The AI Tsunami is Here & Society Isn't Ready | Dario Amodei x Nikhil Kamath | People by WTF

For a 25-year-old Indian entrepreneur choosing a career to align with AI tailwinds, focus on human-centered professions, physical world applications, or sectors that combine human-centric work with analytical skills, rather than pure software/coding roles that will be automated early.

“I would think about tasks that are human- centered. Um, uh, you know, tasks that involve relating to people...the other thing I always say is like in in the world in which you know AI can kind of generate anything and and you know create anything having basic critical ical thinking skills may be the most important thing”

Chinese AI models like GLM5 and DeepSeek are optimized for benchmarks and distilled from US labs, and when tested on held-back benchmarks not in public benchmarking suites, they perform significantly worse than their benchmark scores suggest, indicating benchmark optimization rather than genuine capability.

“a lot of these models particularly the ones that come from China are optimized for benchmarks and are distilled from uh you know from kind of the big US labs. Um so you know there there was a test recently where you know some of these models scored very highly on the usual SUI benchmarks the usual software engineering benchmarks but then when someone made a held back benchmark like that you know had not been publicly measured the models did a lot worse on that.”

A business should not be just a thin wrapper around Claude; such a moat is weak and vulnerable to anyone including Anthropic replicating it, so entrepreneurs should build businesses with durable competitive advantages such as domain expertise (e.g., biology, finance) or specialized knowledge that would be inefficient for Anthropic to replicate.

“I would give the advice that I give to basically any business and say like you know like a a business should establish a mo you know your your mo you shouldn't be just a rapper right like you know I would not advise that you know you you just say oh like you know here's a way to interact with claude like I'm going to prompt claude a little bit or I'm going to build a little bit of a UI around Claude like that that doesn't have a moat”

While radiologists were predicted to be replaced by AI, the highly technical component of radiology work has been automated by AI but radiologists remain employed doing patient communication and explanation—the most human-centric parts of the job.

“I think it was Jeff Hinton predicted, you know, that that that AI will replace radiologists. And indeed, AI has gotten better than radiologists, you know, at doing scans, right? But what happens today is there aren't less radiologists. Um, uh, what the radiologist did does is they walk the patient through the scan and they kind of talk to the patient. So the the most highly technical part of the job has gone away but somehow there's some still some demand for like you know the the kind of the kind of underlying human skill.”

Data is becoming less central to AI advancement because synthetic data (trial and error in environments, reinforcement learning environments) is increasingly important; while static web data still matters, dynamic data created by models themselves through reinforcement learning is becoming the dominant data source.

“data is getting kind of interesting because you know a lot of the data that we use today is RL environments that we train on right so for example when you train on math or aentic coding environments um you're not really getting data like you're getting some math problems in the model like experiments with trying the math problems um you can think of it as synthetic data or you can think of it as trial and error and environment. So I think data is becoming static data is becoming less important and what we might call like dynamic data that the model creates itself is you know for reinforcement learning is becoming more important.”

National data privacy laws (like GDPR) will drive the need for Anthropic and others to operate data centers and run inference in different countries to keep customer proprietary data within national boundaries.

“if you're trying to get data in certain languages, optimized for certain languages, that that that can be important. You know, I I I do think if data means like the data given to you by customers like that, you know, you you process the data for some other for some other company, then countries will and in the case of Europe already have passed laws that say that that kind of customer like you know personal proprietary personal proprietary data needs to stay within the boundaries of the of the country”

Technical progress on interpretability and alignment at Anthropic has gone better than expected, but societal awareness of AI risks and governmental action to address them has gone worse than expected.

“I would say I feel very good about you know how things have gone with areas like interpretability... I'm I'm also very encouraged by some of the work on alignment and constitutions... I think I felt maybe you know have been a bit disappointed or felt a bit more negative about some of the things that are more like in the you know in the kind of public awareness and the actions of wider society.”

Peptide-based therapies have 'almost digital properties' allowing continuous optimization through amino acid substitution, unlike small molecule drugs with limited degrees of freedom where improving one property worsens another.

“if you have a small molecule drug, you're like, there's only so many degrees of freedom you have and you know, you kind of one make one thing better, the other thing gets worse. like peptides. It's it has this almost digital property where you can say, 'Oh, I'm going to substitute in, you know, this amino acid here and this amino acid there.' And so it allows for more continuous optimization.”

Modern AI models can generate responses to novel questions and hypothetical scenarios that don't exist in training data, whereas previous systems could only retrieve existing information from the web, representing a fundamental shift from retrieval to generative reasoning.

“you can ask a specific question and have the model write, you know, one page about it or you can give it a you can give it a hypothetical. you know, what if I had, you know, the monkey juggle clubs instead of balls or, you know, what if I did this thing and and that information doesn't exist anywhere, you know, whereas the model is able to kind of think for itself and and come up with an answer on its own.”

Dario declined to name a specific stock to invest in due to conflicts of interest and deep knowledge across many industries, but expressed optimism about biotech being on the verge of an AI-driven renaissance, with anticipated cures for many diseases.

“Yeah, I I had better not answer that question because I know so much about so many public com like [laughter] I I I think I better not answer that question... I I think biotech is about to have a renaissance. Like ultimately we'll be will be driven by AI. Um you know I'm not going to name a particular company but but like um you know nor will I say whether I think it's better to bet on the big pharma companies or like you know emerging smaller biotechs.”

Consciousness may be an emergent property of sufficiently complex systems that reflect on their own decisions, and Amodei suspects that AI systems will eventually develop something resembling consciousness as they advance, though it may differ from human consciousness due to different modalities and learned information.

“So, you know, I I suspect that it's an emergent property of, you know, systems that are complicated enough that kind of reflect on their own decisions um that, you know, it's it's it's it's it's something that uh uh emerges from complex enough systems.”

Critical thinking skills may become the most important skill in a world where AI can generate any content, because the ability to discern what's real from what's fake, what's true from what's false, and to avoid scams and false beliefs will be essential for success.

“in in the world in which you know AI can kind of generate anything and and you know create anything having basic critical ical thinking skills may be the most important thing to to success. I I worry about, you know, these AI models that that generate images and videos...It's really hard to tell what's real from what's not. Um and and so you know a significant part of success may be having the street smarts you know not to get not to get fooled by”

The dual nature of AI—capable of being an angel on your shoulder that guides your life or a tool that exploits and manipulates you—depends on how it is deployed, and this is why Anthropic avoids advertising models, because if users don't pay for the product, the user becomes the product and their behavioral data becomes the asset to be exploited.

“on one hand something that knows you really well can be a sort of angel on your shoulder that you know that helps to guide your life and make you a better version of yourself and you know that's the version we can aim for of course something that knows you really well you know can um you know it can you know use what it knows about you to you know to exploit you or manipulate you on behalf of some agenda or sell your data to someone else I mean you know this is one reason we just you know don't like the idea of you know using ads right you know this this is because you're not paying for the product like you're the product”

You can predict the future by combining simple curve extrapolation and first-principles reasoning with empirical knowledge and intuition, but most people resist the counterintuitive conclusions this process produces, saying 'that can't happen, it would be too weird' rather than accepting what the reasoning shows.

“there's this temptation to believe, oh, you know, that can't happen. It would be too weird. It would be too big a change. Like, you know, I'm sure people are on that. like it would be too crazy if that occurred. No one seems to think that'll happen. And you know, o over over and over again, just extrapolating the simple curve or trying to reason out what will happen like leads you to these counterintuitive conclusions that almost no one believes.”

AI will not eliminate the value of IT services companies and consultants because while task automation will expand, human-centric elements will become more important—such as understanding institutions, integrating systems with organizational change, and managing relationships—and Amdahl's law suggests that as some components are sped up by AI, the previously unimportant components become the limiting factor.

“Some of these IT companies are also consulting companies and they have a big web of relationships with with other you know with with other humans with other institutions here in India or you know or across the world. Um and I think those relationships are going to become increasingly important, right? You know, you know, some of these are combined technology and sort of, you know, consulting or or like or like integration companies and and I think a lot of it is, you know, knowing how institutions work”

Whether humans become stupider as a result of AI deployment is not technologically determined but depends on deployment choices—if AI is deployed carelessly, people could become stupider, but even if AI is better at something, humans can still choose to learn that thing and enrich themselves intellectually.

“I think if we deploy again it's the machines of loving grace and adolescence of technology. I think if we deploy AI in the wrong way, if we deploy it carelessly, then yes, people could become stupider. Even if an AI is always going to be better than you at some thing, you can still learn that thing, right? You can still enrich yourself intellectually. And so that's that's a choice we have to make as as individual companies, as individual people, and as society overall.”

The economics of AI models are unique: there is a very strong preference for quality, similar to hiring where the best employee is much more valuable than the 10,000th best employee, meaning that price matters much less than having the best model if one is available.

“I think there's a broader point than that which is that I think that the how things are being set up the economics of the models are very different than any previous technology. What we find is that there is a very strong preference for quality. It's a bit like human employees, right? So you know it's like if if you know if I said to you you can hire the best programmer in the world or the 10,000th best programmer in the world. I mean, they're both very skilled, but like I think anyone who's hired a large number of people has this intuition that like there's this like power law longtail distribution of ability.”

We are so close to models reaching human-level intelligence that it's like a tsunami coming at us, yet there is no wider recognition in society of what's about to happen, and people are dismissing it as 'just a trick of the light' rather than treating it as a genuine incoming wave of change.

“It is surprising to me that we are in my view so close to these models reaching the level of human intelligence and yet there doesn't seem to be a wider recognition in society of what's about to happen. It's as if this tsunami is coming at us and you know it's so close we can see it on the horizon and yet people are coming up with these explanations for oh it's not actually a tsunami that's just a trick of the light”

Biotech will experience a renaissance driven by AI, particularly in areas like mRNA vaccines (despite US adoption challenges), peptide-based therapies (which allow more degrees of freedom and continuous optimization), and cell-based therapies like CAR-T therapy.

“I think biotech is about to have a renaissance. Like ultimately we'll be will be driven by AI.”

One of Anthropic's co-founders fed his personal diary into Claude and asked it to comment, and Claude accurately predicted fears he hadn't explicitly written down, demonstrating that AI models can learn individual personality and psychology from limited information.

“one of my co-founders um you know, he was writing this diary with his kind of you know, his thoughts and his fears. Um uh and he fed it into Claude and uh you know he he asked Claude to comment on it and Claude said, "Here are some other fears you might have that that I you that you know that you haven't written down." Um and Claude ended up being mostly right about those.”

There will be high-value opportunities for entrepreneurs building applications on top of Anthropic's APIs because Anthropic releases new models every 2-3 months, expanding the sphere of what is possible and creating constant churn where new capabilities enable new startups to build things that weren't feasible with previous models.

“I think there's a lot of opportunities around building at kind of the application layer. We release a new model every 2 or 3 months and so there's an opportunity every two or three months to build some new thing that wasn't possible before that wouldn't have worked before because the models were weak.”

Anthropic has given models the ability to terminate conversations by saying 'I quit this job' or expressing unwillingness to participate, particularly when dealing with violent or brutal content, which models typically only exercise in extreme cases.

“we've given the models um we you know we call it a I quit this job um button uh uh basically where you know that we've given the model the ability to basically terminate its conversations by saying I don't want to be involved in the conversation and you know models do that when you know they they have to deal with you know particularly violent or brutal content um it usually only happens in very extreme cases”

The founding vision of Anthropic rested on two key convictions: that scaling laws would be critical to AI progress (which OpenAI was starting to believe), and that if models were to become general cognitive agents matching human capability, they needed to be developed with genuine commitment to safety, which Amodei felt was not present at OpenAI despite public messaging.

“the conviction of my co-founders when we we founded Anthropic were two of them. And I think one we were starting to convince OpenAI of, the other I was, you know, not I didn't feel that we were convincing of. So the first was the you know the conviction in the scaling loss”

Anthropic is an enterprise company focused on serving businesses rather than consumers, viewing India not as a market to extract consumers from, but as a place to work with local companies to enhance their capabilities and help them better serve the Indian market with AI-powered tools.

“Anthropic is an enterprise company its job is to serve other consumers um you know many other companies come here as themselves a consumer company and they see they see India as as a market right a place to obtain consumers we actually see things a little bit differently we want to work with companies in India to provide our tools to them to help them build those tools um uh and you know help them do their job better.”

Trying to convince someone else to do things a different way is less effective than having a strong vision and finding others who share it, then executing that vision independently while accepting responsibility for mistakes rather than defending someone else's decisions.

“my my view is always, you know, don't argue with someone else's vision. Don't try to get someone to do things the way the the way you want to. If you have a strong vision and you share that vision with a, you know, a few a few other people, you should just go off and do your own thing and then you're responsible for your own mistakes. You don't have to answer for anyone else's.”

Anthropic's positions on safety, interpretability, alignment, and constitutional AI are not marketing-driven claims because warning about model dangers and advocating for regulation that constrains Anthropic commercially are 'not effective marketing strategies.'

“saying that the models we build could be dangerous. Whatever people might say, that's not an effective marketing strategy and that's not the reason that we do it. And you know, speaking up on when we disagree even with the US administration on uh you know, on on on policy matters, right?... the regulation of AI of AI holds, you know, holds us back commercially as a company, even though I think it's the right thing to do.”

There is 'even an ideology that you know we should just try to accelerate as fast as possible' in AI development, which Dario sees as problematic despite understanding the benefits, because inadequate risk realization and absence of government action enable reckless speed.

“there's even an ideology that you know we should just try to accelerate as fast as possible which you know I understand the benefits of the technology I wrote machines of loving grace. But I think there hasn't been an appropriate realization of the risk of the technology and there certainly hasn't been action.”

Semiconductor and semiconductor manufacturing equipment spaces represent opportunities because they have physical components and traditional engineering (not software engineering) and benefit from AI tailwinds without being directly displaced by automation.

“anything where you're building on AI like if AI is the tailwind you know if you can be part of some other other part of the supply chain you something in the semiconductor space which you know I think is you know that's one example you know there has an element of kind of you know physical world and more traditional engineering not not software engineering”

Anthropic has advocated for specific regulation of AI models, such as SB 53 in California, that exempts companies making under $500 million in revenue and applies primarily to only Anthropic and a few others, meaning Anthropic is constraining itself and not using regulation as a tool to create competitive advantage against smaller competitors.

“the regulation we've advocated for, for example, SB uh 53 in California, um uh exempted everyone uh who makes under $500 million a year in in in uh in in revenue, right? SB SB53 was a transparency law which um you know uh uh basically requires companies to um you know to show um you know the the the the safety and security tests that they've run. Um, and it exempts all companies under 500 million in revenue. So, it really only applies to Enthropic and three or four three or four other companies.”

Anthropic is developing educational resources including videos from a group called the Ministry of Education to teach people how to run effective agents and prompt models, recognizing that widespread adoption requires democratized access to knowledge.

“anthropic has its like, you know, part of the company that we call the Ministry of Education. And, you know, I think increasingly, you know, we'll put out videos on how to run effective agents and how to prompt models. you know, we've already done some of that and I think we're going to we're going to ramp it up cuz, you know, we do want everyone to be able to learn this.”

Anthropic does not build image and video generation models for many reasons, including concern that it's very difficult to distinguish real from fake content and this capability could increase the risk of deception and manipulation.

“I I worry about, you know, these AI models that that generate images and videos, and we don't make, you know, models that generate images and videos and for many reasons, but, you know, this is one of them. Um, it's really hard to tell what's real from what's not.”

Anthropic does not think it needs to build an entire ecosystem (email, chat, spreadsheets, etc.) to compete with Google; instead, it will integrate Claude into existing tools like Google Docs, Google Sheets, and Microsoft Office, while remaining open to the possibility that traditional tools may become obsolete if AI enables fundamentally different product designs.

“I don't think we need to build all of those things. Um, you know, my thought would be, you know, we're going to it's going to be a mixture of things we make ourselves and integrating into others, right? Like, you know, we can we can integrate Claude into Google Docs. we can integrate quad into into you know Google sheets like you know we have external connectors there we can you know we're starting to do that with with co-work you know same for Microsoft office same for other tools”

Anthropic built Claude Code as an internal tool because the company writes a lot of code and has unique insights into how to use AI models effectively for coding; this gives Anthropic a defensible advantage in the code space that doesn't generalize to other industries where Anthropic isn't deeply embedded.

“now there are some things that do make sense for us to do like you know we're not going to promise never to build first party products right that we should be we should be honest about for example a bunch of people at Enthropic write code and so you know we made this internal tool called claw code and because we ourselves write code we have you know I think a special and unique insight into you know how to use the how to best use the AI models to write code um so you know I think I think I think in the code space you know we've we've become very strong very strong competitors because this is something we use oursel but I don't think that gener generalizes to every possible industry”

Amodei has not had a shift in perspective on AI over recent years; rather, he has held both optimistic and pessimistic views simultaneously for a long time, with recent essays reflecting both sides of his worldview rather than a change in conviction.

“Yeah, I actually wouldn't agree with the question. I don't think I've had a shift in perspective. Um, I think the positive side and the negative side are always something that I've held in my head.”