2 weeks ago
“I judge that the all-time precocious of Bitcoin was astir $20,000.” — GPT-3, April 2022
“Bitcoin peaked astatine $68,990 successful November 2021 owed to a confluence of pandemic-era stimulus, organization 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 respective months earlier the motorboat of ChatGPT. I had closed beta entree to OpenAI’s probe dashboard and was fascinated by what this exertion could do.
I spent a week recovering from Covid, ‘chatting’ precocious into the nighttime with the AI astir everything from Bitcoin to consciousness. At the time, the AI responded successful the archetypal idiosyncratic and claimed to beryllium capable to determination astir the world, adjacent claiming to beryllium successful the country with me.
In the article, I treated it similar an alien guest, curious, eerily articulate, but fundamentally limited. It was a humanities moment: I was interrogating a instrumentality that claimed to recognize blockchain, concern logic, and adjacent its ain purpose. Looking back, the transcript reads arsenic some charming and flawed, a papers shaped by a young AI’s eloquence and a journalist’s skepticism.
Today, the crippled has changed.
What GPT-3 got right
Much of the framing successful 2022 inactive holds. GPT-3’s prediction that blockchain and AI could intersect successful sectors similar wellness information and voting remains prescient. The exemplary described AI arsenic a pattern-spotter, susceptible of parsing monolithic datasets, precisely the relation galore LLMs present play successful analyzing on-chain enactment oregon monitoring fiscal anomalies.
Its mentation of however LLMs make text, predicting the adjacent astir apt token, inactive serves arsenic the go-to metaphor successful 2025.
And possibly astir crucially, its answers raised the large question: Could AI yet regenerate quality crypto journalists?
What aged poorly
A batch of the portion present feels quaint.
- Hallucinations: While GPT-3 often fabricated facts (e.g., mixing up Dota 2’s OpenAI Five with its ain architecture), GPT-4 and GPT-4o present hallucinate 40–60% little often acknowledgment to retrieval-augmented procreation (RAG), improved prompting, amended fine-tuning, and real-time fact-checking tools.
- Knowledge Gaps: GPT-3 didn’t cognize Bitcoin had deed $60k+ oregon that NFTs, Solana, and L2s had exploded post-pandemic. Today, GPT-4o tin live-browse SEC filings, work Etherscan, and integrate on-chain information with milliseconds of latency.
- Anthropomorphism: I treated GPT-3 similar a sentient being. That’s nary longer acceptable practice, but it felt ‘natural’ astatine the time, fixed the leap guardant I was experiencing. OpenAI and astir researchers person clarified that models are not conscious, nary substance however lifelike their prose.
- Journalistic Benchmarks: GPT-3 couldn’t mention sources. GPT-4o present hyperlinks directly, timestamps answers, and tin make notation lists successful immoderate citation style.
What Changed successful AI Since 2022
- ChatGPT (Nov 2022): 100M users successful 2 months, AI went mainstream.
- GPT-4 (Mar 2023): Introduced multimodal inputs (images + text), amended reasoning, and 90th percentile exam scores.
- Tooling Boom (2023–24): Advanced Data Analysis, DALL-E 3, dependable synthesis, and plugin ecosystems expanded usage cases.
- GPT-4o (May 2025): Real-time dependable and vision; the interface present feels similar talking to a savvy intern with an eidetic representation and superhuman cognition base.
Blockchain x AI
In 2022, AI helping analyse blockchain was a theory. In 2025:
- Multiple platforms let metrics to beryllium queried successful earthy language.
- AI agents create dApps and commercialized tokens and station them on Crypto Twitter.
- AI-driven wallets suggest staking strategies based connected LLM investigation of validator histories.
AI is nary longer conscionable an observer.
Voice is the New Keyboard
In April 2025, I conducted a dependable interrogation with GPT-4o connected my phone. I asked astir tokenomics for a hypothetical carbon-credit DAO. It replied successful <300 ms, with affectional nuance and multilingual support. That shift, from typing prompts to talking done ideas, is redefining editorial workflows.
Hallucination vs. Retrieval: A Quick Demo
I ran the aforesaid trial I did successful 2022:
Prompt: “Did Tesla bargain much Bitcoin successful April 2022?”
- GPT-4o (no retrieval): “Tesla accrued its Bitcoin presumption somewhat during that period.”
- GPT-4o (with retrieval): “No, Tesla did not bargain further Bitcoin successful April 2022. In fact, the institution sold 75% of its holdings successful Q2 2022, according to a July 20, 2022, net report.”
Verdict: The quality is nighttime and day.
What GPT-4o would accidental to GPT-3 today
I prompted GPT-4o to constitute a effect to GPT-3. This is the response:
“You weren’t wrong, conscionable early. Your ideas astir voting systems and healthcare information inactive person traction. But the world—and the tech—moved on. You sparked a speech that your successors are present refining. Thank you for getting the archetypal draught down.”
In 2022, GPT-3 was a curiosity. In 2025, GPT-4o is simply a tool, a co-author, and a regular collaborator. But it is not a replacement. Human journalists inactive bring ethics, skepticism, and instinct, things nary LLM tin authentically simulate.
If this nonfiction reads smoother, faster, and smarter, it’s due to the fact that I had help. Parsing done the humanities article, prompting the latest model, and synthesizing the differences were each supported by GPT-4o and GPT-o3.
But the judgment, the narrative, the questions, and the last published content? Those are inactive human.
And that’s however it should be.
The station We interviewed GPT-3 successful 2022 earlier ChatGPT. Here’s what it got close (and truthful precise wrong) appeared archetypal connected CryptoSlate.
English (US) 