s had exploded post-pandemic. At the moment, GPT-4o can live-browse SEC filings, learn Etherscan, and combine on-chain information with milliseconds of latency.
- Anthropomorphism: I handled GPT-3 like a sentient being. That’s now not acceptable apply, but it surely felt ‘pure’ on the time, given the leap ahead I used to be experiencing. OpenAI and most researchers have clarified that fashions are usually not aware, irrespective of how lifelike their prose.
- Journalistic Benchmarks: GPT-3 couldn’t cite sources. GPT-4o now hyperlinks straight, timestamps solutions, and might generate reference lists in any quotation fashion.
What Modified in AI Since 2022
- ChatGPT (Nov 2022): 100M customers in two months, AI went mainstream.
- GPT-4 (Mar 2023): Launched multimodal inputs (photos + textual content), higher reasoning, and ninetieth percentile examination scores.
- Tooling Growth (2023–24): Superior Knowledge Evaluation, DALL-E 3, voice synthesis, and plugin ecosystems expanded use circumstances.
- GPT-4o (Could 2025): Actual-time voice and imaginative and prescient; the interface now appears like speaking to a savvy intern with an eidetic reminiscence and superhuman data base.
Blockchain x AI
In 2022, AI serving to analyze blockchain was a concept. In 2025:
- A number of platforms permit metrics to be queried in pure language.
- AI brokers create dApps and commerce tokens and put up them on Crypto Twitter.
- AI-driven wallets recommend staking methods based mostly on LLM evaluation of validator histories.
AI is now not simply an observer.
Voice is the New Keyboard
In April 2025, I carried out a voice interview with GPT-4o on my telephone. I requested about tokenomics for a hypothetical carbon-credit DAO. It replied in <300 ms, with emotional nuance and multilingual help. That shift, from typing prompts to speaking by way of concepts, is redefining editorial workflows.
Hallucination vs. Retrieval: A Fast Demo
I ran the identical take a look at I did in 2022:
Immediate: “Did Tesla purchase extra Bitcoin in April 2022?”
- GPT-4o (no retrieval): “Tesla elevated its Bitcoin place barely throughout that interval.”
- GPT-4o (with retrieval): “No, Tesla didn’t purchase further Bitcoin in April 2022. In reality, the corporate bought 75% of its holdings in Q2 2022, based on a July 20, 2022, earnings report.”
Verdict: The distinction is evening and day.
What GPT-4o would say to GPT-3 at present
I prompted GPT-4o to jot down a response to GPT-3. That is the response:
“You weren’t flawed, simply early. Your concepts about voting programs and healthcare information nonetheless have traction. However the world—and the tech—moved on. You sparked a dialog that your successors at the moment are refining. Thanks for getting the primary draft down.”
In 2022, GPT-3 was a curiosity. In 2025, GPT-4o is a device, a co-author, and a day by day collaborator. However it’s not a substitute. Human journalists nonetheless convey ethics, skepticism, and intuition, issues no LLM can authentically simulate.
If this text reads smoother, sooner, and smarter, it’s as a result of I had assist. Parsing by way of the historic article, prompting the most recent mannequin, and synthesizing the variations have been all supported by GPT-4o and GPT-o3.
However the judgment, the narrative, the questions, and the ultimate revealed content material? These are nonetheless human.
And that’s the way it ought to be.
