What Actually Makes A Startup Durable

Predictions and Expectations:

• The cost of intelligence is expected to drop by a factor of 10x per year for the same level of intelligence.
• Within six to 12 months, the economics of AI are expected to make it more cost-effective than human engineers for all programming tasks.
• There is no chance that a human is more effective than a model for any single programming task today if using the best available intelligence.
• It is predicted that every tool and process can be built internally and instantly within the next months.
• Pure software startups are predicted to be "dead" or easily replicable without hard moats like sales, regulation, or physics.
• A "hard bit" (e.g., brutal B2B sales, regulatory barriers, or hardware physics) is expected to be the primary source of durability for startups.
• Companies are expected to compress from thousands of employees down to below Dunbar's number (around 100 or 50 people).
• It is predicted that AI will make it possible for an average engineer to become a great engineer and for someone who cannot sell to become good at selling.
• AI will not be able to replace founder well-being, community building, or the "power of witnessing" (having someone acknowledge the user's journey).
• The industry is expected to see a surge in demand for AI competitors and sovereignty over AI inference due to US export restrictions.
• It is expected that pure software startups will be less durable and defensible compared to those attacking harder problems requiring significant capital.
• Most AI output is expected to be bad, but feedback loops with AI will make results game-changing for companies within weeks.
• It is expected that software will no longer be the "hard bit" of building a company, making distribution one of the few remaining hard things.
• It is predicted that the fastest way to fail is to stay comfortable in research and not talk to customers.
• The "grass" is not expected to be greener when pivoting from something deep to something shallow due to hidden problems.
• It is expected that high agency is a trait that can be improved through building difficult but tractable projects.
• There is an expectation that founders will be terrified of facing the market because of loss aversion regarding previous work.

Timelines and Milestones:

• Cost reductions in intelligence are expected to make humans obsolete for programming tasks in "six or 12 months."
• Automatic model routing and zero-cost queries for users are expected to occur within "12 months."
• Some companies are expected to reach millions in revenue "in just a few months" or "by the end of WC."
• Deep tech projects (e.g., nuclear reactors) are expected to require capital of "800 million dollars" next year.
• The speaker mentions that a solo founder experiment resulted in a crisis of confidence "six weeks into the batch."
• Ambitious startups are expected to be "10 years" ago where writing software was the hard bit.
• Founders are expected to re-assess goals and hold themselves accountable "every two weeks."
• The speaker suggests that companies built with AI will be "incredibly fast-paced" compared to "back in our times."

Technology and Product Direction:

• The strategy is to shift from pure software to products with "hard bits" such as regulated industries, hardware, or physics.
• The direction is to use AI to empower partners and share superpowers with founders, with the organization building tools to support this rapidly.
• The focus is expected to shift from writing code to "witnessing" and providing community support that AI cannot replace.
• The future of AI interaction is predicted to be fully automatic routing behind the scenes, with users just typing queries without perceiving the underlying models.
• The approach to AI-native companies involves picking a single narrow loop (e.g., automating follow-up emails or prototyping ideas post-call) rather than trying to implement everything at once.
• The direction is to stop delegating "talking to customers" as it is the last thing to automate to maintain context and close the loop.
• The trend is to view "AI wrapper" as a meaningless derogatory term, noting that all current successful ideas have AI at their core.
• There is a shift in belief that technical co-founders do not necessarily need a complementary business co-founder; two deep technical founders are preferred.

Market and Industry Outlook:

• A global hub (specifically San Francisco) is expected to attract the most ambitious people and companies with the biggest visions.
• It is expected that there will always be other regional hubs (London, Paris) but they may lack the critical mass of global hubs.
• The market is expected to respond to US export restrictions by creating demand for sovereignty and competitors to US models.
• It is predicted that pure software is "too easy to replicate" now, reducing the durability of such businesses.
• There is an expectation that capital needs will rise commensurately with the ambition of the problems solved (e.g., curing disease, infinite energy).
• The speaker expects that "more companies are being built" at early stages, making distribution the critical differentiator.
• It is predicted that the "hard bit" for startups will shift from software implementation to sales, regulation, or physics.
• The market for early-stage startups is expected to be highly empirical, requiring rapid hypothesis testing rather than theoretical planning.

Company Plans:

• YC plans to continue funding solo founders but expects higher success rates with co-founders due to the utility of support during low morale.
• The organization plans to experiment with "virtual partners" and office hours using AI while maintaining human partners for deep, tailored advice.
• YC plans to re-evaluate the requirement for a deep software engineer on founding teams, testing if "vibe coders" can succeed.
• There is a plan to fund founders regardless of the idea they apply with, betting on the team's ability to figure out durable moats.
• The company plans to help founders overcome the tendency to research too much by forcing them to "launch early and often."
• There is a plan to provide "25K plus" in credits to student founders to offset the cost of API usage.
• The organization intends to keep reassessing ambitious goals "every two weeks" and help founders overwhelm the biggest bottleneck.

Financial Guidance:

• Founders are advised that they may not need to raise as much capital today as they did "five years ago" to achieve similar results.
• However, ambitious projects (like nuclear reactors) are expected to require massive capital, with one example needing "800 million dollars."
• It is predicted that using VC money to buy tokens and go faster may still be worth it even for software companies.
• The speaker notes that a "one person billion dollar company" is theoretically possible but economically illogical due to the marginal cost of adding a second person being low.

Risks and Caveats:

• There is a risk that founders will be "blindsided" if they only look at AI's capabilities and not what it cannot do (well-being, community).
• A risk exists that founders will pivot not due to evidence but due to "sadness" or lack of enthusiasm, leading to worse outcomes.
• There is a risk of "loss aversion" preventing founders from talking to the market and risking the realization they are building the wrong thing.
• The risk of running a single-person company is cited as lower statistical success and higher vulnerability to confidence crises.
• There is a risk that "pure software" startups will be killed by replication due to the ease of implementation.
• A caveat is raised that advice on fast pacing applies less to deep tech, space, or physics-heavy projects which have different timelines.
• There is a risk of over-relying on AI for "talking to customers" or "talking to co-founders," which must remain human.
• The speaker warns that finding the "hard bit" is difficult and that "atoms" in hardware or "regulatory barriers" are not easy to overcome.

Confidence and Disagreement:

• The speaker "is not sure" if the cost efficiency argument holds today, but "expects" it to be true in 6-12 months.
• There is "not one single programming task" that a human is more effective for today, according to the speaker's strong belief.
• The speaker "thinks" it is illogical to run a single-person company unless breaking a record.
• There is a "misconception" that YC advice is a static database; partners actually provide deep, tailored advice that cannot be replaced by AI.
• The speaker is "not sure" if pivoting has become cheaper, suggesting a general answer is difficult.
• There is a "30 percent" internal disagreement rate at YC regarding which companies to fund.
• The speaker "does not think" high agency can be taught directly but believes it can be trained through difficult projects.
• There is a belief that "AI will not be able to do" certain things like providing emotional support or the "power of witnessing."
• The speaker "expects" that most founders have a bias toward building/researching rather than talking to customers.