Over the next five years, every professional will gain AI 'interns' — e.g. each physician getting five fresh-MD-equivalent assistants — increasing productivity; but these interns will eventually surpass the professionals' expertise, at which point the change becomes catastrophic and impossible to roll back.

In 1996 Khosla backed Juniper to bet on TCP/IP for the public internet against universal expert consensus (telcos and Cisco's CTO insisted on ATM and refused TCP/IP above 155 Mbps), because the data showed TCP/IP usage climbing exponentially near the knee of the curve — the bet returned $7 billion on a $3 million investment (~2,500x).

Khosla's public-transit startup designed self-driving two-to-four-person vehicles that fit in a bicycle-lane width yet provide ~10x the passenger throughput of light rail for the same street width, are hailed on-demand like Uber, are always cheaper and faster than today's public transit, and can be deployed via prefab raised infrastructure rather than years-long tunnel digs.

A multi-center Stanford study on complex diagnoses found physicians alone were 73% accurate, AI alone was 88% accurate, but when AI was given to doctors the doctors only improved to 76% while the AI degraded from 88% to 76% — illustrating that human involvement can drag down AI performance.

The binding constraint on humanoid robots is intelligence, not hardware — China has produced impressive hardware form factors (boxing, running), but they aren't learning robots and fail when the environment changes; the breakthrough needed is intelligence that learns and adapts without being programmed for each task.

Most jobs on the planet are not really jobs but 'servitude' — repetitive survival labor like picking lettuce for 40 years or mounting tires on an assembly line — and freeing people from this servitude to reclaim their time would be a great thing.

The dystopian AI outcome is a choice society makes, whereas the utopian world of abundance will happen anyway — it is only a matter of when, not if.

Large innovation almost never comes from incumbents in the field — across retailing (Amazon vs Walmart), media (Netflix/YouTube/Facebook vs NBC/CBS), transport (Uber, Airbnb), space (SpaceX vs Boeing/Lockheed), biotech (Genentech), and COVID vaccines (Moderna/BioNTech vs big pharma) — because experts extrapolate the past while founder-driven companies have permission to try crazy things.

The 2030s will see a faster rate of demise of Fortune 500 companies than ever before, because most Fortune 500 companies should not be run even remotely the way they're running by 2040, and the transformative reinvention won't come from existing companies — someone new will reinvent each industry.

Calling oneself 'founder friendly' is a real disservice to founders — like telling a five-year-old you'll always say yes — whereas the right approach is teaching founders to make decisions and lengthening the leash over time, giving them the final say while always challenging and debating them.

By around 2040 there will be a billion bipedal robots doing more total work than human labor today — robots that are smarter, never sleep, never complain, and aren't constrained by narrow specializations (the same robot that cleans your house could be your doctor).

Symptom-based medicine will disappear: instead of detecting disease from symptoms, we will detect disease 10 years before symptoms manifest based on what's happening in the body, and providing far more up-front care reduces downstream disease burden and expensive interventions like heart surgery and ER visits.

Even the strongest founders benefit from a debating partner who challenges them but does not make decisions — like a pro athlete with a coach — which is why Khosla refuses to sit on boards or take governance votes, instead doing regular dinners/quarterly sessions to challenge founders while always granting them the final vote.

When we invested in Square in 2010, Jack Dorsey had just been fired from Twitter and had four failed startups on his Wikipedia, so traditional wisdom would say don't back him.

Almost certainly within the next 5 years, any economically valuable job humans can do, AI will be able to do 80% of it, with a few exceptions like heart surgery or brain surgery — meaning 80% of 80% of all jobs can be done by AI, which will make the world a very different place.

The next 5 years (to 2030) will look like productivity improvements — GDP going up and accelerating — but by 2040 and beyond we enter an era of abundance so large it's hard to imagine, where the need to work goes away and people work because they want to, not to pay their mortgage.

The capitated/Medicare Advantage segment of healthcare will be transformed by AI far more effectively than fee-for-service, because for a hospital system reducing cost means reducing revenue, whereas in a capitated system providers benefit from AI-driven cost reduction since they get paid a fixed amount per patient.

Movies and content will eventually be produced for $100,000 instead of $100 million using AI, and anyone who signed restrictive deals like the Screen Actors Guild agreement will go out of business because they locked themselves into an order-of-magnitude less effective content-generation paradigm.

Khosla has high confidence (above 80%, greater than 50% by 2050) that a world of abundance happens within our lifetimes, and is essentially certain it happens before 2060 — timing is hard to predict but the direction is not.

Sentient AI killing humanity is a real worry, but it should be assessed within the full basket of risks humanity faces (asteroids, far worse pandemics that could kill 700 million rather than COVID's ~7 million), and it is a smaller risk than authoritarian leaders like Xi or Putin gaining AI dominance.