These concessions made sense for O1, which was essentially a huge experiment, despite the limitations of the code base. They did not make so much sense for chat, a product used by millions of users that was built on a different and more reliable pile. When O1 launched and became a product, cracks began to emerge in the internal processes of OpenAi. “We thought, ‘Why are we doing this in the experimental code base, we should not do it in the main product research code?’ There was a great internal rejection. “
Last year, the company announced internally the Sputnik project, an effort to review the code and find out what parties they had to merge and which ones had to remain separate. Employees believe that the project was not fully carried out. Instead of merging the batteries, employees were encouraged to prioritize the use of the “Berry” battery, which enraged some people who worked in the chat. An Openai spokesman denies it and refutes that the Sputnik project was successfully deployed.
The problems perceived with the code base had tangible repercussions, add the sources consulted by Wired. Ideally, after an employee launches a training work, the Graphic Processing Units (GPU) involved in that work are released to use by another person. As the Berry code base is structured, that does not always happen. “People stuck in the GPUs. There was a jam,” confesses an old employee.
No one knows what to think about Deepseek
Outside OpenAi, the industry is divided on how to interpret Depseek’s success. Earlier this week, Nvidia’s actions collapsed because investors feared that the industry had overestimated the number of chips necessary to work with AI.
According to experts, that interpretation is mine. If Deepseek discovered a way to develop models more efficiently, as it claims to do so, it could accelerate the models development process, but ultimately the winner will continue to be the company with more chips. “You need less computing per intelligence unit, but people will continue to want more units to climb even more,” explains thousands longage, an independent A -policies researcher who worked in Openai for six years, and recently as the main advisor for the preparation of AGI.
Perhaps Stargate, the new and striking Openai infrastructure project, internally relieve the feeling of scarcity. Crusoe, the company that builds the first Stargate data centers in Abilene, Texas, has already put the first stone of 998,000 square feet facilities, according to Andrew Schmitt, spokesman for the company.
Although the details of the project are opaque, they have told me that it could grow until more data centers, chips and supercomputer manufacturing. OpenAI plans to appoint a new general director to direct the project, at least in theory. The current CEO, Sam Altman, “is very good to make promises of what will happen in the future. And then, in the future, those things are completely unreliable,” says another former staff member.
Article originally published in Wired. Adapted by Alondra Flores.
#OpenAi #launches #O3mini #response #Depseek
