Harper Carroll arrived in AI education from a Stanford computer science background, machine learning engineering at Meta, and a short spell at a small GPU compute startup in late 2023. There she noticed that few people understood how to fine-tune open source models, and she began writing and teaching to drive signups for the startup’s platform.
Her first guide, published right after Mistral 7B was released when she had about 50 followers, received 50,000 views. In March 2024 a video she made explaining the difference between AI and machine learning reached 5 million views; one in 20 viewers followed her afterward. She now has more than 500,000 followers across platforms and works full-time as an AI educator.
AI as a medium
The conversation emphasized that AI is a medium—like language or photography. The tools are available to everyone, but what differentiates outcomes is what people bring to the medium. A simple, well-designed workflow around AI will usually generate better outputs than a single, throwaway prompt.
Carroll describes AI broadly as the mathematics of understanding the world; text-based large language models are only one branch of that larger field.
Brain activity, outsourcing cognition and workflows
A circulated study found reduced brain activity in people who used AI to write essays compared to those who wrote without assistance, which led to alarmist claims about outsourcing cognition. Carroll points out that measured reductions can reflect shallow usage—one-sentence prompts accepted without revision—rather than a fundamental atrophy of thought. Saving time on some tasks can free cognitive resources for others, much like preferring online search to browsing a physical bookstore.
Why fine-tuning changes outputs where prompting doesn't
Carroll and the interviewer agreed on a key technical distinction: models are massive mathematical functions whose parameters are variables during training and constants at inference. Fine-tuning modifies the model weights themselves, shifting the underlying probability distribution of outputs. In-context prompting can pattern-match within frozen weights, and while surface behavior can bend, the core distribution doesn’t shift in the same way.
Her practical suggestion for dataset construction is to take your own writing, have the AI rewrite it with its characteristic tics, and then fine-tune the model using the AI version as input and your original as the target. This teaches the model to undo the telltale signs of machine-generated text.
Experiments and demos
Carroll first tried Claude with about 1,000 of her Instagram captions, video transcripts, and X posts as context; a detection tool marked the resulting output as 100% AI. She then fine-tuned an open-source Llama model on the same data; the fine-tuned output tested as 100% human. She demonstrated this at South by Southwest, and the demo took roughly 20 minutes.
Should people still learn to code?
Both agreed that people should still learn to code, but the focus should shift from memorizing syntax and trial-and-error to understanding systems and practical projects. ‘‘Vibe coding’’ lowers the entry barrier and allows more people to build products, but those with systems knowledge can build much more sophisticated things with the same tools.
Experienced developers also spot failure modes that pure vibe coders miss—Carroll gave an example of an agent that began storing data in a Word document and using it as a makeshift database because the session started with a Word file. An engineer sees the inefficiency immediately; a vibe coder might run the system for months before noticing the problem.
So the recommendation is to learn enough coding to understand what’s happening under the hood and to teach programming through projects that reveal software-architecture concepts.
Intuition remains a differentiator
Silicon Valley’s emphasis on logic, data, and sharp models can lead to intuition being dismissed as ‘‘soft.’’ Carroll argues that intuition—making decisions that run against immediate data patterns—remains a human advantage. Models trained to recognize patterns will struggle with counter-pattern judgment; intuition-informed judgment can therefore complement AI for a long time.
What the field often gets wrong publicly
Carroll says public discourse too often leads with fear: job displacement, imminent AGI, or a disruptive transition that requires universal basic income. While she doesn’t rule out those futures, she considers fear-first framing the wrong introduction to the technology. Fear produces avoidance, and avoidance is the fastest path to falling behind.
A better question is how to raise ambitions: AI not only scales individual capabilities but also what organizations can attempt. For that to happen broadly, everybody needs access and literacy—lower-cost models, serious open-source investment, and companies that do not become mere serfs to major platforms.
Carroll has been communicating this message to audiences ranging from engineers to people who have never coded: right now there is not all that much to fear, and AI can be an incredible productivity tool. Those who refuse to engage with it are likely to struggle.
Organizational practice and events
At O’Reilly, the approach is to start with people’s current jobs and find ways to ‘‘mix in’’ AI to make work more impactful, training both humans and agents to improve productivity together.
On July 9 the article’s author will speak with Trail of Bits cofounder and CEO Dan Guido about the playbook his company used to go AI-native. Guido will present the approach he outlined earlier and take about 40 minutes of audience questions on what worked, what didn’t, and what remains unsolved. The event is free and open to all.
Conclusion
Carroll’s experiences and experiments illustrate that fine-tuning can materially shift a model’s behavior in ways that prompting cannot, and that practical AI literacy—combined with systems understanding and human intuition—is essential if organizations and individuals are to benefit widely from these technologies.
