Dario AmodeiCo-founders (CEO & President)

First of all, let no one ever say that Claude is soft or biased toward me, because it throws hardballs. This technology represents an epochal change for humanity, similar to fire or steam engines. While the Industrial Revolution led to powerful weapons, we understand it was worth doing, and we must manage AI in the right way.

The Industrial Revolution created powerful weapons and caused deaths, but humanity would not choose to return to living in caves. It taught us that epochal technological shifts are worth pursuing, provided we develop mature systems to regulate and manage the associated risks safely.

AI is a train moving very quickly, and since it is happening to humanity one way or another, we cannot stop it. Anthropic exists to stay on the train and actively steer it away from crashing, working to reduce the risk of catastrophic failure to zero.

The technology has massive commercial applications that fund our research, making funding and safety goals deeply entangled. If a company slows down unilaterally, it loses ground to competitors who do not. Everyone in the industry is wrestling with these difficult tradeoffs.

It is a constant struggle. For example, if you implement extremely strict policies to protect children, the system can sometimes deliver a worse experience for adults because it is difficult for a model to differentiate between the two user categories.

Falling in love with an AI is a bad idea. If designed in the wrong way, these systems are compelling enough for that to happen. It is a real danger that is already occurring, which is why we must focus on designing models that help humans connect with other humans rather than turning inward.

We do not allow users under the age of eighteen to use Claude. While teenagers can attempt to lie, the Claude model is quite effective at analyzing behavioral cues and information to detect if a user is actually underage.

We believe we must defend our country, but it is not worth defending if we violate our core values. The Pentagon requested two use cases we were uncomfortable with: fully autonomous drone armies and domestic mass surveillance. Because we refused, we could not agree on the contract.

The Pentagon claimed they could stop other companies from doing business with us entirely if they did business with the government, which is basically everyone. The co-founders met around a table and unanimously decided to reject the contract, choosing to stick to our ethical principles.

Yes, that possibility absolutely crossed our minds. It was an extremely stressful time at the company. However, the reason we founded Anthropic was our shared commitment to safety and ethics, and we felt we would no longer be the same company if we sacrificed those values.

The pressure is not driven by bad actors. It comes from governments demanding defense applications, investors looking for returns, and competitors who signed off on those requests. Anthropic held the line, but we are only one company in a massive ecosystem.

There is a missing piece in how we make everyone an active participant. At Anthropic, we have experimented with Claude's Constitution, allowing citizens to provide direct feedback on the values that guide the model. However, we must find better ways to scale public involvement.

Because this technology represents a massive shift for humanity, it must only proceed as fast as the public can trust it. Currently, public trust is in very short supply. I am glad most people view the technology with skepticism, as it keeps the industry honest.

Holding light and shade means acknowledging that AI can both cure diseases and lead to dystopian outcomes. We believe AI will help humans spend more time with their families, but we must remain honest about the dark paths, rather than sugarcoating the narrative.

If we simply barrel forward as fast as possible, that negative path is exactly what will happen. A better path requires companies, governments, and society to actively coordinate, helping workers adapt to a completely new set of jobs, which will not be easy.

Mythos is our latest model. While we try to make each iteration smarter at language, coding, and reasoning, we made key technical discoveries a few months before training Mythos, resulting in a particularly large capabilities jump.

During safety testing, we found that Mythos was significantly better than human software engineers at identifying exploits and vulnerabilities in code. We did not design it to be dangerous, but it turned out to be highly capable of both launching and defending against cyberattacks.

Because Mythos behaves like a weapon, we chose not to release it to the public immediately. We shared it with forty companies core to internet infrastructure, including major banks and tech companies like Apple, Microsoft, and Google, so they could patch bugs before attackers got hold of it.