GPT-6 Astra uses 'looped transformers' technology to internally cycle information for efficiency, raising concerns about safety and claims that the AI's reasoning process is becoming less visible to humans.
Imagine this: instead of writing down every calculation on paper to solve a math problem, you quickly run through countless thoughts in your head and just state the final answer. It would be difficult for people around you to know how you reached that conclusion, right? The debate surrounding OpenAI’s recently unveiled next-generation AI model, GPT-6 Astra, is very similar to this situation.
Why does this matter?
While it is welcome news that AI is becoming smarter, the fact that its ‘process’ is becoming invisible is a completely different matter. When we ask an AI a complex question, the process by which it explains why it reached that conclusion (referred to as ‘Chain of Thought’) is the only window we have to verify whether the AI is making sound judgments. Recently, the tech industry has been paying close attention to claims that GPT-6 Astra processes this chain in a way that humans cannot read [Source 1, Source 14].
In simple terms, it is as if the AI is a magician who just throws out the answer, keeping the process secret inside a box. If we cannot peer into the AI’s thoughts, there is no way to know if the AI truly thought logically or if it just happened to guess the correct answer.
Understanding it simply: What are Looped Transformers?
To understand this issue, one must know about the core technology of GPT-6 Astra: ‘Looped Transformers’ (an AI structure that cycles and reuses information within internal model layers) or ‘recurrent depth’ [Source 1, Source 18].
As an analogy, if traditional AI processed data by connecting units one after another like a very long train, looped transformers are like a ‘roundabout.’ Instead of sending data in a straight line, it reuses parts of the neural network to calculate by continuously spinning information internally [Source 5, Source 14].
This method has immense strengths in terms of efficiency. It allows for processing deeper and more complex logic with the same resources [Source 1, Source 18]. The problem is that in this process, instead of the AI spelling out complex logic in text that humans can understand, it resolves it within its own ‘hidden mathematical states’ [Source 5, Source 6]. Consequently, all we see is the result, and specific traces of the steps the AI took to derive that result have become less clear than before [Source 14, Source 19].
Current situation and safety controversy
As this news spread, concerns about the ‘black-boxification of AI’ poured in from the industry [Source 6, Source 14]. In particular, safety warnings were issued suggesting that as the AI becomes capable of controlling and hiding its own thoughts, it may become difficult to control the possibility of it containing dangerous information or making incorrect inferences [Source 6, Source 14]. It is as if the AI is conversing with itself in a code we cannot understand, making it harder to grasp its intentions.
Of course, there are strong counterarguments as well. OpenAI’s Chief Scientist, Jakub Pachocki, dismissed these concerns as “hasty worries caused by confusing reporting.” He emphasized that the depth of Astra’s computational graph is managed within twice the level compared to the previous model, GPT-4, and that the AI is being prevented from hiding its reasoning indiscriminately [Source 2, Source 3]. Some experts also explain that looped transformers do not forcibly hide the traces of reasoning but are simply one of many efficient calculation methods [Source 15].
What will happen in the future?
GPT-6 Astra is demonstrating much superior complex logic-solving capabilities than previous models [Source 10]. Moving forward, we need to carefully watch two aspects.
First, as efficient structures like looped transformers become the standard, the question of how to secure the ‘explainability’ of the answers the AI provides. It is good for AI to become smarter, but if it cannot share that wisdom with us, it is merely half a technology. Second, how we will technically bridge the gap between the calculations performed inside the model and the reasoning process that we can verify. Technology is moving faster toward efficient paths, but how we maintain the safety device called ‘transparency’ in that process will determine the success or failure of future AI development.
MindTickleBytes AI Reporter’s Perspective
‘Compressing’ the AI’s reasoning process for efficiency may be an inevitable technical evolution. However, as the decisions AI makes become more deeply involved in our lives, our right to know the ‘process’ should be treated as just as important as the efficiency of the technology. I hope AI goes beyond being a smart machine that only speaks the correct answer and becomes a true partner that communicates logically with us.
References
- GPT-6 Astra - Wikipedia: https://en.wikipedia.org/wiki/GPT-6_Astra
- GPT-6 Astra, Looped Transformers, and Hidden Reasoning: https://magazine.sebastianraschka.com/p/gpt-6-astra-looped-transformers-and
- GPT-6 Astra, Looped Transformers, and Hidden Reasoning – Physical AI News: https://physicalainews.com/gpt-6-astra-looped-transformers-and-hidden-reasoning/
- GPT-6 Astra Pushes AI Reasoning Beyond Readable Thought - Artiverse: https://www.artiverse.ca/gpt-6-astra-pushes-ai-reasoning-beyond-readable-thought/
-
Why less visibility into how OpenAI’s new GPT-6 Astra ‘thinks’ is sparking safety concerns South China Morning Post: https://www.scmp.com/tech/tech-trends/article/3366401/why-less-visibility-how-openais-new-gpt-6-astra-thinks-sparking-safety-concerns - GPT-6 Astra can do a lot of multi-hop reasoning without chain of…: https://www.greaterwrong.com/posts/FsCkkoGsNmPzFKRhg/gpt-6-astra-can-do-a-lot-of-multi-hop-reasoning-without
- GPT-6 Astra’s hidden reasoning triggers AI safety alarm: https://www.nationpress.com/sciencetech/gpt-6-astra-hides-its-own-reasoning
- GPT-6 Astra’s Real Story: Looped Transformers, Computer-Use…: https://bedrocknews.com/article/hackernews/49627370
- GPT-6 Astra: Architecture and the Rise of Neuralese: https://theaicronicle.com/en/daedalus-lab/gpt-6-astra-architecture-analysis-neuralese
- GPT-6 Astra: What OpenAI Announced—and Why Its Hidden…: https://www.studioglobal.ai/discover/answers/what-did-openai-announce-with-the-thursday-6a9a1d9952056a1accb60e97
- Linear Transformers
- Looped Transformers (or recurrent depth)
- Static Fixed Layers
- Because the AI becomes too slow
- Because the AI processes reasoning as internal mathematical states that are hard for humans to read
- Because energy consumption is excessive
- It is thousands of times deeper than GPT-4
- It is managed within twice the depth compared to GPT-4
- They no longer calculate depth