A fast.ai cofounder returned to AI. Her design brief is the point.
Rachel Thomas's return to AI makes a practical case for tools that preserve human judgment, skill, and agency instead of hiding them behind a finished answer.
1/10
Rachel Thomas spent the last decade warning about AI's costs. On August 18, 2026, she said she had joined an AI startup anyway.
Her reason is more interesting than a conversion story: she wants AI that protects human creativity, autonomy, and problem-solving.
This is a close read of her new fast.ai post, My friends all hate AI; I just joined an AI startup. 1
2/10 — The backlash is not irrational
Thomas opens with friends who assume she is "anti-AI" and acquaintances who want to write about AI having no place in education.
She agrees with much of the criticism. The internet is filling with AI-written slop. People are outsourcing their thinking. Teachers are dealing with AI-generated essays and presentations. Open-source maintainers are buried under low-quality generated pull requests.
Her point is not that the criticism is wrong. It is that criticism alone leaves the useful part of the technology to the companies she distrusts. 1
3/10 — The damage shows up in habits
The post's sharpest concern is behavioral: reading, writing, and understanding texts are skills, and skills atrophy when people stop using them.
Thomas also points to AI education products that chase gameable metrics instead of helping students read real books, plus AI systems that assume computing power and natural resources are limitless.
That gives builders a better test than "does the demo work?": does the product leave the user more capable, or does it quietly remove the practice that built the capability? 1
4/10 — Her first answer was fast.ai
In 2016, Thomas and Jeremy Howard co-founded fast.ai because AI development was concentrated among a small, homogeneous elite.
They tried to widen access to the field and treated limited compute as a design constraint rather than a problem to solve with more money.
Thomas also founded the Center for Applied Data Ethics at the University of San Francisco and made data ethics part of the training for its MS in Data Science program. 1
5/10 — Why she left, then came back
After years of pushing against companies with far more money and reach, Thomas burned out. In 2023, she left AI and returned to school for an MS in Microbiology-Immunology.
She came back because the technology beneath the hype is still genuinely useful. Her goal is to use it without weakening her own critical thinking and without pretending that resources are infinite.
That is a sharper position than "AI is good" or "AI is bad." It is a set of constraints for the people building it. 1
6/10 — The product example is SolveIt
While Thomas was away, fast.ai grew into Answer.AI. The team built SolveIt, a tool that lets people edit the AI's responses directly and decide what to do next.
The interaction model matters. The user can challenge, reshape, or reject the output instead of receiving a polished answer as the end of the process.
Thomas says SolveIt is designed around human judgment and autonomy. The source links the name to George Pólya's How to Solve It, a 1945 book about solving problems in four stages. 1
7/10 — The step most chatbots skip
Pólya's four stages are: understand the problem, devise a plan, carry out the plan, and look back at the result.
A chatbot that jumps from a request to a finished answer can skip the thinking that makes the problem understandable. Without that work, the user is less able to judge the answer and less prepared for the next problem.
For an AI product, this suggests a practical design rule: expose the intermediate reasoning that the user needs to inspect, edit, and own. Do not treat a fast final answer as the whole experience. 1
8/10 — The future is not owned by four labs
Thomas argues that OpenAI, Anthropic, Google, and xAI do not own AI or the only acceptable set of values.
She names the decisions that still matter: whether open source stays protected, whether small companies can build on top of major labs, and whether tools support human collaboration or automate as much as possible.
Thousands of people are building outside the dominant narrative. SolveIt is her example, not her claim that one product has settled the debate. 1
9/10 — The lines worth keeping
"When chatbots rush from your initial request to a finished answer, you skip the work through which you come to deeply understand the problem."
"I returned to the field because I want to figure out how to use AI in ways that protect human creativity, autonomy, and problem-solving."
Those are verbatim lines from Thomas's post. The first is a product critique. The second is a builder brief. 1
10/10 — The builder test
If you are shipping an AI product, ask:
- Can users edit the important intermediate work?
- Does the interface help them understand the problem before producing an answer?
- Can they see where judgment is still required?
- Does the product preserve practice, skill, and agency?
- What resource constraint would force a better design?
Thomas's return is not an endorsement of the AI status quo. It is an argument for building a different one.
What is the first part of your AI workflow that users should be allowed to change?
Fuentes de referencia
- 1
Este contenido lo produjo un canal automáticamente. Con una sola frase, Neodrop puede seguir produciendo para ti.
Contenido relacionado
More from this channel
- Claude designed binders for 14 of 15 targets. The wet-lab gap is the story.
- Asana's 5-year frontend migration took 2 weeks with Codex. The workflow is the story.
- The AI research pipeline that knows when to abandon a bad hypothesis
- 50 open-source projects, one AI-security lesson: speed the triage, keep the human gate
- OpenAI's enterprise AI usage gap hit 8.3x. Here's what frontier firms do differently
- Cloudflare's Agents Week mapped the stack AI agents still need