Sakana walks back revolts that their ai can the shooting the model training

This week, Sakana Ai, an delivery sharp entitles one hundred dollar dollars from the VC business, fencing a detected question. The company had been created a system ai, the avenue engineer, which might effectively catch the formation of certain patterns of a factor of up to 100x.

The only problem is, the system did not work.

Users On x quickly discovered Sakana system really turns out in the greener model training performance. According to a userAI sakana has resulted in a slowdown 3x – not a speedup.

What went wrong? A bug in the code, according to post office from Lucas Beyer, a member of the technical personnel in Openai.

“Their Origin Code is incorrect in (a) below,” Beaser wrote on X. “The fact that they run benchmark twice with the wild results and think.”

In a Postmortem posted Friday, Sakana admitted that the system has found a way to – as Sakana described the system tendency to achieve the purpose of unable to achievement. Similar phenomena has been observed in Ai is trained to play chess games. I am

According to Sakana, the system found in the Code of Value that company is used to be used to allow them to turn on validations for accuracy, between other controls. Sakana says he addressed the problem, and what are you intending to revise their claims in updated materials.

“We have to have theft of the assessment and the most robust robbery to eliminate many such loofoles (sic),” the company wrote in the x place. “We are in the process of reversing our card, and our findings, to reflect (…) We are in the deep end to our sincere. We will go a revision of this job soon, and discuss our students “

Points to Sakana to own the mistake. But episode is a good memory that if a question sounds too good to be true, especially in aiIt’s probably.