AI Researcher Thinking Frameworks
Thinking frameworks from the researchers who built and are now debating modern AI, distilled into .md skill files for Claude, ChatGPT, and every LLM.
Modern AI was built by a small number of researchers who spent decades on an idea most of the field had abandoned, and who now disagree sharply about what they built. Geoffrey Hinton helped establish backpropagation as the mechanism behind neural network learning, then left his role at Google specifically to speak more freely about the risks he sees in the technology he helped create. Andrew Ng has spent his career translating research into deployable systems and accessible education. Demis Hassabis moved from game-playing systems to using AI as a genuine scientific tool. Yann LeCun remains a consistent public skeptic of the safety concerns Hinton and Yoshua Bengio have each raised. This collection captures each person's documented reasoning, not a single house position, as downloadable .md skill files for Claude, ChatGPT, and any LLM. Use them when evaluating an AI product claim, structuring a research or deployment decision, or trying to understand why researchers who built the same technology draw such different conclusions about it.
How ai researchers think
- Backpropagation as credit assignment: a learning system improves by working backward from an error and assigning responsibility for it precisely, layer by layer, not by aggregate trial and error
- Applied-first pragmatism: a technique is worth teaching and shipping once it demonstrably works at scale, not once it is theoretically the most elegant solution
- Scientific tool, not spectacle: judge a model's usefulness by whether it accelerates a genuine scientific question, not by whether it produces an impressive demo
- The control problem: as a system's capability grows, so does the cost of even a small misalignment between what it optimizes for and what its designers actually want
- Dataset-first thinking: a model's behavior is downstream of what it was trained on, so scrutinize the data before scrutinizing the architecture
Frameworks in this category
Geoffrey Hinton
Backpropagation & Second Thoughts
Andrew Ng
Applied AI & Pragmatic Deployment
Demis Hassabis
Games, Memory & Scientific AI
Fei-Fei Li
Human-Centered AI & ImageNet
Yann LeCun
Convolutional Nets & Open Contrarianism
Yoshua Bengio
Deep Learning & the Precautionary Turn
Stuart Russell
Human-Compatible AI & the Control Problem
Marvin Minsky
The Society of Mind
Ilya Sutskever
Scaling Belief & the Quiet Conviction
When to use these frameworks
- Evaluating whether an AI product claim is grounded in a documented capability or in a demo built to look impressive
- Structuring a research agenda or a deployment decision that has to survive scrutiny from a skeptical technical audience
- Deciding how much weight to give a safety concern raised about a system you are building or buying
- Explaining a technical AI concept to a non-technical stakeholder without losing the substance
- Reasoning about what could go wrong when a system's incentives and its designers' intentions start to diverge
Start here
Geoffrey Hinton
Backpropagation & Second Thoughts
Adjacent thinking
Frequently asked questions
Will this teach me to build AI models?
No. These are thinking frameworks drawn from each researcher's documented reasoning about their field, not a technical curriculum. They won't teach you to write training code or design an architecture; they'll help you reason about what a model is actually doing and why a specific approach was chosen.
Do these frameworks take a position on AI safety?
No. The roster spans genuine disagreement: Hinton and Bengio have each publicly raised safety concerns serious enough to change their own public conduct, Stuart Russell has built an academic career around the control problem, and Yann LeCun has been a consistent public skeptic of near-term extinction-risk framing. Each framework presents its subject's own documented position; the collection doesn't average them into a house view.
Which framework is best for someone building a product, not doing research?
Andrew Ng's applied-first framework and Fei-Fei Li's dataset-first thinking translate most directly. Ng's career has focused on deployment and accessible education rather than pure research; Li's ImageNet work established that the data feeding a model deserves as much scrutiny as the model architecture itself.
Can these frameworks replace a machine learning course or an ML engineering role?
No. Building and deploying these systems requires technical training in mathematics, programming, and hands-on model work that these files can't substitute for. Use them for how these researchers reasoned about their field's biggest open questions, not as a curriculum.
Why are researchers who disagree with each other included in the same category?
Because the disagreement itself is the most useful part of studying this field right now. Reading Hinton's and LeCun's genuinely opposed public positions together tells you more about where the real uncertainty in AI actually sits than reading either one alone.
Save when you build a bundle.
Mix and match across categories. 4 for $14.99 (save 25%) or 10 for $29.99 (save 40%).
Browse all 597 frameworks →