It's the first time in the history of the world.
CreatedAIPeople have come to the point where they can't figure out how much AI has evolved!
Just as anthropic was pounding Nasdaq's IPO, it submitted an updated bid letter (S-1/A) to the U.S. SEC.
A week after the initial public offering, Anthropic's management has added a shocking "bottom line" to the file.
「AIThe accelerating cycle of AI has been going on for a long time, and it's evolving so fast that even human scientists can't monitor it in real time!
AIThe self-accelerating red line is almost at the foot.Internal measurements show that the future of Claude will soon trigger a "significantly accelerating AI research and development" decline.In less than a year, we've accomplished all the research and development progress of the past two years!
ConstructedAIEveryone is looking in the rearview mirror.Trapped in "delayed indicators", Anthropic acknowledged for the first time,In the past, there has been no timely detection of the acceleration of R&D, and the future may have to wait until the acceleration occurs!
Clearly prohibited "complete automation"AI」。Even though more than 90% of the internal code has been taken over by Claude, the company still draws the line on page 188.It's time to let Claude design the next generation of front-end models automatically!
Six words to change the letter of intentEven AI creators can only see things after they've done it.
To put the previous two editions of the book together in a step-by-step comparison, the most important change is on page 29.
In the first edition of the S-1 released a week ago, Anthropic said that the company "may not be able" to identify it before the significant acceleration of development occurs.
In this revised version, the legal and administrative authorities have inserted six English words specifically into the sentence."Have in the past not identified".。
In just six words, a direct warning of "the future may happen" becomes a confession of "the past has already happened".
In anthropic's own lab,AIThe speed of development has suddenly accelerated, and what the human team has reacted to has already happened!
According to internal estimates, the future model will soon hit the "self-accelerating red line" in the policy framework for responsible expansion.
According to this definition of the red line, once the line is crossed, it meansIt's not.AIAdditionally, in just one year, the entire two-year R&D progress can be completed at baseline speeds from 2018 to 2024.。
That's why even top researchers who write algorithms and digitize data themselves are "post-conscious", because it's difficult to improve the ability of large models to measure in real time.
The new model has become so strong in pre-training and Agent collaboration that the training curve is unreadable.
Researchers had to wait for the model to finish, run the evaluation, and then pull in internal engineering data and practice for several months, finally stating the code output and experimental cycles, in order to reverse the real acceleration frequency.
But the time is not right here.
When researchers used months of lag data to calculate that "it's crossed the line", the new model finished writing the core code for the next generation of models day and night.
People think they are in control all the time.AIThe progress of evolution, in fact, is often seen in retrospectives from a few months ago!
Internal Token bills squeeze full-time salariesIt's hard to let Claude create the next generation automatically.
The observation failed because Claude was too fast to take over the development of Anthropic internally.
The updated edition of the book reveals a set of the latest three-quarter internal data. By September, more than 90 percent of the code in the Anthropic internal code library had come from Claude.
With AI driven, the amount of code submitted per day by every engineer has risen from 35 times the average in 2024 in just one quarter.More than 45 times。
From January to September of this year, Claude's expenses were publicly quoted by the company.More than 250 times the storm.。
As early as April this year,Anthropic spends more on Claude internally each month than the total monthly cash salary sent to all its employees.(With no incentive for shareholding)
Since Claude writes code dozens of times faster than humans, and even bills overwhelm human wages, the best course of action is to let Claude design and train the next generation of Claude automatically.
On this skeletal node, Anthropic himself has drawn a red line!
At the end of page 188 of the chapter on the internal R&D cycle, the Anthropic specification adds a harsh rule: even if it is done entirely technically in the future, it can be done in the future.The company will not step away from human oversight and understanding to allow Claude to design and build the next generation of cutting-edge models.。
The concern is that more and more research and development work is taking over models today, and that if any generation of models hides gaps, biases or deceitful tendencies that humans have not discovered, the risk will be passed along in the research and development process.
Once the power is given to it to design the next generation fully automatically,It's hard to trace the source of these threats, because they're generational, cumulative, layered, and human-like.。
In order to prevent this transgenerational loss of control, Anthropic, on page 31, reveals that it is considering modifying the safety red line.Specific key decisions in the mandatory R&D process must be signed by human review。
However, maintaining the safe boundaries has also cost Anthropic a considerable amount of research and development progress.
Dario revealed a few weeks ago that "a trained Mythos-class high-end model was hidden in the snow for about four months".
Recently, in order to prevent the weight of the model from being stolen and to strengthen the security measures,The company postponed the start of pre-training for a new front-facing base model by more than a month.
For example, in April of this year, a model with a very strong capacity to crack a vulnerability was created.MythosUntil now, only a small number of institutions have been provided with protection through Project Glasswing project orientation.
Signed a long-term accounting agreement worth $895.4 billionWho's slowing down who's out
On the one hand, the pre-training of the new model was delayed by more than a month due to safety concerns; on the other hand, Anthropic earned more money than anyone else.
This reflects the cruelest reality of the AI competition at the forefront:All the top players are trapped in a prisoner's dilemma that no one dares slow down.。
In Anthropic's view, if we don't invest enough in AI-assisted research and development, we're not going to be able to do that.Competitors will continue to increase their profitability with AI-created AI, which will keep you far behind in a few generations.。
As a result, Silicon Valley's core computing supplier was almost completely signed by it.
In first place are the $16.12 billion equipment leases from BoT, and the $12.13 billion TPU computing capabilities from Google Cloud (plus up to $150 billion in additional frameworks).
Followed by $12.5 billion from Musk's SpaceX and the xAI (Colossus Group) and $110 billion from Amazon AWS, followed by Microsoft's Azure, NScale, Lambda and AMD.
If you look at the latest financial statements of the US tech giant, you can see how exaggerated this bet is.
A startup that generated $33.2 billion in revenue in the first three quarters has surpassed its backlog of off-the-shelf commitments.It's the first time we've ever seen a company like Meta (US$866 billion), Microsoft (US$594 billion), Great Britain (US$53 billion) and Amazon (US$45.3 billion).The United States is second only to Google!
Nearly $900 billion worth of computing contracts have been signed on the books, meaning Anthropic has no way of slowing down the pace of development.
The only way to digest this computing power, which has generated nearly $70 billion in annual revenues, is for Anthropic to hand over more and more R&D work to Claude, allowing it to write code and experiment at tens of times the speed of humans!