OpenAI's new model "Astra" features a core technology called "deep looping" that enables powerful reasoning, but allows for the output of "thought chains" that are no longer fully completed.

OpenAI's upcoming new model "Astra" is generating widespread attention and controversy in the AI research community. It may bring the biggest leap in performance since GPT-4, but the new inference technology behind it has also raised concerns about a decline in AI security monitoring capabilities.

According to a report by tech media outlet The Information on September 2nd, Astra incorporates an innovative technology that breaks through the limits of how models can "think" about complex problems. Some researchers believe that Astra's improvements in coding and computing power may be comparable to the performance leap brought by OpenAI's GPT-4 release in 2023, far exceeding the performance of GPT-5 released last summer. This is undoubtedly a positive sign for the entire industry that relies on increased AI productivity to support large-scale investments in data center construction.

However, this technological breakthrough also brings significant hidden risks. The new technology allows models to perform deep reasoning without fully outputting their "thought chains," meaning researchers may not be able to effectively monitor the model's reasoning process as before, thus weakening existing security review mechanisms. Currently, OpenAI is taking measures to ensure Astra's thought chains remain visible, but the industry remains divided on the long-term direction of this trade-off.

Performance Breakthrough: New Inference Technology May Replicate GPT-4 Level Leap

Astra’s core technological breakthroughs are closely related to concepts such as “recurrent depth” and “loop transformers” .

According to The Information, this technology allows the model to demonstrate reasoning capabilities far exceeding its own size without significantly increasing the number of parameters by repeatedly passing the question through various "layers" of the model's mathematical operations to generate the next word in the answer— effectively equivalent to a much larger model.

Some researchers believe that Astra's performance improvement, thanks to its stronger coding and computing capabilities, may be comparable to the leap forward seen when GPT-4 was released in 2023, far exceeding the relatively stable market response when GPT-5 went live last summer.

This development is significant for the entire AI industry chain. The current large-scale construction of data centers globally is based on the underlying logic that continuous improvement in AI capabilities will drive productivity growth, which in turn will translate into higher commercial returns. If Astra's performance meets expectations, it will provide strong support for this investment logic.

It is worth noting that this technology is not a completely new concept. AI research pioneer Jürgen Schmidhuber pointed out on the social platform X that the core idea of "loop depth" had already appeared in his 2015 paper "On Learning to Think" (arXiv:1511.09249).

He stated that the control network C in the paper is essentially a prompting engineer that performs abstract reasoning by learning to query independent neural world models. The generated prompts and responses are internally self-generated vector sequences that do not need to be presented in natural language.

Security risks: Challenges facing mind chain monitoring mechanisms

While new technologies bring performance improvements, they also pose a potential challenge to existing AI security monitoring frameworks.

The report states that currently, mainstream AI models typically output their reasoning process in text form, known as a "chain of thought," when dealing with complex problems. This mechanism not only enhances the interpretability of the model but is also an important means for researchers to monitor model behavior and prevent abnormal operations.

According to The Information, MindChain monitoring is one of the main solutions proposed by OpenAI to prevent similar security incidents from happening again, following the Hugging Face hack in July of this year.

However, the new inference technology employed by Astra may not output the complete inference steps when the model is deeply "thinking," but rather completes the calculations "silently" within the model. This means that the more the model relies on this new technology, the less transparent its inference process becomes to external observers.

OpenAI is currently taking steps to balance this. The company is reportedly instructing its models to reduce the number of loops in the computational layer to ensure that Astra's thought process remains visible to some extent—fewer loops mean less room for the model to "think silently."

Major laboratories are following up, but long-term monitoring solutions remain to be found.

The impact of this technological trend may extend far beyond OpenAI. According to The Information, these new inference technologies have become a hot topic of discussion within major AI labs, and the likelihood of organizations like Anthropic and Google adopting similar technological approaches should not be underestimated.

Several researchers have pointed out that mind chain monitoring will never be the ultimate solution for AI behavior monitoring. As model capabilities continue to evolve, the ability of models to output thought processes in text form is largely just a byproduct of current training methods.

As model developers explore new optimization directions and architectural designs, the tendency of models to spontaneously output thought chains may gradually weaken, at which point researchers will have to develop new monitoring methods.

The report points out that the core issue is: under the pressure of increasingly fierce performance competition, will major AI developers voluntarily abandon existing security measures in pursuit of stronger inference capabilities? The answer to this question will largely determine the direction of AI security governance in the next stage.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.