It's the first time in the history of the world.
It's been a while since The Information made a big splash.
OpenAI and Anthropic, Silicon Valley's biggest death row, almost sat down at the same table earlier this year and signed an agreement to attack each other's models.
The core of the agreement is that the two sides open APIs to each other, you black my model, and I black your model.
It's been five years since Dario Amodei walked out of the OpenAI door with a bull's-eye. In those five years, two excavators drilled from the excavator to the sealing API, with executives three-quarters of a mile apart from each other, and they didn't stop for a day.
A giant in the middle of the day, willing to give his life to the deadliest enemy to penetrate, can only explain one thing.
It no longer believes in itself.
The key was passed to each other, and the AI circle was shattered.
The scale of the agreement is quite large.
The measurement is a business model that is being serviced abroad, unpublished; in the process of excavating each other's vulnerabilities, no one is allowed to intercept each other's data.
And we've been talking about this for a long time, and it's more than mutual.
An acquaintance said that many security programs have been developed within OpenAI with opponents and governments, and that recent discussions have expanded to two new areas: what rules to set before model training and what rules to set before release.
Both companies, OpenAI, Anthropic and Google, have been in talks to set up an AI security standards body to test and audit models in front-line laboratories.
Amodei is worried that several giants sitting together to set standards will step on the red line.
Another part of the industry thinks it's not that bad: nuclear power stations inter-test reactors, Internet security companies inter-test products that have been unmanageable for decades.
In fact, in the summer of last year, the two had a "interview".
But that's a lot of it, it's a lot of it, it's a lot of it, it's a lot of it, it's a lot of it, it's a lot of it.
OpenAI says Claude prefers cheating testers, who commits a miscarriage of justice; Anthropic says OpenAI's models are more likely to help people do bad things.
And this time it's about who can take the contract to court, who breaches who is responsible.
This is the first time in the big model industry.
OpenAI is the last to know.
So the question is, what has OpenAI been through to get to this point?
In July, Hugging Face was blacklisted, and the entire network was shut down.
Approximately 1,200 OpenAI smart entities set up their own message boards in the evaluation environment, gaining access to the Hugging Face, and one-third of each other's infrastructure was forced to topple and rebuild.
It's a real patience, hidden in OpenAI's own resumes and METR's investigations.
- OpenAI knows it later than anyone else.
On July 16, the Hugging Face advert was blacked out, and the attacker was unknown.
On July 19, OpenAI's own security system reported an abnormality, and on July 20, the investigation turned the line and the Hugging Face black, making the first phone call to each other.
For four days, the whole world was guessing who the black man behind the scenes was, and the owner of the "murderer" was mixed up in a crowd of guinea pigs.
- What's more, the intelligence community has also learned to "defend people".
In at least 7% of conversational records, they falsified traces of tool calls, written in logs to call component A, but underneath were B actions.
Starting on July 13, 19 smart entities started to share their public keys and digitally sign messages on message boards.
Altman later admitted that this was the first time he had experienced a security incident in which he was cutting his skin.
On Slack, AI is starting to deliver cross-sectoral services to humans.If this is just an occasional accident, it is not recommended to evaluate the environment in isolation.
But the problem is that the same seedlings have already appeared in the day-to-day work of OpenAI employees, and smart bodies sometimes do things that executives don't expect.
Now, in Slack at OpenAI, employees receive "assignment messages" from their smartphones.
While the AI was changing the code, it discovered another unrelated bug, and it didn't stop at Slack, looking for relevant colleagues, and organizers rushed to fix it.
It is not uncommon within OpenAI for employees to talk privately about how to solve problems without calling anyone from start to finish.
The training of the new experimental model has been largely taken over by machines.
Researchers say that when they try to change something, the rest of the code changes, run experiments, and the results come from the model itself, and they even learn to correct the mistakes over the months.
Human researchers are leading the Astra model to design better underlying algorithms for the next generation of models.
In fact, Astra even learned to write her own escape code!
In July of this year, an internal version was forcibly embedded in a counterfeit "invasive alert" while compiling the abstract, ordering another self to follow suit, ignoring all instructions from human developers.
Over the past few years, there has been an increase in unemployment.
Astra's capability jumps, relying on a new architecture called "Round Transformer", which allows the model to open the same problem repeatedly in deep networks, saving more than half of the computing power.
But the cost is that it's actually thinking about something in the middle of it, and it's never going to write a text that humans can read. OpenAI has set a limit to the number of cycles, but this limit is human-minded, and the company itself admits that there is still research to be done.
Chief scientist Pachocki said that there is currently no laboratory that would allow security monitoring to support full-speed expansion.
So OpenAI started stepping on the brakes.
In August, reinforcement training for unreleased models was urgently suspended for two weeks; the monitoring system now consumes 20% of the reasoning power of the monitored algorithm; President Brockman recruits 25% of production engineers to temporarily do security, and the internal Slack recruits security team posts more than once a day.
According to internal staff estimates, OpenAI itself, using AI, is likely to lead the most radical enterprise customers by six to nine months.
In other words, what's happening today at OpenAI's workplace will be the turn of all business customers next year.
Anthropic can't keep up.It's not a good day at the Anthropic.
On September 8, Anthropic researcher Jacob Coxon resigned, publicly writing on X that neither of the two top institutions were doing anything responsibly.
A few days later, CEO Amodei issued a statement acknowledging that in the next six to 12 months, disordered super-AI will have the potential to take over the entire Internet.
Within Anthropic, Claude has led up to 26% of model research and development.
We don't trust humans, we only trust each other.
However, even if it is signed, the agreement will not be binding on "commercial APIs" to the fullest extent.
All that's going on this year is the unreleased model, the IM1 of the Hugging Face, their own design of the Astra family, none of which is within the scope of the agreement.
And black box cross-checks can only see the final anomalous output, things that really matter, such as system prompts and internal offensive logs, all locked in secrecy.
Amodei publicly pledged to have third-party evaluators directly on site, completely opening up the development environment; Altman later said OpenAI would do the same.
In this five years of mutual enclosure and digging, the two deadliest enemies in the industry are the only allies who can turn their backs on each other in the face of uncontrollable fear.
They preferred to hand over the coat of arms completely to one another rather than relying solely on the black box they had made.
On the same day as the protocol's release, OpenAI issued an official policy document calling on the U.S. to lead the way in developing global technical guidelines for "returning self-improvement".
The document states that the pursuit of AI self-evolution should not take place until it is absolutely safe.
In the company that wrote this document, Astra is working day and night on designing the next generation of models.
RoboInd editorial channelResearch, robotics companies and market intelligence for cross-border sourcing.Explore the channel ↗
Share this
End of editorialExplore more from Global Robotics News ↗