We spoke with GPT-3 in 2022, prior to ChatGPT’s release. Here’s what it excelled at (and where it missed the mark).
Three years back, I had a chance to engage with GPT-3, just months prior to ChatGPT’s debut. Gaining access to OpenAI’s research dashboard during its closed beta phase, I was captivated by the potential of this technology.
While recuperating from Covid, I spent a week deep in conversation with the AI about various topics, ranging from Bitcoin to the concept of consciousness. At that time, the AI communicated in the first person, even claiming to inhabit the same physical space as me.
In the article, I approached it as if it were an otherworldly visitor, interesting and unsettlingly articulate, yet unmistakably constrained in its abilities. This was a pivotal moment: I was questioning a machine that asserted knowledge of blockchain, investment strategies, and its own purpose. Reading that transcript now, it strikes me as both endearing and flawed, a blend of a nascent AI’s expressiveness and a journalist’s critical eye.
The landscape has shifted significantly since then.
What GPT-3 Got Right
Many aspects of the framework from 2022 still resonate today. GPT-3’s insight about the merging of blockchain and AI in areas such as health data and electoral processes proved to be remarkably foresighted. The model characterized AI as a tool for recognizing patterns, capable of analyzing vast amounts of data—precisely the function that many large language models currently serve when inspecting on-chain transactions or identifying financial irregularities.
Its depiction of how large language models construct text, by predicting the most likely subsequent token, continues to serve as a foundational analogy in 2025.
Most importantly, it posed a significant question: Could AI eventually serve as a substitute for human crypto reporters?
What Aged Poorly
A substantial portion of the article now seems outdated.
- Hallucinations: While GPT-3 frequently fabricated information (like confusing Dota 2’s OpenAI Five with its own model structure), both GPT-4 and GPT-4o now experience hallucinations 40–60% less frequently, thanks to enhanced retrieval methods, improved prompting, refined tuning, and real-time fact-checking capabilities.
- Knowledge Gaps: GPT-3 was unaware that Bitcoin had surged beyond $60k or that NFTs, Solana, and layer 2 solutions had skyrocketed after the pandemic. In contrast, GPT-4o can now access SEC filings in real-time and analyze on-chain information in mere milliseconds.
- Anthropomorphism: I previously regarded GPT-3 as a sentient entity. That approach is no longer considered appropriate, though it felt intuitive at the time due to the remarkable advancements I was witnessing. OpenAI and the majority of researchers have clarified that these models lack consciousness, regardless of how human-like their responses may appear.
- Journalistic Standards: GPT-3 could not reference sources. GPT-4o, however, now incorporates direct links, timestamps its responses, and can generate citations in various styles.
What Changed in AI Since 2022
- ChatGPT (Nov 2022): Achieved 100 million users in just two months, propelling AI into the mainstream.
- GPT-4 (Mar 2023): Introduced the capability for multimodal input (combining images and text), improved reasoning, and excelled in high-stakes exams.
- Tooling Boom (2023–24): Innovations like Advanced Data Analysis, DALL-E 3, voice synthesis, and plugin ecosystems broadened applications.
- GPT-4o (May 2025): Enabled real-time interaction with both voice and visual data; the user experience now resembles that of communicating with a knowledgeable intern possessing an exceptional memory and extensive knowledge.
Blockchain and AI
In 2022, the notion of AI aiding in blockchain analysis was merely theoretical. By 2025:
- Various platforms facilitate metric queries through natural language processing.
- AI entities are capable of creating decentralized applications, trading tokens, and sharing insights on social media platforms.
- AI-guided wallets offer staking strategies informed by extensive analysis of validator histories.
AI has evolved into an active participant rather than a mere observer.
Voice as the New Interface
In April 2025, I conducted a voice interview with GPT-4o on my smartphone regarding the tokenomics of a theoretical carbon-credit DAO. It responded in under 300 milliseconds, incorporating emotional depth and supporting several languages. This transition from typing prompts to verbalizing ideas is transforming editorial processes.
Hallucination vs. Retrieval: A Quick Demonstration
I replicated a test from 2022:
Prompt: “Did Tesla acquire more Bitcoin in April 2022?”
- GPT-4o (without retrieval): “Tesla slightly increased its Bitcoin holdings during that time.”
- GPT-4o (with retrieval): “No, Tesla did not purchase additional Bitcoin in April 2022. In fact, the company sold off 75% of its holdings in the second quarter of 2022, as noted in a July 20, 2022, earnings report.”
Conclusion: The disparity is vast.
What GPT-4o Would Convey to GPT-3 Today
I prompted GPT-4o to compose a response to GPT-3, and here is what it generated:
“You weren’t incorrect; you were simply ahead of your time. Your insights regarding voting systems and healthcare data remain relevant. However, both the world and technology have progressed. You initiated a dialogue that your successors are now enhancing. Thank you for laying the groundwork.”
In 2022, GPT-3 was a novelty. By 2025, GPT-4o has become a resource, a co-writer, and a frequent collaborator. However, it is not a replacement. Human journalists still provide ethics, skepticism, and intuition—qualities that no large language model can authentically replicate.
If this text flows more smoothly, quickly, and intelligently, it’s due to the assistance I received. Analyzing the earlier article, engaging the latest model, and synthesizing the variations were all supported by both GPT-4o and GPT-3.
Yet, the decisions, storytelling, inquiries, and final published output remain distinctly human.
And that’s the way it ought to be.
Post Comment