It is a little more than a week since Deepseek used up the AI world. The introduction of its open weight model-apparently trained on a fraction of the specialized computer chips that deposited leaders of the industry of the industry-held shock waves in OpenAI. Employees not only claimed to see hints that Deepseek had wrongly distilled the models of OpenAi to create themselves, but the success of the startup wondered whether companies like OpenAi Wild spend too much on the calculation.
“Deepseek R1 is the Sputnik moment of AI,” wrote Marc Andreessen, one of the most influential and provocative inventors of Silicon Valley, on X.
In response, OpenAI is preparing for launching a new model today, prior to the originally planned schedule. The model, O3-Mini, will debut in both API and Chat. Sources say that the O1 levels have reasoning with speed on 4o-level. In other words, it is fast, cheap, smart and designed to crush Deepseek. (OpenAi spokesperson Niko Felix says that the work on O3-Mini long before the debut of Deepseek started and the goal was to launch at the end of January).
The moment has Galvananed OpenAi staff. Within the company there is a feeling that – in particular if Deepseek dominates the conversation – OpenAi must become more efficient or the risk of running out on his newest competitor.
Part of the problem stems from the origin of OpenAi as a non-profit research organization before becoming a profit-seeking powerhouse. A continuous power struggle between the research and the product groups, claims employees, has resulted in a gap between the teams working on advanced reasoning and those who work on chat. (OpenAI spokesperson Niko Felix says that this is “incorrect” and notes that the leaders of these teams, Chief Product Officer Kevin Weil and Chief Research Officer Mark Chen, “Meet every WEK and work closely to get up product and research priorities to vote. “)
Some within OpenAI want the company to build a uniform chat product, a model that can see if a demand advanced reasoning requires. That has not happened so far. Instead, a drop-down menu in chatgpt requires users to decide whether they want to use GPT-4O (“great for most questions”) or O1 (“Capturing faltered reasoning”).
Some employees claim that, although chat provides the lion's share of OpenAi's turnover, O1 receives more attention – and calculating sources – from leadership. “Leadership does not care about chat,” says a former employee who (you guessed it) chat. “Everyone wants to work on O1 because it is sexy, but the code basis is not built for experiments, so there is never a momentum.” The former employee asked to remain anonymous, with reference to a non -dismay agreement.
For years, OpenAi has experimented with learning reinforcement to refine the model that was ultimately the advanced reasoning system called O1. (Strengthening education is a process that trains AI models with a system of fines and rewards.) Deepseek built the learning work that OpenAi had pioneer to create his advanced reasoning system, called R1. “They benefited from knowing that reinforcement learning, applied to language models, works,” says a former OpenAi researcher who is not authorized to speak about the company publicly.
“Learning the reinforcement [DeepSeek] DID is similar to what we did at OpenAI, “says another former OpenAi researcher,” but they did it with better data and cleaner stack. “
OpenAi employees say that research that went to O1 was done in a code base, called the “Berry” pile, built for speed. “There were considerations-experimental strictness for transit,” says a former employee with direct knowledge of the situation.
Those considerations were logical for O1, which was essentially a huge experiment, despite the limitations of the code basis. They were not so logical for chat, a product used by millions of users that was built on a different, more reliable stack. When O1 was launched and a product was started, cracks started in the internal processes of OpenAi. “It was like:” Why do we do this in the experimental codebase, should we not do this in the main product research code base? “, The employee explains.” There was a big pushback for that. “
