SUMMARY

Andrew Ng delivers a talk at Y Combinator's AI Startup School on accelerating startup execution with AI, emphasizing speed, agentic workflows, concrete ideas, and responsible innovation.

STATEMENTS

AI Fund, as a venture studio, builds an average of one startup per month by actively participating in coding, customer discussions, feature design, and pricing decisions, providing hands-on experience in startup creation.
Execution speed is a strong predictor of startup success, and new AI technologies enable entrepreneurs to operate much faster, increasing their odds of achieving product-market fit and overall viability.
Best practices for speed in AI startups evolve rapidly every two to three months, requiring founders to adapt continuously to leverage emerging tools and workflows for competitive advantage.
Opportunities in the AI stack are greatest at the application layer, as applications generate revenue to sustain underlying semiconductor, cloud, and foundation model layers, despite less media attention on apps.
The rise of agentic AI represents the most important trend in AI technology, allowing iterative workflows that improve output quality over linear prompting, opening vast startup opportunities in workflow automation.
Traditional LLM usage mimics forcing humans to write essays linearly without revisions, limiting quality, whereas agentic workflows enable outlining, research, drafting, critiquing, and iterating for superior results.
Agentic workflows have proven essential for complex tasks like compliance document extraction, medical diagnosis, and legal reasoning at AI Fund, turning potential failures into successes through iteration.
An emerging agentic orchestration layer simplifies building applications by coordinating calls to underlying tech layers, making app development easier while reinforcing the value of the application layer in the AI stack.
Concrete product ideas, specified in enough detail for engineers to build immediately, enable rapid execution and validation, contrasting with vague ideas that slow progress and invite unfocused efforts.
Vague ideas like "AI for healthcare optimization" receive praise but hinder speed, while concrete ones like "software for online MRI booking in hospitals" allow quick building and testing, even if ultimately flawed.
Subject matter experts who have deeply pondered a domain for years develop reliable gut instincts for decisions, serving as a faster alternative to data collection in early startup stages.
Successful startups pursue one clear hypothesis at a time with full commitment, pivoting decisively when data disproves it, avoiding resource dilution across multiple unproven paths.
Frequent pivots on every new data point indicate insufficient domain knowledge, suggesting the need for deeper expertise to generate high-quality concrete ideas and maintain momentum.
AI coding assistants accelerate the build-feedback loop in application startups, enabling rapid engineering iterations toward product-market fit by drastically reducing development time and costs.
For quick prototypes, AI tools make developers 10 times faster or more, due to lower integration needs and relaxed requirements on reliability, scalability, and security during initial testing.
In production code, AI assistance yields 30-50% speedups, but prototypes benefit far more, allowing startups to systematically test 20 ideas cheaply, accepting that many won't reach production.
The mantra "move fast and break things" should evolve to "move fast and be responsible," balancing speed with accountability, especially as AI lowers barriers to rapid experimentation.
New generations of agentic coding assistants like Claude, Cursor, and Devin, evolving every few months, boost productivity significantly, with lagging tools creating substantial competitive disadvantages.
As coding costs plummet, codebases are rebuilt frequently—sometimes three times a month—reducing code's perceived value and enabling flexible decisions on tech stacks without high reversal costs.
Software architecture choices, once irreversible "one-way doors," are becoming "two-way doors" due to cheaper engineering, allowing teams to experiment and pivot tech stacks weekly if needed.
Despite AI automating some coding, learning to code remains crucial, as easier tools democratize programming, empowering non-engineers like CFOs and recruiters to enhance their roles through automation.
Domain knowledge enhances AI tool usage; for instance, an art history expert generates superior Midjourney images by specifying genres and palettes, while novices produce generic outputs.
The ability to precisely instruct computers—best learned through coding—will be a key future skill, enabling users to leverage AI effectively for desired outcomes across professions.
With engineering speed surging via AI, product management becomes the new bottleneck, shifting team ratios from 1 PM per 4-7 engineers to potentially 1 PM per 0.5 engineers in some cases.
PMs who code or engineers with product instincts excel in this shifted landscape, as rapid feedback tactics like gut checks, stranger interviews, and A/B tests guide feature decisions efficiently.
Understanding AI provides a speed advantage in emerging tech, unlike mature fields like mobile or marketing where knowledge is widespread; wrong AI decisions can lead to months-long dead ends.
GenAI building blocks like prompting, evals, guardrails, RAG, and fine-tuning combine combinatorially, enabling novel applications impossible a year ago, with continuous learning unlocking exponential possibilities.
Startup success correlates strongly with execution speed, achieved through concrete ideas, rapid engineering, quick feedback, and AI literacy, positioning teams to capitalize on vast untapped opportunities.

IDEAS

Execution speed predicts startup success more than ideas alone.
AI stack's biggest opportunities lie in revenue-generating applications.
Agentic AI shifts from linear prompts to iterative, high-quality workflows.
Concrete ideas enable engineers to build immediately, boosting velocity.
Vague ideas garner praise but stall progress due to ambiguity.
Gut instincts from domain experts outpace slow data collection.
Pursue one hypothesis doggedly, pivot decisively on evidence.
AI coding makes prototypes 10x faster than production code.
Rebuild codebases frequently as engineering costs plummet dramatically.
Learning to code empowers all roles, enhancing productivity universally.
Precise computer instructions, via coding, define future power skills.
Product management now bottlenecks faster AI-accelerated engineering teams.
Feedback tactics range from gut to A/B testing in speed-accuracy trade-offs.
AI knowledge gives edge in emerging tech over mature domains.
Building blocks combine exponentially for novel AI applications.
Hype narratives distort AI realities for promotional gains.
Responsible AI focuses on application, not inherent technology safety.
Moats evolve post-product; prioritize user love initially.
Agentic workflows integrate blocks, but token costs rarely constrain early.
Flexibility in model switching via evals accelerates adaptation.
Education experiments with AI tutors, but end-state remains unclear.
Ethical kills of viable projects ensure societal benefit alignment.
Empower non-engineers with AI to bridge productivity gaps.
Protect open source from hype-driven regulatory gatekeeping threats.
Hyperpersonalized education via AI demands ongoing workflow mapping.

INSIGHTS

Speed trumps perfection in startups, amplified by AI tools.
Concrete specificity unlocks rapid validation, falsifying ideas efficiently.
Agentic iteration mirrors human creativity, elevating AI outputs.
Domain expertise fuels intuitive decisions, bypassing data delays.
Bottlenecks shift from engineering to human-centric product judgment.
Coding literacy democratizes AI command across all professions.
Hype inflates risks, masking responsible innovation's true potential.
Combinatorial building blocks spawn unprecedented application possibilities.
Ethical responsibility integrates into fast execution without slowing progress.
Open source preservation ensures equitable AI knowledge diffusion.

QUOTES

"Execution speed is a strong predictor for startup's odds of success."
"The biggest opportunities have to be at the application layer."
"Agentic workflows are really a huge difference between it working versus not working."
"Concreteness buys you speed."
"Vague ideas tend to get a lot of kudos... but when you're concrete, you may be right or wrong."
"Gut instincts... can be actually a surprisingly good proxy."
"We're easily 10 times faster... building quick and dirty prototypes."
"Move fast and be responsible."
"It's time for everyone of every job role to learn to code."
"The ability to tell a computer exactly what you want... seems like it will remain the best way."
"Product management work... is increasingly the bottleneck."
"If you flip the wrong bit, you're not twice as slow... you spend like 10 times longer."
"Knowing all these wonderful building blocks lets you combine them... exponentially."
"AI is neither safe nor unsafe. It is how you apply it."
"Focus on building a product that people want, that people love."
"Don't worry about how much tokens cost... only a small number of startups... have users use so much."
"Education will be hyperpersonalized... but the workflow is still not clear."
"Look in your heart and if fundamentally what you're building... don't do it."
"Trying to bring everyone with us to make sure everyone is empowered to build with AI."
"Hype narratives do keep on getting amplified... used as a weapon against open source."

HABITS

Build one startup per month through hands-on involvement.
Focus exclusively on concrete, engineer-ready product ideas.
Pursue single hypotheses with full team determination.
Pivot decisively upon disconfirming data or evidence.
Rebuild codebases multiple times monthly as needed.
Encourage insecure prototypes for initial laptop testing.
Adopt latest agentic coding tools every few months.
Learn continuously via courses on AI building blocks.
Interview strangers in high-traffic spots for feedback.
Update gut instincts using A/B test data insights.
Architect software for easy model and provider switching.
Kill projects ethically despite strong economic viability.
Empower all staff, including non-engineers, to code.
Experiment rapidly with 20 prototypes per innovation cycle.
Stay atop AI trends to avoid technical blind alleys.
Use evals to benchmark and switch foundation models weekly.
Map complex workflows to agentic AI iterations.
Protect open source through advocacy against regulations.
Balance speed with responsibility in all executions.
Diffuse AI knowledge to bridge productivity inequalities.

FACTS

AI Fund co-founds startups at one per month rate.
Agentic AI trend emerged prominently over last year.
Prototypes with AI achieve 10x speed over production.
Production coding speeds up 30-50% via AI assistants.
Traditional PM-to-engineer ratio was 1:4 to 1:7.
New ratios propose 1 PM to 0.5 engineers recently.
GenAI building blocks list includes 15+ tools like RAG.
Deep learning courses teach combinatorial AI applications.
California bill SB 1047 proposed burdensome AI regulations.
Coursera Coach uses AI for effective student support.
Duolingo leverages AI for clearer language learning paths.
Hype narratives aided certain firms' fundraising goals.
Open source threats amplified false AI danger claims.
Mobile ecosystem hampered by Android-iOS gatekeepers.
Token costs rarely problematic for early-stage startups.
Agentic workflows integrate prompting, evals, guardrails.
Education AI experiments ongoing but end-state unclear.
AI Fund kills projects on ethical, not financial, grounds.
Marketing teams code to outperform non-coding peers.
Regulatory fights ongoing to preserve open source freedom.

REFERENCES

Y Combinator AI Startup School event in San Francisco.
AI Fund as venture studio building monthly startups.
Coursera platform and its AI Coach feature.
Deeplearning.ai courses and avatar chatbot.
Midjourney for generating background art images.
GitHub Copilot for code autocomplete.
Cursor and Windsurf as AI-enabled IDEs.
Claude using o3 for coding assistance.
Devin and Cloud Code as agentic coding tools.
OpenAI models and switching via evals.
Duolingo for AI-transformed language learning.
Khan Academy's Key Learning for K-12 education.
Jeff Bezos's two-way door vs. one-way door decisions.
Wall Street Journal article on AI losing control.
California SB 1047 regulatory bill on AI.
Jasper AI company facing business troubles.
Andrej Karpathy's mentions of AI coding.

HOW TO APPLY

Identify a specific problem in your domain and spend extended time thinking about potential solutions to build deep expertise.
Formulate product ideas with precise details, ensuring an engineer could implement them without ambiguity for immediate action.
Assemble a small team of subject matter experts to leverage gut instincts for quick feature and direction decisions.
Select one concrete hypothesis as the sole focus, allocating all resources to validate or falsify it rapidly.
Use AI coding assistants to prototype quickly, ignoring initial security and scalability for laptop-bound testing.
Build 20 low-fidelity prototypes in parallel to test multiple ideas, accepting high failure rates for speed.
Interview 3-10 strangers in coffee shops or lobbies respectfully to gather unbiased product feedback swiftly.
Run evals on new AI models weekly and switch providers seamlessly to maintain cutting-edge performance.
Analyze A/B test data not just for winners but to refine mental models and improve future gut decisions.
Learn one new GenAI building block monthly through courses, combining it with existing ones for novel apps.
Shift team ratios toward more PMs if engineering accelerates, prioritizing those with coding and product skills.
Architect software modularly for low switching costs between models, orchestration layers, and tech stacks.
Evaluate projects ethically beyond finances, killing those with potential societal harm despite viability.
Train all non-engineering staff in basic coding and AI prompting to boost cross-functional productivity.
Advocate publicly against hype-driven regulations threatening open source AI development and access.
Map existing workflows to agentic iterations, starting with outlining, research, drafting, and revision loops.
Experiment with AI in education by testing autograding, chatbots, and personalization on small scales.
Update technical decisions frequently to avoid blind alleys, consulting AI experts for architecture choices.
Pursue rapid feedback loops by alternating gut checks with stranger inputs before scaling to tests.
Rebuild entire codebases if a better tech stack emerges, treating it as a reversible two-way door decision.

ONE-SENTENCE TAKEAWAY

Prioritize execution speed with concrete AI ideas for startup success.

RECOMMENDATIONS

Focus on application-layer startups for maximum revenue potential.
Adopt agentic workflows for superior AI task outcomes.
Specify ideas concretely to enable engineer-led rapid builds.
Cultivate domain expertise for reliable gut-based decisions.
Test one hypothesis fully before pivoting resources elsewhere.
Prototype insecurely on laptops to accelerate initial validation.
Upgrade to latest agentic coding tools every few months.
Rebuild codebases freely as engineering costs continue dropping.
Teach coding to all roles for universal productivity gains.
Hone precise computer instruction skills through AI practice.
Increase PM ratios as engineering speeds outpace product work.
Use coffee shop interviews for quick, diverse user insights.
Learn AI deeply to sidestep costly technical missteps.
Combine GenAI blocks combinatorially for innovative apps.
Debunk hype narratives to maintain realistic AI views.
Apply AI responsibly, focusing on beneficial use cases.
Build products users love before worrying about moats.
Ignore early token costs; optimize only at scale.
Architect flexibly for seamless model and provider switches.
Experiment ethically, killing harmful projects proactively.

MEMO

Andrew Ng, the pioneering AI educator and entrepreneur behind Coursera and Deeplearning.ai, addressed aspiring founders at Y Combinator's AI Startup School in San Francisco on June 17, 2025. Drawing from his hands-on experience at AI Fund—a venture studio launching one startup monthly—he emphasized execution speed as the linchpin of success in the AI era. "Speed is a strong predictor of a startup's odds," Ng asserted, noting how AI tools now enable unprecedented velocity, from coding to customer validation. Yet, he cautioned, this acceleration demands concrete ideas over vague aspirations; specificity allows engineers to build prototypes 10 times faster, weeding out flops swiftly.

The talk delved into the AI stack, where Ng pinpointed applications as the richest opportunity, sustaining foundational layers like semiconductors and clouds through revenue generation. A pivotal shift, he explained, is the rise of agentic AI—iterative systems that outline, research, draft, critique, and revise, far surpassing linear prompting. At AI Fund, these workflows have transformed challenges like medical diagnostics and legal analysis from impossible to achievable. An emerging orchestration layer further simplifies app-building, but Ng stressed that true value lies in deploying these for real-world problems, not hype.

Ng shared tactical wisdom for rapid iteration: pursue one hypothesis doggedly, pivot on data, and leverage AI coding assistants like Claude and Devin, which evolve monthly. Engineering bottlenecks have eased—codebases now get rebuilt thrice monthly as costs plummet—shifting constraints to product management. He advocated flipping traditional ratios, proposing twice as many PMs as engineers, and arming all roles with coding skills. "It's time for everyone to learn to code," Ng urged, illustrating how art historians outperform novices in AI image generation through precise prompting.

Amid AI's promise, Ng debunked overhyped narratives—extinction risks, job obsolescence, nuclear-only compute—that serve promotional agendas. Safety, he clarified, stems from responsible application, not technology itself, and he decried efforts to regulate open source into oblivion via bills like California's SB 1047. In Q&A, Ng advised focusing on user-loved products before moats, empowering non-engineers with AI to curb inequality, and mapping education workflows to hyperpersonalized tutors, though the endgame remains experimental.

Ultimately, Ng's message was pragmatic optimism: AI democratizes building, but speed through concrete execution, ethical vigilance, and continuous learning will separate thriving startups from the rest. As opportunities explode at the application layer, founders must move fast, stay responsible, and protect open innovation to shape a flourishing future.

Regardez moins, lisez plus avec

Andrew Ng: Building Faster with AI