AI News

GPT-6.2 Delayed for Weeks Due to Safety Alignment Issues, OpenAI Accused of Disrupted Release Cadence

36Kr · AIUpdated Nov 16 21:10

AI summary

Leaks claim that OpenAI's GPT-6.2 has been delayed for several weeks due to safety and alignment issues, with GPT-6.5 also affected; the core challenge is how to turn pre-training results of over 10 trillion parameters into a product usable by the public. Previously, GPT-6.1 Astra, originally scheduled for an October release, was also canceled due to safety and alignment standards not being met. OpenAI officially stated that a boundary-crossing incident on September 20 prompted the company to pause training and evaluation of some frontier models. These claims currently come from community leaks, and OpenAI has not yet officially confirmed them.

Body · AI translation

OpenAI has finally ended up just like Google did in 2022:

One teaser and promise after another, yet the new models are nowhere to be seen, and GPT-6.2 keeps getting delayed again and again.

And the core reason behind it is, once again, "safety."

In a widely circulated "deleted" post, OpenAI engineer Joe spilled the beans:

OpenAI has become one of the companies that takes safety most seriously in the whole world. OpenAI is now a safety company.

What that means is that at OpenAI, safety anxiety has been pushed to the top of the list, and releasing new models has instead become a secondary task.

Frankly, this has a bit of a Google flavor to it.

(Google in 2022 held leading technology but kept hesitating to release products out of concerns over safety and brand reputation, and could only watch as OpenAI snatched away the first-mover advantage in generative AI.)

And just now, AI tipster Token Gremlin offered an even more specific account: OpenAI's next-generation models are being delayed due to safety and alignment issues, and the delay has already lasted several weeks.

Moreover, it is not just GPT-6.2, which was next in the release order; he also named GPT-6.5 at the same time, and pointed out that OpenAI is facing a thorny problem:

How to turn the results of pre-training with more than 10 trillion parameters into a product that the public can actually use.

At present, this tip has sparked heated discussion in the community, and related topics have also climbed onto the Twitter trending list.

Speaking of which, the reason this matter is worth paying attention to is that as early as early November, people in the community had been claiming one after another that they had already been given gray-scale access to GPT-6.2.

Even The Information revealed on November 4 that OpenAI's new Sol, Astra and cyber models might be launched as soon as that Friday, which was November 6.

As a result, before anyone knew it, half a month had passed, and not only had the new Astra still not arrived, but other models had also been delayed again and again.

In contrast, its good brother Anthropic released Claude Fable 5.2 and Mythos 5.2 in succession on November 9~

Could it really be that OpenAI has truly drawn a prison cell for itself over safety issues, Lina!

Or is it that... Altman is preparing to wait until Anthropic goes public on the 19th and then look for a chance to snipe at it?

GPT-6.2 is still on the way

According to Token Gremlin's latest account, the big listing gift package Altman prepared for Dario will probably have to be delayed because of safety alignment issues.

In addition to the delay of several weeks, he also offered a more pessimistic judgment: these problems will be difficult to resolve in a short period of time, and the pace of model releases will also be hard to restore quickly.

Even if OpenAI launches GPT-6.2 in the near future, it may have to go online with a large number of guardrail restrictions, while other versions continue to be pushed back.

But interestingly, just a day earlier, Gremlin had also shared a case of using GPT-6.2 Astra to make a game.

We do not know what information made Gremlin suddenly flip-flop, but just over a week ago, he was actually much more optimistic about OpenAI's new models.

On November 7, Token Gremlin said that GPT-6.2 Astra and Sol failed to launch as expected because serious safety and alignment issues were discovered at the last minute and needed additional fixes.

At that time, his expectation was that after a few more days of polishing, the new models should be able to meet everyone.

Almost at the same time, OpenAI's Sebastian Sanso also replied in a quote saying:

It is worth spending a few more days to get safety right. Be patient, good models are coming soon.

But as things stand now, no new model has been released, and OpenAI officials have not responded further on this matter.

Gremlin's tone, meanwhile, has shifted from waiting a few more days to multiple generations of models being blocked one after another.

So why is OpenAI watching Anthropic keep the music playing and keep dancing, while it itself makes no move at all?

In this regard, in another post, Token Gremlin also directly cited Joe's post:

He linked its description of changes in the internal safety culture with the delay of GPT-6.2, arguing that after the Hugging Face incident, OpenAI became especially cautious about clearing new models for release.

Although this is somewhat riddle-like, safety issues affecting R&D and releases do have precedents at OpenAI.

Previously, OpenAI suspended part of its training due to out-of-bounds model behavior, and canceled the October release plan originally set for GPT-6.1 Astra because safety and alignment fell short.

OpenAI officials also said on October 23 that an out-of-bounds incident on September 20 prompted the company to suspend training and evaluation of some frontier models.

It can be said that, at least judging from OpenAI's previous public actions, safety issues are indeed an important variable affecting the iteration pace of frontier models.

Is OpenAI scared, or is there really a problem

Regarding OpenAI's repeated delays, many netizens have also started speculating.

For example, Lumina posted:

Is OpenAI trying to wait for Anthropic to launch Fable 5.2 first and see what cards its opponent is holding? But instead the other side got ahead, and even GPT-6.2 Astra had to continue to be delayed?

Token Gremlin, meanwhile, attributed the difference between the two companies to their accumulated safety work.

In a follow-up post, he said that although he has no goodwill toward Anthropic, he must admit that the company's long-term emphasis on safety has paid off:

In his judgment, Anthropic's new models are already able to be delivered publicly, while OpenAI has fallen months behind at this stage.

Of course, these claims still come from community tips and personal judgments, and OpenAI has not yet made an official confirmation.

But setting aside exactly when GPT-6.2 will be released, a more noteworthy question has already emerged:

As model capabilities grow stronger and stronger, competition among frontier labs may be shifting from who can train a stronger model to who can deliver a stronger model safely and stably to users.

Especially as RSI (recursive self-improvement) accelerates the model R&D process, pre-training scale and model capabilities are still climbing, but that does not mean capabilities can be directly converted into products.

Safety guardrails, alignment training, deployment plans, and the reliability of models in real usage scenarios will all affect the final delivery speed.

Reference link

This article comes from the WeChat public account "Quantum Bit", author: henry, published by 36Kr with authorization.

Source

36Kr · AI