Anthropic CEO hits back fiercely in a lengthy post, revealing that he has kept Mythos under wraps for 4 months, voluntarily cutting off a lucrative revenue stream for the sake of AI safety.
In the wee hours of Silicon Valley, the air is thick with the smell of gunpowder.
Just moments ago, Dario, the head of Anthropic, rarely lost his composure on X.
He personally posted a nearly 1,000-word hardcore long post, directly hitting back at the most piercing question the public has raised about Anthropic —
Is your so-called AI security just a fig leaf for raking in money frantically?
In response, Dario revealed for the very first time —
A fully trained Mythos-class advanced model had been kept under wraps for about four months before the public was allowed to access it.
He laid out every penny of the money his company "actively lost due to safety reasons" over the past year on the table.
In an industry where model rankings can be rewritten every few weeks, four months is long enough for wave after wave of new models to make their debut.
Choosing to wait very likely made Anthropic give up a significant part of the months-long lead it could have had.
"Anthropic Only Wants to Win the Superintelligence Race"
This time, the person who pushed Dario to break his usual calm is Nathan Calvin, a well-known investigator in the AI security field.
A few hours ago, Calvin's thousand-word scathing essay instantly blew up the entire AI circle.
He published a long list, detailing the "nine sins" of Anthropic over the past year with every point hitting the nail on the head.
Spending 500 billion US dollars on a computing power order, secretly forming a self-developed chip team, letting Claude take over 26% of internal R&D tasks for "self-iteration", quietly modifying safety commitments......
Calvin's underlying message is extremely sharp: Stop pretending, you just want to build superintelligence before everyone else!
The trigger for this debate was originally a draft federal AI security act that leaked over the weekend.
After reading the draft, Calvin and several policy researchers were furious, believing that this act, which Anthropic pushed behind the scenes, seems to call for regulation on the surface, but hides hidden tricks —
It not only tries to override stricter laws in various states, but also sets an extremely high threshold for the federal government to call a halt to dangerous AI.
Anthony Cimino, head of federal affairs at Anthropic, originally intended to post a response to calm the situation over the weekend, but completely ignited Calvin's anger.
Calvin directly threw out a series of soul-searching questions: What kind of company is Anthropic exactly?
If you go by what the executives said in the open letter, this is a "safety institution" that is always ready to hit the brakes for humanity.
Not long ago, Dario even publicly warned that AI clusters could take over the entire internet within the next 6 to 12 months, calling on the whole industry to slow down.
Calvin sarcastically pointed out: If these 9 things were done by OpenAI, Anthropic's employees would have long since come out to denounce them, and take them as solid evidence that the rival is "greedy for money and disregards human lives".
Dario Shows the Books Late at Night
How Much Revenue Did We Voluntarily Give Up For Safety?
Facing the accusations that almost overturned the company's moral foundation, Dario did not hide.
At the beginning of his response, Dario first showed high EQ, thanking Calvin for his supervision, but after the polite remarks, he directly made his point.
Right after that, he listed 9 "self-harming operations" that would make countless Silicon Valley investors' hearts ache.
After Getting the Divine Model, We Kept It Under Wraps for 4 Months
This is the most revealing bombshell in the entire long post.
Dario publicly admitted for the first time: Back in February this year, Anthropic had already trained a high-level model of "Mythos class".
But Anthropic held back resolutely.
At that time, the cybersecurity defense mechanism for this class of model had not been fully polished, so they chose to only open it to a very small number of partners, specifically for enhancing cybersecurity defense tests.
It was not until the safety guardrails were in place in June this year that this model was officially opened to the public.
A full 4-month leading window, in the AI race where every second counts, was cut off by their own choice.
Dario said frankly that it was precisely because of this 4-month buffer that many enterprises managed to reinforce their own safety infrastructure before AI models with strong cyberattack capabilities became widespread.
On June 9 this year, Fable 5 launched by Anthropic for ordinary users belongs to this tier, with strict protection.
Mythos 5 released in the same launch event was first provided to a small number of cybersecurity defense and infrastructure partners with partially unblocked capabilities.
Therefore, there was still restricted use during the "under wraps" period; when it was opened in June, there were also different access boundaries.
Of course, he did not give a quantifiable account of how much revenue the wait cost, and how many risks it avoided.
But this choice at least shows that the release timing itself has become part of AI safety.
"Zero Tolerance" for Biological Risks
The cost in the biological sector is especially painful for Dario. He himself has a background in biology.
He stated that the company has seen worrying results in biological capability tests, and has intercepted potential bioweapon abuse attempts, so strict protection is adopted for high-risk use cases.
The problem is that it is very difficult for filters to accurately read everyone's intention.
Letting dangerous requests pass is called false negatives; blocking normal research is called false positives. To minimize the former, Anthropic accepted a large number of the latter.
Normal researchers are thus frustrated, and competitors even take "fewer restrictions" as a selling point.
This cost is also borne by users.
Anthropic's previous official statement confirmed that after Fable 5 triggers the biosafety filter, the request will be handed over to a less capable model for processing.
The August update reduced bio-related rollbacks by about 85% in the company's tests, but restrictions on dual-use research are still retained.
Another account is regulation.
Dario said that Anthropic promoted legislation such as California's SB 53, and also supported stricter proposals than the final passed version. In June this year, he publicly advocated that the government should have the power to block model releases.
This is equivalent to supporting external forces to hold the switch of its own products.
Finally, it is 5% of the equity. Just a few days ago, Anthropic announced that it plans to put 5% of the company's shares into a public interest fund, pending shareholder approval. After the company's valuation exceeds the specified benchmark, the board of directors may additionally add 2% at its discretion for every 50% increase in valuation.
AI is Getting Faster and Faster, Who Will Hit the Brakes?
Why does a company that insists on safety still spare no effort to buy computing power and accelerate R&D?
The core logic Dario gave is called "race to the top", competing towards higher standards.
His judgment is that a company at the cutting edge of technology is more capable of pushing peers to raise standards only by proving that advanced AI can be developed responsibly.
Calvin's concern is also very straightforward: if you have to win every round of competition, how much can the company give up for safety?
Especially when AI starts to participate in the R&D of the next generation of AI, capability improvement is getting faster and faster, and the time for human beings to check problems and fix loopholes may be getting tighter and tighter.
Today, Anthropic is willing to wait four months.
Next time, in the face of a more powerful model and larger orders, will it still be able to wait?
References:
https://x.com/DarioAmodei/status/2112732230124818891?s=20
https://x.com/_NathanCalvin/status/2112295219765137570?s=20
This article is from the WeChat Official Account "Xinzhiyuan", author: ASI Revelation, editor: Taozi, published with authorization from 36Kr.