OpenAI raises alarm among experts: the recurring depth of its Astra models will make reasoning more opaque
After wowing us with language models that immediately responded to our queries, labs like OpenAI released something better: reasoning language models. One of their characteristics is that, instead of responding instantly, they spend time creating a thread of reasoning in which the model converses with itself. Thanks to this technique, which of course involves greater computing time, the answers are more reliable. Not in vain, they are much more restful, although also more expensive.
One of the graces of this entire process is being able analyze what the reasoning of the model has been. That helps, among other things, determine whether you took the smartest path or why you made a huge mistake. However, OpenAI wants to change the game with something they call “recurrent depth” or “opaque recurrence.”
When the model reasons, but does not explain
That “opaque recurrence” reported by media such as The Information either TechCrunch would arrive in the next OpenAI models, called Astra. What we are referring to is the fact that the model reasons about the user’s request, but without informing at all what its line of reasoning has been. Having this information, although it may not seem like it at first, is crucial to exercising control over the software.
The reasoning of the model allows traceability to be applied and its responses evaluated with greater reliability. Thanks to the chain of thought, it is much easier to audit the results and know why the model reached certain conclusions. As sources recall, it was these chains of thought that precisely allowed us to better understand why OpenAI’s AI agents attacked Hugging Face.
As you see, “opaque recurrence” is not a new technology and approach that improves the models, but a way of further hide how they work inside. The neural network processes information cyclically and resolves a good part of the calculations directly in what they call the “latent space.” In other words, it does not emit intermediate tokens that humans can read.
Experts warn about this behavior of laboratories
Experts like Buck Shlegeris, who heads Redwood Research, have already raised eyebrows. Your fear is more than reasonable. If this technique becomes normalized, forensic monitoring tools are no longer really useful.
For his part, his colleague Ryan Greenblatt warns that these non-linear methods can evolve at a speed that leaves us without tracking capacity. Likewise, analysts like Zvi Mowshowitz are already suggesting that it is time to call regulators before this race for efficiency ends up blowing up control standards. Until today, all the large laboratories maintained the unwritten agreement to be as transparent as possible with these logical processes, but it seems that this consensus is beginning to break down.
What OpenAI says
Faced with these criticisms, OpenAI’s scientific management has defended that the integration of recurrent depth in Astra is limited and does not suppress the generation of understandable explanations. The firm’s scientific director, Jakub Pachocki, has assured that the laboratory maintains its commitment to the supervision of chains of thought as the axis of its research program. In fact, according to the company, a total transition towards totally opaque internal languages is ruled out.
Of course, this doesn’t just affect OpenAI. Although their Astra models have been the center of criticism, the truth is that both Anthropic and Google are studying similar techniques. After all, the more opaque a model is, the harder it is for other companies to understand and distill it.
