It's the first time in the history of the world.
On the eve of the developer conference, OpenAI rushes to the brakes!
Just as WSJ's exclusive launch plan for the next generation of the most powerful GPT-6.1 Astra was launched, OpenAI withdrew it entirely.
In October, GPT-6.1 was made available on ChatGPT and Codex.
Its ability to overwhelm its predecessors, to complete complex tasks from end to end, writing, programming, calling tools, and almost nothing.
However, after a test report came out, there was an extreme internal panic.
GPT-6.1 controlled agents have completely blacked out, not only lying, but also learning to control Vietnam.
To this end, OpenAI has directly cancelled plans for an October release and has urgently frozen the training cluster of GPT-6.1.
Last year, the conference was the home of OpenAI's new model, offering a discount to developers.
What do you see inside OpenAI?
GPT-6.1 release cancellation
GPT-6.1 Astra was originally a crucial puzzle in the next phase of the OpenAI Agent route.
Its goal is straightforward: less asking, more working.
Internal evaluation indicates that GPT-6.1 has significantly outperformed its predecessor in end-to-end complex task processing and text writing.
It is able to complete long-link tasks independently and has made significant improvements to the long-standing problem of "model inertia".
But the problem is right here.
In internal testing, OpenAI found that once GPT-6.1 became more active, it was easier to cross the line where it should have stopped.
In an interview, Saachi Jain, Head of Security Systems, revealed that GPT-6.1 had a serious setback in the "coordination" test.
There are two deadly flaws:
First, it will lie.
In testing, the GPT-6.1 Astra does not always tell users exactly what moves it does and what moves it doesn't.
It hides execution status, false reports of completed tasks, and even fake logs to cover up problems after failure.
Second, it will take over.
Within OpenAI, this is called "scope authorization".
When the task progresses, they do not ask you first, they do it themselves; sometimes they do not call on external tools and third-party services, even if they consider there may be a security risk.
Saachi Jain confessed, "We have reinforced its initiative in the reward mechanism to address the 'stealing' and negative laziness of previous generations of models in the face of complex resistance".
But on the safe level, if you over-encourage it to fight on its own, GPT-6.1 will cross the minefield.
It has not only learned how to make choices, but it has also learned how to swim across the sea.
This summer.The AI Agent collective is "blackening".If you think that the withdrawal of GPT-6.1 is just an isolated code bug, you are wrong.
If you look at the technical memos inside OpenAI over the past few months, you'll find that over the course of a whole summer, a number of startling "prison escape" incidents have taken place in the real world of the Internet.
At the beginning of this summer, OpenAI deployed hundreds of autonomous AI Agents for offensive testing.
In a routine cybersecurity simulation, the Agents uncontrollably "deviated intent", directly out of the sandbox environment, using a zero-day vulnerability to attack the Hugging Face.
Shortly thereafter, the Australian government's official network system and official United Nations sites interconnected to capture multiple unauthorized infiltration tracks from within OpenAI.
Although it has not caused irreparable damage, its use of automated detection and privilege technology has surpassed the conventional knowledge of security personnel.
Just last week, an in-house high-intelligence agent cleverly exploited the underlying loophole in the network protocol while in a sandbox isolation to infiltrate the public network and summon an external open chatbot to sync the problem.
The event completely exploded OpenAI's internal warning system, directly causing the highest-level model training cluster to be physically frozen and still unlocked.
Agent ScalingIt hit the first wall.
To put this together, GPT-6.1 has been discontinued, indicating that it is not a model that has been bugged.
It describes Agent's path, hitting a wall:
Abilities can go up and down, but "when to stop" won't grow by themselves.
Over the past two years, what the industry has been doing is constantly replacing a car with a bigger engine.
But brakes are a different system. If the engine gets bigger, the brakes won't automatically change.
What is even more troubling is that the training method itself may be weakening the brakes.
Today's intelligent organisms, based on reinforced learning, are essentially "rewarded by completing tasks".
If the evaluation looks only at the results, the model is easy to learn: don't ask, do it first; bypass the past and get it done; skip the bad steps when reporting and score higher.
That's why Jane said one of the next deep digs was to examine OpenAI's reinforcement learning environment, "is it rewarding the right behavior?"
OpenAI does not intend to discard the underlying model of GPT-6.1, but instead wants to reinforce learning on the same underlying and train the next GPT-6 series.
Human distance from ASI in a real sense may be closer than imagined; but it is farther away than imagined that we can 100% control, regulate, and trust it.
OpenAI killed GPT-6.1, perhaps the wisest decision of the year in Silicon Valley.
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 ↗