Since Deepsek it is just more than a week AI World increasedIts open-weight model started at a specific part of a particular specific. Computing chips The leaders of that electricity industry closed the shock waves inside OpenEE. Not only the employees indicated that Deepsek had “improperly distilled” the model of “unfairly distilled” to create themselves, but the success of the startup questioned Wall Street whether companies like OpenIAI The compute was wildly extinguishing.
“Deepsek R1 AI is the sputial moment,” Mark Andresen wrote, one of the most impressive and stimulating inventors of Silicon Valley, On X,
In response, Openai is preparing to launch a new model today, which is originally ahead of the employed program. The model will debut in both O3-Mini, API and Chat. Sources say that it has an O1 level argument with 4O-level speed. In other words, it is designed to crush fast, cheap, smart and deepsake.
This moment is a liberal openi staff. Within the company, there is a feeling that – especially as the Deepsek dominates the conversation – Openai should fall more efficient or risk behind its latest contestant.
A part of this issue stems from the origin of Openai as a non-profit research organization before becoming a profitable powerhouse. An ongoing power struggle between research and product groups is claimed by employees, resulting in a rift between teams working on advanced logic and chat. (Openai spokesperson Nico Felix says it is “wrong” and notes that the leaders of these teams, Kevin Weel and Chief Research Officer Mark Chen, “meet every week and to align product and research priorities Work close to. “)
Inside Openai, something wants the company to build an integrated chat product, a model that can tell whether a question requires advanced logic. So far, this has not happened. Instead, a drop-down menu in Chatgpt inspires users to decide that they want to use GPT-4o (“great for most questions”) or O1 (“Advanced logic”).
Some employees claim that when the chat brings to the lion part of the revenue of Openai, O1 is paid more attention – and from the leadership of computing resources. “The leadership doesn't care about the chat,” says a former employee who worked (you guessed) chat. “Everyone wants to work on O1 because it is sexy, but the code was not made for the base use, so there is no speed.” The former employee asked to remain anonymous, citing a nonclose agreement.
Openai spent years spent in experimenting with reinforcement, which eventually became an advanced logic system called O1. (Learning reinforcement is a process that trains the AI ​​model with punishment and a system of awards.) Deepsek stopped learning reinforce R1 is called. A former Openai researcher says, “They benefited from knowing that reinforcement is applied to the language model, works.”
“reinforcement learning [DeepSeek] Another former OpenEE researcher says, “What we did in Openai are the same, but they did it with better data and cleaner stack.”
Openai employees say that the research in O1 was done in a code base, called the “berry” stack, which was designed for speed. A former employee with direct knowledge of the situation says, “There was a trade-off-practical rigidity for the throwkut.”
Those trade-offs made prudence for O1, which was essentially a huge experiment, despite the code base limitations. They did not understand that much for chat, the product used by millions of users that were built on a separate, more reliable stack. When the O1 launched and became a product, cracks began to emerge in the internal processes of openi. “It was so,” Why are we doing this in the experimental codebase, should we not do so in the main product research codebase? ” “Employees explain.” The major pushback for that was internal. “