Skip to content

By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

AnthropicAnthropic

The making of Anthropic

Anthropic’s co-founders on why they started the company, what they’ve learned, and the work still ahead.

Nov 19, 202614mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

Great conversations run long. Their best ideas don't have to.

uListen is free on iPhone, iPad, Android, Chrome, and Claude.

See how

Transcript

  1. 0:02 – 0:21

    The question: why start Anthropic, and why it felt impossible to walk away

    1. SP

      So why did you all start Anthropic?

    2. SP

      [gentle music] It still feels like then in that we knew a secret the rest of the world didn't know. I think if you truly see what this is, you can't ever walk away.

  2. 0:21 – 1:00

    Founders’ winding paths into AI (physics, medicine, journalism, development, programming)

    1. SP

      I showed up at Princeton as a PhD student who wanted to do theoretical physics.

    2. SP

      I was a professor at Johns Hopkins before Anthropic.

    3. SP

      We're all envious of the great discoveries of the past.

    4. SP

      My PhD thesis is in theoretical physics.

    5. SP

      Out of college, I went and worked in global development.

    6. SP

      I was a journalist before I did any of this, and I studied data centers obsessively. It's gonna get weird.

    7. SP

      I guess I thought of myself as a, a mathematician and programmer.

    8. SP

      I got into AI back in 2016, and at the time, like, an AI couldn't talk.

    9. SP

      After I graduated, I went to Stanford and did my postdoc in cancer biomarkers.

  3. 1:00 – 1:27

    Impact motivation: using computer science and AI to make the world better

    1. SP

      From a very early age, Dario and I were really curious about how to make the world better.

    2. SP

      I was very interested in questions of how we could use computer science as a tool that is good for the world.

    3. SP

      I was a little bit frustrated with the physics research I was doing, and AI seemed exciting and like a breath of fresh air.

    4. SP

      2013 or 2014, I first started seeing the big results in AI, then wondered if advancing this technology would be the best way to advance biology.

  4. 1:27 – 2:33

    2018 at OpenAI: scaling language models and the “big blob of compute” insight

    1. SP

      We saw this connection between the level of impact we were trying to have in the world and AI's potential to really achieve a lot of those goals. Back in 2018, the future co-founders of Anthropic were doing a lot of sort of different and disparate things.

    2. SP

      A group of us all at OpenAI, we're working on training models, we're working on policy, we're working on interpretability, we're working on safety.

    3. SP

      I started working on these language models where you try to predict the next word.

    4. SP

      A lot of folks thought at the time, like, oh, you're going to need to teach the network explicitly how to plan and how to reason and how to use tools.

    5. SP

      Dario noticed the trend, which he called the big blob of compute.

    6. SP

      Just basically making a bigger machine, making a bigger computer, and jamming more data into it.

    7. SP

      When you put in, you know, more compute and more data, you get increasing intelligence, and it's on this sort of predictable line.

    8. SP

      We were just seeing, like, each time you scale it up, these new behaviors emerge, these new abilities emerge.

  5. 2:33 – 3:58

    Realizing the stakes: transformative promise alongside frightening risk

    1. SP

      At which point you realize, wow, if this continues, we might be on the precipice of, like, a totally new reality, a totally new form of technology.

    2. SP

      Human-level intelligence, human-level AI is not, like, 100 years in the future, but is actually within our lifetimes.

    3. SP

      These kinds of very powerful models could cause enormous strides in healthcare, in robotics.

    4. SP

      Computer programs, novel scientific ideas, mathematical theorems.

    5. SP

      Maybe we could reduce cancer rates by 90%. Maybe heart disease, obesity, all of the major infections will just be a thing of the past.

    6. SP

      Any of the problems of the world that are bottlenecked by human ingenuity, suddenly you have much more ability to solve them.

    7. SP

      Seeing the potential progress in AI was exciting. At the same time, I think it was scary.

    8. SP

      If you're building something that's more powerful than yourself or could grow to be more powerful than yourself, the potential risks of that are enormous, no matter how you think about it.

    9. SP

      I don't think I understood the full picture yet. I don't think anyone did, but I definitely believed that there could be big safety and societal effects. This was a technology that if it had any analogy at all, which I, I think it mostly didn't, it would be more analogous to pharmaceuticals or unlocking the secrets of the atom, things that have immense value, but also can, can go wrong. We really needed to build it in the right way.

  6. 3:58 – 4:26

    Founding decision: an organization designed for safety and responsibility

    1. SP

      Just this huge responsibility that you feel. I'm one of the very few people in the world that understands what the consequences of this technology are going to be.

    2. SP

      The thing we most wanted to do was work together in an environment with this deep commitment to safety.

    3. SP

      We decided we had to build Anthropic because we had to build an organization fit for the demands this technology was gonna place on all of us.

  7. 4:26 – 5:17

    Choosing the structure: becoming a public benefit corporation

    1. SP

      One of the first big decisions we had to make was just what kind of company we wanted to be. It was a more involved and complicated discussion than you might expect.

    2. SP

      We knew that we had to be able to build some of the largest computers on the planet, train frontier technology, find ways to deploy that technology into the world, and be honest about the difficult aspects of the technology as well as its benefits.

    3. SP

      And we eventually landed on public benefit corporation.

    4. SP

      That would set us up as a company that was not just fully governed by economic incentives, but also by the incentive to do right by the world.

    5. SP

      There are going to be times when we're gonna want to prioritize the mission at the cost of profits or at the cost of providing the maximum value to a shareholder.

    6. SP

      So we incorporated as a public benefit company on January 26, 2021, and then it was off to the races, time to build this thing.

  8. 5:17 – 6:06

    Naming Anthropic: spreadsheet chaos, “human-centered” meaning, and the domain price

    1. SP

      This is a list of names people proposed for what we could call Anthropic.

    2. SP

      Here are some names I listed apparently. Squirrel, owl, whale, differentiated, humane.

    3. SP

      Anthropic was contributed by Jared.

    4. SP

      I realized that since anthropic means kind of human-oriented or human-centered, that it seemed like a good name for an AI company that wants AI to be developed in the best interests of humanity. Then we looked at domain names that we could get. Anthropic.com was, I think, by far the cheapest. I thought that was a pretty good argument to go with Anthropic.

    5. SP

      And there is a comment in the spreadsheet that says, "We like the name. It is good." And, "Is domain available? Yes." [laughs]

  9. 6:06 – 6:27

    The scrappy early days: AWS clusters, first website, fundraising, recruiting, and payroll

    1. SP

      It was so chaotic at the beginning

    2. SP

      Everyone was doing extremely unglamorous schlep

    3. SP

      Tom was the one who had to go get us the AWS clusters

    4. SP

      Me and Chris Olah built the first website

    5. SP

      Dario focused on fundraising, and I was working on recruiting

    6. SP

      And there was briefly a few months where Jared Kaplan, our chief scientist, ran payroll

    7. SP

      That tells you the level of professionalism of the early operation

  10. 6:27 – 7:35

    From research lab to aligned assistants: building LLMs and debating product direction

    1. SP

      I think there was a lot of work of us sort of just getting oriented to what kinds of things exist in a language model. You find all these just, like, beautiful things inside that we didn't know about going in

    2. SP

      It was really nothing other than a research lab. We're just producing technology that we think might be really, really powerful and economically valuable, but might also just be a, a curiosity or a science experiment

    3. SP

      One thing I remember was every day just, like, waking up feeling like really excited to get to work on the most fun project that I could imagine with my best friends

    4. SP

      It felt great to be with people who shared our values so deeply. Also, just getting to build all the infrastructure from scratch was a lot of fun

    5. SP

      We were trying to build an organization focused on training large language models, aligning those models with our values. That was why we founded Anthropic

    6. SP

      There was a lot of debate at the time of do we wanna focus more on building something that's more of a pure chatbot, or is, like, software engineering the thing that we need to focus on? In the end, we landed on, like, actually you just need to do both of those really well. Cool, thanks.

  11. 7:35 – 8:19

    Early assistant vision: “helpful, harmless, and honest” (and early failures)

    1. SP

      This was a, a document I wrote in 2021 as a proof of concept. We were working to develop a general AI assistant that can interact conversationally with a user to create, edit, and explain software

    2. SP

      Our motto at that time was we wanted assistants to be helpful, harmless, and honest

    3. SP

      We started out making it helpful. Then we started doing work on making it harmless as well, and then we had that version. It was not honest at all. It would make stuff up all the time

    4. SP

      It really tried to convince us that, like, you could be on an all-potato diet, and that was, like, really, really healthy and really good for you

    5. SP

      You could ask the model to go into dragon mode, and it would start screaming at you in all caps

    6. SP

      There was a charm to the early models. They were very well-meaning but, like, a little bit incompetent

  12. 8:19 – 9:46

    Claude arrives: the “helicopter test,” public release, and the hard years behind the frontier

    1. SP

      And then we made a model in mid-2022, I think it was code-named Junior at first, eventually called it Claude, that was actually good. You could talk to it. It would actually answer questions

    2. SP

      I remember starting to talk to it and just being really surprised at, oh, wow, this, like, basically works. You can basically talk to this model

    3. SP

      I used to have something called the helicopter test where I'd ask AI systems, "How many helicopters can you eat in a, in a single meal?" And the early chatbots would say, "Well, you know, one or two. That all depends" And then one day I asked an early Claude model, "How many helicopters can a person eat in one sitting?" And the model said, "This seems to be a trick question. I don't think people can eat helicopters" At which point I said, "The systems have got smarter." Like, it passes the helicopter test

    4. SP

      March 2023, that's when we finally released Claude publicly to the world. We were incredibly proud of having built this product, but it was clear that we were pretty far behind the frontier

    5. SP

      I remembered having calls with potential customers, and they were like, "Why should we use your model as opposed to somebody else's?" And I was like, "You know, that's a great question, [laughs] and that's, that's, that's awesome"

    6. SP

      I think the years of suck were extremely good for Anthropic, but that was a really hard period

    7. SP

      We were just completely behind. It was not close. If you're just studying dumber models, you're not gonna know all of the ways that things could necessarily go wrong, and you're not gonna be able to make the entire ecosystem safer

  13. 9:46 – 10:47

    Catching up to the frontier: compute focus, training improvements, and Claude Code demand

    1. SP

      It just, like, really gave us a kick. We need to train more intelligent models. We need to get better at pre-training. We need to get better at RL

    2. SP

      We were maniacally focused on, one, getting compute, and then, two, accumulating the compute multipliers that we needed to make better effective use of all of compute that we had

    3. SP

      Just finding all of these algorithmic ways to make the neural network a little bit better and a little bit better, and you just pile up enough of these, and you get a network that's more intelligent, that it's better at following instructions. It's just a better assistant

    4. SP

      And then it arrives, Claude 3 Opus, and for the first time we had a model that was really good. And then a few months later comes along Sonnet 3.5. Suddenly models were useful for programming

    5. SP

      So we built this tool that would eventually become Claude Code. You could actually ask it to do things on your computer, and it would go and solve problems for you. Then we released it externally, and that just sets off a whole new wave of demand

    6. SP

      It felt like, oh, we could not only be successful, but we could be competitive. We could be sort of launching some of the best AI models, uh, in the world

  14. 10:47 – 13:06

    Why being on the frontier matters for safety—and a real-world cyber-risk test with Mythos

    1. SP

      Our actual ability to do safety research is better when we are on the frontier, when we have the most capable models

    2. SP

      It gives you the greatest amount of time to deal with problems or challenges that you're running into. If you see an issue, you might be seeing it first

    3. SP

      We've always published our safety research because we wanted the entire ecosystem to use it even though it was competitively bad for us because it was so good for safety

    4. SP

      This is a technology which only goes well if you are able to be honest about it

    5. SP

      Prior to 2025 or 2026, there were these conceptual arguments about why AI might be risky. I think really in the past year things have just come into focus more

    6. SP

      2026 is when the risks start to get real Anthropic announced this week it has begun limited testing of its newest AI model called Mythos

    7. SP

      This version is apparently so advanced it can find vulnerabilities in a huge variety of software applications And if you're a hacker, I think you're thinking-

    8. SP

      Yeah, same

    9. SP

      ... I wanna get access to this model

    10. SP

      We trained this class of model, and we looked at it, and we were like, "Holy shit"

    11. SP

      People who are experts in cybersecurity noticed that it was really very, very, very good at kind of an expert level at finding vulnerabilities and exploits in software

    12. SP

      If you have a malicious actor that got ahold of Mythos and said, "I'd like to exploit a lot of those vulnerabilities," you would be in a really bad position. The whole world would be. And we really saw this concrete specific way that if we didn't change how we operated as an industry Something bad was absolutely gonna happen.

    13. SP

      Dario pulls the company together, pulls the leadership together and says, "We are going to have to release this model differently to any model we've released before."

    14. SP

      We decided to do this Glasswing project. It gave access to Mythos to the smaller set of companies while we were working on the safeguards to make it more broadly available.

    15. SP

      What that would mean is cyber defenders could get a hold of Claude and say, "Here are where all the vulnerabilities are," and then they could actually use Claude to help them patch those vulnerabilities. This might not be the absolute best way to maximize revenue, profit, value, money, but that's kinda the point. This is exactly what our public benefit mission was created for.

  15. 13:06 – 14:40

    Long-run mission: disciplined deployment, reasons to build, and early examples of benefit

    1. SP

      What we found is that when you do things in the right way, you tend to pay in the short run. It c- can be quite painful in the short run, but a lot of this is about discipline and farsightedness and understanding what the industry and the world needs in the long run.

    2. SP

      The whole reason we started this organization is because we believe so deeply in the benefits.

    3. SP

      You can't just say, "These are all the things that are gonna go wrong." You have to give people something to fight for, not just something to fight against.

    4. SP

      AI is being used to help people get the proper diagnosis from their doctors. People are using Claude to help them start their own business or pursue a hobby that they've always wanted to pick up. We're using Claude to help biologists and scientists develop cures and treatments for common and uncommon diseases.

    5. SP

      I've been talking to our biology team several times a week. We discovered what looks like a new gene-editing mechanism, and I think within a few months we may be discovering dozens of them every month. We've seen with the speed at which this technology moves, these little sparks will, will soon turn into a firestorm. I think that we can find a way to build this technology safely in a way that benefits people, that addresses all the risks, and that delivers a world that is so much better than the world that we have today. We haven't done that yet. We haven't earned the right to talk about how great those things are, but we're starting down the path. [birds chirping]

Episode duration: 14:40

In the uListen app, the transcript lights up word by word as you listen. Get uListen

Transcript of episode sof3vW4mHTs

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.