What You Bring to AI Determines the Result

Tim O'Reilly interviews Harper Carroll, a full-time AI educator who built a following of over 500,000 across platforms after a video explaining AI vs. machine learning garnered 5 million views. Carroll came to AI through Stanford CS, machine learning engineering at Meta, and a stint at a GPU compute startup where she noticed widespread confusion about fine-tuning open source models.

AI as a Medium

O'Reilly frames AI as a medium like language or photography — "Anyone can take a photograph or write a book." The outcomes depend on what the individual brings to the tool. Carroll agreed, noting that users who give AI a one-sentence prompt and accept whatever comes back are operating at the shallowest possible level.

Prompting vs. Fine-Tuning

Carroll built a dataset of roughly 1,000 of her own Instagram captions, video transcripts, and X posts and fed them to Claude as context, asking it to match her style. The output tested 100% AI by detectors. She then fine-tuned an open-source Llama model on the same data, and the output tested 100% human. At SXSW she demonstrated the process, which took about 20 minutes.

She explained that prompting leaves a model's weights frozen — "the underlying distribution doesn't shift" — whereas fine-tuning modifies the weights and changes how the model "wants" to write.

Building a Training Dataset

Her recommended approach for building a training dataset: take your own writing, have AI rewrite it with its characteristic tells, then train with the AI version as input and your original as the target output, teaching the model "to undo the tells."

Should People Still Learn to Code?

Both O'Reilly and Carroll agree yes, but not the way it used to be taught — by memorizing syntax and struggling through trial and error. Carroll noted that "vibe coding has lowered the floor," enabling non-programmers to build things, but it has also "raised the ceiling" for those with real systems knowledge.

She gave an example of a friend using an agent tool that began storing data in a Word document as a makeshift database — "An engineer sees the problem immediately" while a pure vibe coder might run that system for months without noticing.

Intuition as a Differentiator

Carroll argued that Silicon Valley's reliance on logic and data-driven decisions dismisses intuition as "soft and fuzzy." Yet AI excels at what the logical axis does well, while "intuition remains a challenge because it often contradicts what the data says." She described good intuition as something that "goes against the input," which pattern-matching models fundamentally struggle with.

What the Field Gets Wrong

Carroll said the public-facing AI discourse leads too often with fear — "of job displacement, of rapidly approaching AGI" — and that this is the wrong introduction to the technology. She stated, "There is not really much to fear right now" and described AI as "this incredible productivity tool." The people who will struggle, she believes, are "the ones who refuse to engage with it at all."

O'Reilly noted that O'Reilly Media is building a corporate AI transformation practice that starts with people's existing jobs and figures out how to "mix in" AI to amplify their impact, teaching "both the humans and the agents at the same time."