Design for Human Error
ProductSystems should use unambiguous character sets like Base58 to prevent users from confusing similar-looking characters when transcribing data.
3 more supporting quotes are in the member records.
Hall of Masters
Andrej Karpathy
안드레이 카르파티
AI and frontier techBorn 1986
AI researcher, founding member of OpenAI, former director of AI at Tesla
705 records in all
- Principles 250
- Episodes 127
- Quotes 328
Ask with this master's words
3 free questions a day, no sign-up.
Principles
38 basic records of 250
Principles are our summary of their public statements.
Security Through Obscurity and Hashing
TechnologyProtect sensitive data by revealing it only when necessary; for example, Bitcoin hides public keys behind hashes until the moment funds are spent.
3 more supporting quotes are in the member records.
Build a Binary Classifier
LearningTo master a skill, study both excellent and poor examples to recognize exactly what makes something fail or succeed.
Supporting quotes
People have a strange bug built into psychology: a 10x more important or impactful problem intuitively feels 10x harder (or 10x less likely) to achieve. This is a fallacy - in my experience a 10x more important problem is at most 2-3x harder to achieve.
Source Andrej Karpathy A Survival Guide to a PhD, 2016 (opens in a new tab)
2 more supporting quotes are in the member records.
Leverage Vibe Coding
ProductivityUse high-level LLMs to rapidly prototype and build tools, focusing on the vision while the AI handles the bulk of the implementation.
3 more supporting quotes are in the member records.
Be the customer not the product
MindsetPrefer paying for services to ensure your data is not the primary source of monetization for the company.
3 more supporting quotes are in the member records.
Audit Outbound Communication
TechnologyRegularly monitor which applications are sending data from your machine to identify and remove suspicious software.
3 more supporting quotes are in the member records.
Resist the platform trap
StrategyDo not force a simple, effective tool to become a complex platform or a subscription service if it degrades the user experience.
3 more supporting quotes are in the member records.
Build a Simple Skeleton First
StrategyStart with a simple model and a fixed random seed to establish a baseline before adding any complex features or data augmentation.
3 more supporting quotes are in the member records.
Visual Source of Truth
TechnologyAlways visualize the data exactly as it enters the model to catch preprocessing bugs that metrics alone might miss.
3 more supporting quotes are in the member records.
Prefer Random Search for Tuning
ProductivityUse random search instead of grid search because neural networks are often more sensitive to some parameters than others, making fixed-point sampling inefficient.
3 more supporting quotes are in the member records.
Acknowledge the effort-accuracy tradeoff
ProductivityPerformance is not a fixed number but a result of the time, patience, and expertise invested in the task.
3 more supporting quotes are in the member records.
Leverage Simple Models for High ROI
ProductPrioritize simple architectures that yield disproportionately high-quality results before moving to complex systems.
Supporting quotes
What made this result so shocking at the time was that the common wisdom was that RNNs were supposed to be difficult to train
Source Andrej Karpathy The Unreasonable Effectiveness of Recurrent Neural Networks, 2015 (opens in a new tab)
2 more supporting quotes are in the member records.
Learn by small adjustments
LearningImprovement comes from repeatedly nudging parameters in the direction that reduces error until the system converges.
3 more supporting quotes are in the member records.
Prioritize Interpretability and Debugging
ProductivityAvoid tools or processes that sacrifice the ability to log and debug effectively for the sake of minor efficiency gains during development.
Supporting quotes
The concept of attention is the most interesting recent architectural innovation in neural networks.
Source Andrej Karpathy The Unreasonable Effectiveness of Recurrent Neural Networks, 2015 (opens in a new tab)
2 more supporting quotes are in the member records.
Democratize complex knowledge through teaching
ContentKarpathy scales his impact by creating accessible educational content, from Rubik's Cube tutorials to advanced AI courses, reaching thousands of learners.
Understand the First Principles
LearningOne should look past surface-level myths to understand the underlying chemical and physical processes of a system, like how weight loss is actually a gas exchange process.
Supporting quotes
the energy input to your body is upper bounded by the number of food calories that you eat.
Source Andrej Karpathy Biohacking Lite, 2020 (opens in a new tab)
2 more supporting quotes are in the member records.
Focus on trends over fluctuations
DecisionsDaily measurements can swing wildly due to noise like water weight; look at the long-term rate of change to understand true progress.
3 more supporting quotes are in the member records.
Question the Measurement Tools
DecisionsOne should not blindly trust data from sensors or tools if the results contradict observable reality, such as strength gains versus reported muscle loss.
3 more supporting quotes are in the member records.
Beware of automated bloat
TechnologyAI models tend to over-engineer by adding defensive code and deprecated APIs that increase complexity without adding value to custom projects.
Supporting quotes
They increase accessibility to languages or paradigms that you might not be as familiar with. I think they're very helpful there as well.
Source Andrej Karpathy Andrej Karpathy — AGI is still a decade away, 2025 (opens in a new tab)
2 more supporting quotes are in the member records.
Avoid outcome only feedback
LearningRelying solely on the final result to judge a long process is noisy and inefficient because it ignores which specific steps were actually correct.
3 more supporting quotes are in the member records.
You are seeing 106 basic records. Membership answers from all 705 of this master's records.
See membershipMore masters in this field
See the whole fieldAI and frontier tech
These records are our compilation of public statements and sources. We are not affiliated with or endorsed by Andrej Karpathy or any related company. If something should come down, tell us. Takedown request form