AI

If AGI is real, the question of who builds it and with what values is the most important question of the 21st century.

OpenAI's mission is to ensure that artificial general intelligence benefits all of humanity. That means sometimes our products will not be optimized for short-term revenue.

Our AI chatbot does the work of 700 customer service agents. That's not a talking point — that's our actual headcount reduction.

The average company wastes 5% of their spending on duplicate subscriptions, overpriced contracts, and inefficient processes. We find it and kill it.

I interned at OpenAI and researched at DeepMind. I learned that the best application of AI isn't chatting — it's knowing.

Search hasn't changed in 25 years. It's still a list of links. AI makes that obsolete.

If you can't beat AI, join it. Literally. With a chip in your brain.

AIP lets a supply chain manager ask a question in English and get an answer from a billion rows of data. That's not incremental. That's revolutionary.

AI doesn't kill RPA. AI makes RPA smarter. A robot that can also understand a document, read an email, and make a judgment call is exponentially more valuable.

The security questionnaire is the bane of every startup's existence. We're building AI that fills them out automatically. You're welcome.

Prior authorization was designed to reduce unnecessary care. Instead it became a system that delays necessary care. We're fixing that with AI that can read clinical evidence faster than any human reviewer.

AI in healthcare isn't about replacing doctors. It's about replacing the administrative burden that makes doctors want to quit medicine.

Everyone in this space says they have AI. We have peer-reviewed, published algorithms that predict treatment response. There's a difference between a marketing claim and a clinical tool.

AI is not just a feature you add to existing software. For the most interesting companies, it changes what's possible to build, who can build it, and how economics work.

Everyone in AI is trying to make software smarter. We said: make the chip bigger. Literally. Our chip is the size of a dinner plate. That's not a metaphor.

NVIDIA has the market. We have the physics. A single wafer-scale chip replaces thousands of GPUs. The math works. It just takes time for the market to catch up.

The AI boom created a GPU shortage. Every company is begging NVIDIA for chips. We offer an alternative. That's either terrible timing or perfect timing — depends on who you ask.

I was 20 years old when I co-authored Attention Is All You Need. I didn't fully understand what we'd built. Nobody did. It took five years for the world to catch up.

OpenAI wants to be the AI. We want to be the AI that enterprises actually trust with their data. Those are two very different companies.

We let companies run our models on their own infrastructure. Their data never leaves their servers. In regulated industries, that's not a feature — it's the only option.

The Transformer paper had eight authors. Three of us started AI companies. That might be the highest founder-per-paper ratio in history.

ChatGPT is a consumer product. It's impressive. But enterprise AI is a completely different game. Banks don't want a chatbot that hallucinates their compliance policies.

ChatGPT is great at answering questions about the internet. Glean is great at answering questions about your company. Those are very different problems.

We hit a $4.6 billion valuation. For an enterprise search company. Five years ago, investors would have laughed. Then AI happened and suddenly everyone understood why search matters.

Speed is the feature. When you can run inference 10x faster than anyone else, every AI application changes. Real-time conversations. Instant code generation. Things that feel sluggish on GPUs feel instant on Groq.

Everyone is fighting over who has the best AI model. We're fighting over who can run any model the fastest. That's a better fight to be in.

I invented the TPU at Google. The Tensor Processing Unit. Then I left because I knew I could build something faster. That chip is the LPU.

OpenAI invested in us directly. Sequoia led our Series B. We went from founding to $700 million valuation in 18 months. Legal AI isn't a niche. It's one of the biggest markets AI will ever touch.

I was a research scientist at DeepMind and Meta AI. Winston was a lawyer. The combination sounds weird until you realize legal AI needs both — someone who understands the models and someone who understands the law.

Lawyers bill $1,000 an hour to read documents. AI can read them in seconds. The math is so obvious it's almost embarrassing that it took this long.

I was a corporate lawyer at O'Melveny. I spent 80% of my time on work that didn't require being a lawyer. Research, document review, drafting boilerplate. Harvey does that 80%.

AI should not be controlled by three companies in San Francisco. Open source is how we keep AI democratic. That's not idealism — it's survival strategy for everyone who isn't OpenAI or Google.

Google, Meta, Microsoft, Amazon, and Salesforce all invested in us. Every big tech company. At the same time. That doesn't happen unless you're building something everyone needs.

We host over 500,000 models. Half a million. Every researcher, every startup, every company experimenting with AI — they come through us. We're the infrastructure layer nobody talks about.

OpenAI closed their models. We opened ours. They chose profit. We chose community. The market will decide who was right, but I know which side has more developers.

We started as a chatbot app for teenagers. It failed. Then we open-sourced our NLP library and accidentally became the GitHub of machine learning. Best pivot in AI history.

AI will be the most transformative technology in human history. More than the internet. More than electricity. If you don't believe that, you're not paying attention.

The AI race isn't about who builds the smartest model. It's about who builds the model people actually want to talk to. Empathy is the moat, not benchmarks.

Reid Hoffman and I started Inflection together. A DeepMind co-founder and a LinkedIn co-founder building an AI company. The investor pitch meetings were very short.

We raised $1.3 billion and built one of the most powerful AI models in the world. Then I left to run Microsoft AI. People called it a failure. I call it a $650 million acqui-hire that changed the industry.

Pi is not an assistant. It's not a search engine. It's a conversation partner. The difference matters. People don't want a tool. They want something that listens.

We raised $125 million at a $1.5 billion valuation in October 2022. ChatGPT launched in November 2022. One month. The timing was either perfect or catastrophic depending on which quarter you look at.

We hit $45 million ARR in 18 months. Fastest-growing company in Y Combinator history at the time. Then ChatGPT launched and the whole category changed overnight.

We laid off people. We restructured. We pivoted to enterprise. Everyone wrote us off. But we're still here, still growing, and still profitable. In AI, survival is underrated.

When your underlying technology becomes free and commoditized, you have two choices: die, or move up the stack. We moved up the stack.

Marketing teams don't want a chatbot. They want a tool that knows their brand, writes in their voice, and plugs into their workflow. That's what separates enterprise AI from a toy.

We were a wrapper on GPT-3. I know how that sounds. But the wrapper was the product — templates, brand voice, workflows. ChatGPT gave away the raw model. We sold the workflow.

Our models are smaller and faster than GPT-4 for most tasks. Bigger is not always better. Efficiency is the real breakthrough — not throwing more GPUs at the problem.

Europe has the talent. We have some of the best AI researchers in the world. What we didn't have was a company willing to compete with OpenAI head-on. Now we do.

We open-sourced our first model via a torrent link. No blog post. No press release. Just a magnet link on Twitter. The AI community lost its mind. That was the point.

I left DeepMind because Google moves too slowly. Guillaume left Meta AI because Meta moves too cautiously. We started Mistral because we wanted to move at startup speed with frontier-lab talent.

We raised $113 million with no product, no revenue, and a four-week-old company. Just a PDF and three ex-DeepMind and ex-Meta AI researchers. Europe was desperate for its own AI champion. We volunteered.

We went from founding to $2 billion valuation in six months. In France. People keep saying Europe can't do AI startups. I think we just did.

Big banks use one credit model for everyone — a dry cleaner gets the same analysis as a hotel chain. We built bespoke models for every sector. A hotel has seasonality, occupancy rates, booking patterns. Our AI understands that.

We raised $141 million from Google, NVIDIA, and Salesforce. All of them want to own AI video. None of them have a product as good as ours. When your investors are also your would-be competitors, you know the tech is real.

Hollywood is terrified of us. And they should be. Not because we're replacing filmmakers — we're replacing the $50,000 VFX budget that prevents most filmmakers from even trying.

We co-created Stable Diffusion. Then we built Gen-2, the first commercially usable AI video generator. Text to video. Type a sentence, get a movie clip. That's not incremental progress. That's a new medium.

Everything is Alive won an Emmy. It was made with Runway. An AI-assisted film won an Emmy. That's not the future. That already happened.

Everyone is fighting over training. We focused on inference — running the models, not building them. Training happens once. Inference happens a billion times a day. The economics favor inference.

We built a chip from scratch specifically for AI. Not a GPU repurposed for AI — a chip designed from the ground up for dataflow computing. That distinction matters when you're running trillion-parameter models.

AI is only as good as its training data. Everyone obsesses over the model. We obsess over the data that feeds it. Scale provides the labeled data for OpenAI, Meta, the US military — all of them.

The Pentagon is our customer. The US Army, Air Force, and intelligence community all use Scale to process data. National security runs on good data. We provide it.

I'm a CPA. I know bookkeeping. I genuinely believed AI could automate it. But the technology wasn't there yet, and instead of admitting that, we hired more humans and called them AI. That was the mistake.

We told investors AI was doing the bookkeeping. Investors loved it. The problem was, humans were doing the bookkeeping. The AI wasn't ready. We shipped the marketing before we shipped the technology.

Stable Diffusion was downloaded 10 million times in two weeks. Ten million. It powered thousands of apps. And we captured almost none of that value. Turns out giving away your core product makes monetization tricky.

We open-sourced Stable Diffusion and gave it away for free. OpenAI charged for DALL-E. We gave it to the world. That decision made us famous. It also made it very hard to build a business.

Most companies do not need to train their own model from scratch. They need fast, cheap inference on the best open models. That is the real market.

The future of AI is open source. Closed models will always have a place, but the ecosystem that wins is the one everyone can build on.

The cost of running AI models drops by half roughly every year. That means what is expensive today becomes accessible tomorrow. We want to accelerate that curve.

We are building the cloud for AI. Not a general-purpose cloud — one specifically optimized for training and running large models.

FICO scores are a 30-year-old technology pretending to be modern. They ignore education, employment, and dozens of variables that actually predict whether someone will repay a loan.

The credit system in America is broken in a very specific way — it punishes people for being young, for being new to the country, for not having a history. Our AI fixes that.

I spent 7 years running Google's enterprise business. The thing I learned is that AI is not magic — it is math applied to better data. Lending just happened to have the worst data practices of any industry.

We approve 27 percent more borrowers than traditional models and at lower interest rates. That is not a small improvement. That is a fundamentally different outcome for millions of people.

The AI gold rush needs picks and shovels. Everyone wants to build the model. We build the infrastructure that makes the model actually work in production.

OpenAI, DeepMind, Toyota, Samsung — they all use us. When the best AI teams in the world choose your tool, you know you are building the right thing.

Grok is the anti-ChatGPT. It will answer spicy questions that other AIs are too politically correct to touch. An AI that is afraid to say anything interesting is useless.

AI will be smarter than any single human by next year and smarter than all humans combined by 2029. That is either the best thing or the worst thing that will ever happen. Probably both.

The goal of xAI is to understand the true nature of the universe. That sounds grandiose but every other AI company is just trying to sell ads or automate customer service.

I co-authored "Attention Is All You Need" — the paper behind every modern AI model. Then I pivoted to blockchain. People thought I was crazy. But I saw that decentralized AI needs decentralized infrastructure. The two are inseparable.

Our dashcams do not just record accidents. They coach drivers to prevent accidents. That is the difference between a camera and an AI platform.

AI is the new electricity. Just as electricity transformed every industry 100 years ago, AI will transform every industry now.

The answer is not to fight AI. The answer is to use AI better than anyone else. CheggMate is our bet that human-verified plus AI equals something neither can do alone.

We built the world's largest homework help database. Then AI learned to do the same thing in real time for free. That is a problem.

ChatGPT did to Chegg what Chegg did to textbooks. Disruption is only fun when you are the one disrupting.

Our stock dropped 48% in one day because I mentioned ChatGPT on an earnings call. The market priced in our disruption before we even felt it.

The GPU was designed for gaming. It turns out the math you need for photorealistic explosions is the same math you need for artificial intelligence. Lucky us.

The iPhone moment for AI has arrived. Everything that happened in mobile over the last 15 years is about to happen in AI over the next 5.

Google invented the transformer — the architecture that powers GPT. We should have productized it faster. We are fixing that now.

The hardest part of running Alphabet is that you are defending a $175 billion ad business while simultaneously trying to reinvent it with AI. Both things have to be true at once.

Microsoft missed mobile. We will not miss AI. That is the one sentence that explains every decision I have made since becoming CEO.

The $13 billion OpenAI bet was not a bet on AI. It was a bet on building the platform that AI runs on. That is what Microsoft does — platforms.

AI is the most profound technology humanity will ever develop. More than fire. More than electricity. We need to get it right.

Hivemind is not just a drone brain. It is an AI pilot that can fly any aircraft — a quadcopter today, an F-16 tomorrow. The platform is the product.

The future of air combat is unmanned. Not because humans are bad pilots, but because AI can make decisions in microseconds that humans cannot.

Our drones fly without GPS, without a pilot, and without communication. The AI figures out the environment in real time. That is autonomy, not remote control.

We are the number two in endpoint security. CrowdStrike is number one. But number two with autonomous AI will eventually beat number one with human analysts.

We built Darktrace on a simple idea: instead of looking for known threats, learn what normal looks like and flag everything that is not normal.

A security platform that needs humans to function is not a platform — it is a consulting service with software attached.

The future of cybersecurity is autonomous response. Humans cannot respond fast enough. By the time an analyst sees an alert, the attacker has already moved laterally.

Every organization has a digital immune system whether they know it or not. Darktrace makes it visible, intelligent, and autonomous.

AI can generate a website in 30 seconds. The question is whether that website is any good. We are betting the answer is yes.

AI will be more transformative than the internet. I have reoriented my entire investment strategy around that belief.

Healthcare administration is messy. AI that is 95 percent accurate means 5 percent errors. In healthcare 5 percent errors means denied claims and delayed care.

We called it an AI employee. It was an AI assistant that needed human supervision. The marketing wrote checks the product could not cash.

AI in healthcare requires 99.9 percent accuracy. We were at 95. The gap between 95 and 99.9 turned out to be unbridgeable.

AI is driving the biggest infrastructure buildout since the internet. Every AI model needs servers. We build the servers.

We were one of the first to turn GPT-3 into a product. Then OpenAI released ChatGPT and did it for free. Welcome to AI.

Every AI wrapper company has the same existential question: what happens when the foundation model company gives your product away free?

AI in healthcare works when it handles paperwork not patients. The moment you try to replace clinical judgment you fail.

The AWS Lambda moment for AI. Run a model pay per prediction. No GPUs to manage. That is Replicate.

NVIDIA invested because they understand that healthcare AI needs specialized models not general-purpose ones.

Sam Altman invested personally. AI will manage the flight systems that make hypersonic travel safe and efficient.

a16z invested because they believe AI-powered hardware will define the next decade. We are the proof point.

Beyond Meat engineered plant-based food by hand. We let AI find combinations no human would think of. That is the difference.

We named our AI Giuseppe after an Italian artist who painted portraits using fruits and vegetables. It felt right.

Natural language processing is the bridge between human communication and machine understanding.

The best AI voice is the one you can't tell isn't human.

AI democratized image creation. AI democratized text creation. Music is next.

An AI avatar doesn't call in sick, doesn't need makeup, and speaks 130 languages. That's hard to compete with.

AI won't replace programmers. It will turn everyone into a programmer.

We're not replacing human creativity. We're replacing the expensive, boring parts of video production.

The future of software development is collaborative, cloud-based, and AI-powered. Replit is all three.

As AI gets better, proving you're a real human becomes more valuable. World ID solves that.

AI can underwrite insurance better than humans. We just have to prove it with the numbers.

We're building proof of personhood for the AI era. If you can't prove you're human, you can't participate.

In six months we went from a demo to a product that millions of people use. AI moves at a different speed.

We dropped out of Stanford because AI video generation was moving too fast to wait for a PhD.

We're not building AI to replace filmmakers. We're building AI so that everyone can be a filmmaker.

AI can analyze medical records and identify cancer risk faster than any doctor. Not to replace doctors — to help them.

AI will be the most transformative technology in human history, but we should approach it with both excitement and caution.

Every student deserves a personal tutor. AI might finally make that possible.

Whoever controls the AI infrastructure layer will control the next decade of technology.

Search organized the world's information. AI will understand it.

AI will reshape every industry. The question is not if, but how fast.

GitHub Copilot proved that AI could write production code. That was the moment everything changed.

AI talent is the scarcest resource in the world. Finding it early is the highest-leverage activity.

Open source is the most powerful force in software. AI will be the same.

Go was supposed to be unsolvable by AI for another decade. AlphaGo solved it in 2016.

AI chips are not one-size-fits-all. Different workloads need different architectures.

Our mission is to solve intelligence and then use it to solve everything else.

ChatGPT changed the conversation. We had to pivot from AutoML to AI governance — fast.

Enterprises want both: RPA for structured tasks, AI for unstructured ones. We bridge that gap.

We're evolving from where is my shipment to what should I do about it. From visibility to intelligence.

From digital flashcards to AI tutor. Same mission — making studying effective — completely different product.

BoweryOS optimizes every variable — light, water, nutrients, harvest timing. AI makes farming precise.

Google uses machine learning to optimize our plasma. AI is accelerating fusion research.

Q-Chat turns flashcards into conversations. AI transforms passive studying into active learning.

AI is going to transform how every company manages their content. Box is going to be at the center of that transformation.

AI is not going to replace all jobs. But it is going to replace a lot of jobs that people pretend it won't.

AI safety is the most important product problem of our generation. That's why I joined Anthropic.

We process petabytes of data daily. A human trader could not analyze what our systems process in a single second.

The future of investing is machines making decisions faster and better than humans. We are building that future.

India's first AI unicorn in under a year. The world underestimates what Indian engineers can build.

Artificial intelligence will change the nature of work faster than any technology we have ever deployed at scale. The question is not whether — the question is whether we are ready.

The Internet is the world's largest ungoverned space. Governments are going to have to deal with that, but so will businesses.

AI will change everything about software, but the companies that win will be the ones with proprietary data, not proprietary models.

The SaaS playbook is well-understood now. The AI playbook is being written. That’s where the alpha is.

We use AI to vet developers so companies do not have to spend months interviewing. Our system evaluates skills in hours, not weeks.

AI won't replace lawyers. But lawyers who use AI will replace lawyers who don't.

We've trained our AI on millions of legal documents. It's read more contracts than any lawyer alive.

The legal industry spent decades digitizing paper. Now it's time to actually make the digital documents intelligent.

Autopilot doesn't just review contracts. It negotiates them. That's a fundamentally different capability.

AI in ediscovery isn't about replacing reviewers. It's about making sure humans spend their time on documents that actually matter.

The foreman says everything is on track. The AI says 23 tasks are behind schedule. Who do you trust?

AI doesn't just adjust the weight. It knows when you're struggling at rep 8 and reduces the load by 5 pounds so you can finish the set with proper form.

Hundreds of millions of people in Nigeria and India need credit and can't get it. Their phone data can predict creditworthiness better than a FICO score ever could.

Chess taught me to think deeply about complex problems. Neuroscience taught me how the brain does it. AI is about building both.

AlphaFold is a proof point that AI can accelerate scientific discovery by orders of magnitude.

Data is the moat. Every customer interaction makes the algorithm smarter. You cannot replicate years of data overnight.

AI for social good is not a contradiction. The same techniques that power consumer apps can be turned toward health, agriculture, and education in the developing world.

Poe is not about picking a winner in the AI race. It is about giving users the best experience no matter which model delivers it.

Spatial intelligence is the missing dimension of AI. Current models understand language and generate images. They do not understand the physical world.

AI is not going to replace humans. But humans using AI will replace humans not using AI.

There is nothing artificial about AI's impact on real people and real lives.

If we want machines to think, we need to first give them the ability to see. ImageNet was my answer to that problem.

We are at an inflection point. The question is not whether AI will transform every industry. The question is whether the transformation is equitable.

Decentralized AI is the only way to prevent the technology from becoming a tool of control rather than liberation.

Every child deserves a brilliant friend who has the knowledge of a doctor, lawyer, and financial advisor. AI can be that friend.

Generative AI in the creative and in the art space is going to augment human ingenuity and not replace it.

This is the perfect time for security companies to make sure customers can deploy AI bravely.

The rise of AI and automation is going to be the defining trend of the 21st century and will be a disruptive force in society.

I believe that enterprises have been watching and wanting to adopt AI, but I don't think they've been able to really take full advantage of it.

AIs are only as good as the data they are trained on.

The breakthroughs in model quality sparked a new wave of generative AI, and that will have an incredible impact on cryptocurrency.