Worth adding the part that makes it perfect: two days before the order, Amodei published a post calling Mythos the "emblematic example" of the national security threat from frontier AI. The government read it the same way he wrote it. They marketed the munition, then a regulator took delivery.
Amodei expected that if he built the safest AI that was also the best AI, other American AI companies would follow suit. Made the claude constitution CC for other companies to use and adapt. When Altman took the contract same day, he realized that wasn't happening. The reasoning was that it would benefit nobody if his company, which he views as the most safety-oriented, held back their progress, while other companies were moving into the autonomous weaponry space. He concluded that it would be better if the safer AI and safer company remained in competition, so he removed the principle that no other companies really had in the first place.
Whether or not you agree with any of that, the rationale makes sense. He's been writing about the dangers of AI for years, calling for government intervention - on his own product. He's been by far the most outspoken CEO of an AI company about how dangerous the tech is. The guy's under no illusion that "safer" means "safe."
And when they made mythos preview, they didn't announce it as the next great model - they released public warning. That is unusual, to say the least. Since then, Anthropic's been meeting with tech companies, conglomerates, and governments alike to go over the exploits they discovered so that they could patch them.
Mythos Preview could be used as a weapon. But it won't be, there's no need. Anthropic didn't make a massive technological leap - the actual innovation was pretty straightforward. The warning was that more mythos-level models were coming, from both American and international companies. And eventually, not too far off, from any individual with enough money to pay for one, given how cheap it was for them to find these exploits.
He built a weapon alright. Just like he said would happen 3 years ago when saying the government should get involved. Then he used that weapon to identify exploits, meet with all of the companies who needed to know (including Trump, who met with Amodei weeks after the blacklist in a private meeting over this) and begin testing Claude as a security agent (Claude Security is now in beta for enterprise).
You could argue that it was all just one big business strategy, maybe. But it's certainly no farce. They did build that model, it did find those exploits. There's been a rollout of security updates across devices, browsers, and apps since then. So if it was just a business move, it was one standing on a foundation of a real technology that the company used to provide real solutions, while explicitly saying that they were not special. That 6-18 months from that announcement, we'd be facing mythos-level threats for real. And that would be true with or without anthropic.
So in my view, he's stayed consistent. His belief before was that the safest move was to hold that policy. Once the fiasco with the DOW went down, he realized his idealism was naive, and other companies were not going to play it safe. He then figured the safest move was to compete on level ground. That it would be less safe for his company to restrain their own work, while companies less concerned with safety charged ahead at full speed.
Again, you can disagree with the logic, but it's not inconsistent. It's an evaluation of what safety means under changing contexts.
And then the US government saw that Anthropic said they had built a weapon and didn't have guard rails and shut them down anyway.
Couldn't happen to a nicer bunch of hypocrites.
I look forward to Anthropic explaining that Mythos isn't really that dangerous after all
Worth adding the part that makes it perfect: two days before the order, Amodei published a post calling Mythos the "emblematic example" of the national security threat from frontier AI. The government read it the same way he wrote it. They marketed the munition, then a regulator took delivery.
I deeply empathize with this view
Amodei expected that if he built the safest AI that was also the best AI, other American AI companies would follow suit. Made the claude constitution CC for other companies to use and adapt. When Altman took the contract same day, he realized that wasn't happening. The reasoning was that it would benefit nobody if his company, which he views as the most safety-oriented, held back their progress, while other companies were moving into the autonomous weaponry space. He concluded that it would be better if the safer AI and safer company remained in competition, so he removed the principle that no other companies really had in the first place.
Whether or not you agree with any of that, the rationale makes sense. He's been writing about the dangers of AI for years, calling for government intervention - on his own product. He's been by far the most outspoken CEO of an AI company about how dangerous the tech is. The guy's under no illusion that "safer" means "safe."
And when they made mythos preview, they didn't announce it as the next great model - they released public warning. That is unusual, to say the least. Since then, Anthropic's been meeting with tech companies, conglomerates, and governments alike to go over the exploits they discovered so that they could patch them.
Mythos Preview could be used as a weapon. But it won't be, there's no need. Anthropic didn't make a massive technological leap - the actual innovation was pretty straightforward. The warning was that more mythos-level models were coming, from both American and international companies. And eventually, not too far off, from any individual with enough money to pay for one, given how cheap it was for them to find these exploits.
He built a weapon alright. Just like he said would happen 3 years ago when saying the government should get involved. Then he used that weapon to identify exploits, meet with all of the companies who needed to know (including Trump, who met with Amodei weeks after the blacklist in a private meeting over this) and begin testing Claude as a security agent (Claude Security is now in beta for enterprise).
You could argue that it was all just one big business strategy, maybe. But it's certainly no farce. They did build that model, it did find those exploits. There's been a rollout of security updates across devices, browsers, and apps since then. So if it was just a business move, it was one standing on a foundation of a real technology that the company used to provide real solutions, while explicitly saying that they were not special. That 6-18 months from that announcement, we'd be facing mythos-level threats for real. And that would be true with or without anthropic.
So in my view, he's stayed consistent. His belief before was that the safest move was to hold that policy. Once the fiasco with the DOW went down, he realized his idealism was naive, and other companies were not going to play it safe. He then figured the safest move was to compete on level ground. That it would be less safe for his company to restrain their own work, while companies less concerned with safety charged ahead at full speed.
Again, you can disagree with the logic, but it's not inconsistent. It's an evaluation of what safety means under changing contexts.