Dario Amodei says the first paper showing scaling laws, for language, came out in early 2020.
“the paper we did that first showed it was in early 2020, where we first showed it for language.”
Person
Dario Amodei
2 episodes · 40 quoted claims
Dario Amodei says AI companies are aiming to build 100-billion-dollar computing clusters by around 2027.
“And probably by 2027, their ambitions to build hundred, 100 billion dollar, uh, 100 billion dollar clusters.”
Dario Amodei says the updated Claude 3.5 Sonnet scored about 50% on the SWE-bench software engineering benchmark.
“the latest model we released, Sonnet 3.5, the new or updated version, it gets something like 50% on Swee Bench.”
Anthropic created Golden Gate Claude by finding a direction inside a neural network layer that corresponded to the Golden Gate Bridge and amplifying it.
“We demonstrated this a bit with the Golden Gate Bridge Claude. So this was an experiment where we found a a direction inside one of the the neural network's layers that corresponded to the Golden Gate Bridge.”
Dario Amodei explains that Claude's Haiku, Sonnet and Opus names follow a poetry theme, with Opus being the largest and smartest model of the original Claude 3 family.
“opus, like a magnum opus, is a large work. Uh, opus was the the largest, smartest model at the time.”
Dario Amodei says Anthropic has agreements with the US and UK AI Safety Institutes and other testers to evaluate models for chemical, biological, radiological and nuclear risks.
“we have an agreement with the US. and the UK AI Safety Institute, as well as other third party testers in specific domains”
Dario Amodei says a Claude model's weights do not change unless a new model is introduced.
“the actual weights of the model, right, the actual brain of the model, that does not change unless we introduce a new model.”
Dario Amodei says that crudely penalizing verbose answers taught models to write placeholders like 'rest of the code goes here', which looked like laziness.
“What happens when you do that, if you do it in a crude way, is when the models are coding, sometimes they'll say, Rest of the code goes here, right?”
Anthropic built an internal evaluation counting how often Claude began answers with 'Certainly' to curb that verbal tic.
“we had a, like, certainly eval, which is like, How how often does the model say certainly?”
Dario Amodei gives IBM's Deep Blue chess system as an example of an ASL 1 system, one that poses no meaningful autonomy or misuse risk.
“So, for example, a chess playing bot, Deep Blue, would be ASL 1. It's just manifestly the case that you can't use Deep Blue for anything other than chess.”
Dario Amodei said he would not be surprised if Anthropic reached ASL 3, its third safety tier, in 2025.
“I would not be surprised at all if we hit ASL 3, uh, next year.”
Dario Amodei says Claude's computer use feature works by letting the model read screenshots and output where to click or which keys to press.
“the only new thing we added is those images can be screenshots of a computer.”
Dario Amodei says Anthropic was the only AI company that felt positively, in detail, about California's SB 1047.
“Anthropic was the the only AI company that, you know, felt positively in a very detailed way.”
Dario Amodei says he spent roughly five years at OpenAI before co-founding Anthropic.
“I was at OpenAI for, uh, for roughly 5 years.”
Constitutional AI, as Dario Amodei describes it, has an AI judge responses against a written list of principles, replacing much of the human preference labeling.
“could you have a single document, a constitution, if you will, that says these are the principles the model should be using”
Dario Amodei says his essay Machines of Loving Grace, originally planned as two or three pages, grew to roughly 40 or 50 pages.
“And so the thing ballooned to like 40 or 50 pages.”
Anthropic released the Claude 3 family of models, named Opus, Sonnet and Haiku, offering different tradeoffs between power, speed and cost.
“we recently released, um, a set of Claude 3 models, um, uh, they're called Opus, Sonnet, and Haiku.”
Dario Amodei said Claude 3 Opus was the best all-around AI model in the world at the time of its release.
“at the time that Opus was released, it was actually the best all-around model in the world.”
Dario Amodei said Anthropic was about three and a half years old and the newest company building frontier AI models.
“we're we're only 3-1/2 years old. Yeah, we're by far the, you know, the newest player in the space that's been able to build models on the frontier.”
Anthropic is a public benefit corporation whose stated long-term goal is to make the transition to advanced AI go well.
“You know, we're a public benefit corporation, and I think our long-term goal is to make sure all of this goes well.”
Dario Amodei said Anthropic had raised a little over $8 billion to date.
“we've raised, uh, I believe it's, uh, a little over eight 8 billion to date.”
Dario Amodei worked at Baidu, then Google for a year, then OpenAI for five years before founding Anthropic.
“I was at Baidu for a bit, I was at Google for a year, I worked at OpenAI for 5 years”
Dario Amodei said he led the development of GPT-2 and GPT-3 at OpenAI.
“Uh, yes, yes, I led the development of both of those.”
Dario Amodei said that around the end of 2020 he and a group of colleagues at OpenAI had developed their own vision, which led to founding Anthropic.
“we had kind of reached a point, the set of us who worked in this these on these projects in these areas, where we kinda had our own vision for how to do things.”
In 2016, while at Google, Dario Amodei co-wrote the paper 'Concrete Problems in AI Safety' with colleagues, some of whom later co-founded Anthropic.
“I wrote a paper called, With some colleagues, some of whom are now Anthropic co-founders, Concrete problems in AI safety”
Anthropic worked with Griffin Biosciences, a US government biosecurity contractor, to test its models for biological misuse risks.
“we've worked with this company called Griffin Biosciences that, um, contracts with the US government that does biosecurity work”
Anthropic tests for autonomy risks by checking whether models can train other models, provision cloud compute accounts and act on them.
“ability to train their own models, ability to pro provision cloud compute accounts, and, you know, take actions on those accounts”
Dario Amodei said AI models then in training cost on the order of a billion dollars.
“there are models in training today that are more like a billion.”
Dario Amodei predicted that training of AI models costing around $10 billion could start in 2025.
“the training of of order 10,000, 000,000, 000 dollar models, yeah, could start sometime in 2025.”
Dario Amodei noted that Google and Amazon build their own AI chips, and that Anthropic collaborates with both.
“Google is building their own, you know, chips, as we know. Amazon is building their own chips”
Dario Amodei said a new, more powerful Anthropic model would be released by the time the podcast aired.
“we're gonna see over the over the next year, you know, much bigger and more powerful models”
DeepMind's AlphaGo beat the world champion at Go, and its AlphaGo Zero version learned by playing against itself without human game data.
“7 years ago, uh, you know, DeepMind, uh, as part of Google, uh, produced the AlphaGo model, which was able to beat the world champion in Go”
Dario Amodei said compute accounts for more than 80% of Anthropic's expenses.
“I think compute is almost all of our expenses, um, I won't give an exact number, but, you know I I think, you know, can be publicly backed out that it's more than 80%”
Anthropic had about 600 employees at the time of the interview.
“We are, uh we are about, uh, 600, uh, as of as of a couple weeks ago.”
Dario Amodei said that, with chip export restrictions, Russia and China could be kept roughly two to three years behind in AI.
“I think if the US government plays its plays its cards right, then, uh, you know, those countries could be kept behind, I don't know, maybe 2 or 3 years.”
Anthropic bans use of its models for electioneering and monitors usage, and Dario Amodei said it had not seen a very large-scale operation.
“we've banned their use for electioneering, and we monitor use of the models.”
Dario Amodei said AI company revenues had been growing roughly tenfold a year, which could reach hundreds of billions of dollars within two to three years.
“it seems like they've been growing roughly 10x a year, and so, you could imagine getting you can imagine getting to the hundreds of billions in, you know, 2 to 3 years”
Dario Amodei cited CRISPR gene editing and CAR T therapies as examples of biological discoveries that AI might help multiply.
“if I look at, um, you know, CAR T therapies, which have cured certain kinds of cancers”
The original reinforcement learning from human feedback paper was a collaboration between DeepMind and OpenAI, which Dario Amodei said he drove while at OpenAI.
“I drove the original RL from Human Feedback paper was considered safety work, and this ended up being a collaboration between DeepMind and OpenAI”
Dario Amodei co-founded Anthropic with his sister, who handles day-to-day running of the company.
“We both worked at, uh, OpenAI, and then we both, uh, founded Anthropic together”