Join the event trusted by corporate leaders for almost two decades. VB Transforma connects people building AI Real Enterprise. Learn more
GoogleA recent decision to hide the raw reasoning of its flagship model, Gemini 2.5 Pro, caused a violent response from developers who relied on this transparency for the construction and debugging of the application.
The change, which resembles a similar Openai movement, replaces the reasoning of the model step by step with a simplified sum with a simplified summary. The answer emphasizes the critical tension between the creation of the polished impression of the user and ensuring observable, trustworthy tools that the enterprise needs.
Because companies integrate huge language models (LLM) with more convoluted and critical mission systems, a debate on how much the model’s internal operation should be revealed, it becomes a decisive problem for the industry.
“Fundamental reduction” in transparency AI
For programmers, this reasoning often serves as a necessary diagnostic tool and debugging. When the model provides an incorrect or unexpected exit, the thought process reveals where its logic was lost. And this happened one of the key advantages of Gemini 2.5 Pro over O1 and O3 OpenAI.
On the AI Google programmers forum, users called the removal of this function “massive regression. “Without this, programmers remain in the obscure. Another described that he is forced to” guess “, why the model has failed, which leads to” extremely frustrating, repetitive loops trying to fix “.
Models that provide full access to their reasoning chains provide enterprises with greater control and transparency over the behavior of the model. The decision on the CTO or AI cable no longer relies on which model has the highest comparative results. It is now a strategic choice between the highest level, but the cloudy model and the more clear, which can be integrated with more certainty.
Google answer
In response to indignation, members of the Google team explained their justification. Logan Kilpatrick, senior product manager at Google Deepmind, Explained that the change was “purely cosmetic” and does not affect the internal performance of the model. He noticed that in the case of the Consumer Gemini application, hiding a long thought process creates a cleaner user experience. ” % Of people who will or read thoughts in the Gemini application are very small,” he said.
For programmers, fresh summaries were to be the first step towards program access to traces of reasoning via the API interface, which was not possible before.
The Google team confirmed the value of strict thoughts for programmers. “I hear that you all want harsh thoughts, the value is clear, there are cases of use that requires them,” wrote Kilpatrick, adding that restoring this function to the AI studio focused on the program is “something we can discover”.
Google’s reaction to programmers’ reaction suggests that the middle ground is possible, perhaps through a “programmer mode”, which again includes raw access to thinking. The need for observation will augment only with the augment in AI models in more autonomous agents that operate tools and make convoluted, multi -stage plans.
As Kilpatrick summed up in his comments: “… I can easily imagine that raw thoughts become a key requirement for all AI systems, taking into account the growing complexity and the need to observe + tracking.”
Are reasonable tokens overrated?
However, experts suggest that there is a deeper dynamics in the game than just the user’s experience. Subbarao kambhampati, professor AI in Arizona State UniversityHe asks if the “indirect tokens” produced by the reasoning model before the final answer can be used as a reliable guide to understand how the model solves problems. AND paper Recently, he will co -authored that anthropomorphic “intermediate tokens” as “traces of reasoning” or “thoughts” may have unsafe implications.
Models often enter into endless and incomprehensible directions in the process of reasoning. Several experiments show that models trained in false signs of reasoning and correct results can learn to solve problems, as well as models trained in the scope of well -crazy signs of reasoning. In addition, the latest generation of reasoning models is trained by reinforcement learning algorithms, which only verify the final result and do not assess the “trace of reasoning”.
“The fact that sequences of indirect tokens often look like better formatted and writing human work … does not tell us much about whether they are used anywhere near the same purposes to which people use them, not to mention whether they can be used as an interpretable window, what LLM” thinks “, i.e. as a reliable justification of the final answer,”, the researchers write.
“Most users can not understand anything from the volumes of raw intermediate tokens that threw these models,” said Cambhampati Venturebeat. “As we recall, Deepseek R1 produces 30 pages of pseudoangles in solving a simple planning problem! A cynical explanation why O1/O3 decided not to show raw tokens originally, because they realized that people would notice how unstable!”
After saying, Cambhampati suggests that summaries or explanations after fact can be more understandable to end users. “The problem becomes to what extent they actually indicate internal operations through which LLM has undergone,” he said. “For example, as a teacher, I could solve a new problem with many false starts and withdrawal, but explain the solution in the way I think it makes it easier to understand students.”
The decision to hide COT also serves as a competitive moat. Strict signs of reasoning are extremely valuable training data. As Kambhampati notes, a competitor can operate these traces to perform “distillation”, the training process of a smaller, cheaper model to imitate a stronger capabilities. Hiding raw thoughts significantly hinders the rivals to copy the secret sauce of the model, which is a key advantage in the intensive industry.
The debate about the thoughts chain is a preview of a much greater conversation about the future of AI. There is still a lot to learn about internal actions regarding reasoning models, how to operate them and how far models suppliers are ready to go to allow programmers to access them.