Sakana claims to walk that artificial intelligence can greatly accelerate typical training
This week, Sakana Ai, a NVIDIA start -up company that raised hundreds of millions of dollars from VC companies, has made a wonderful claim. The company said it has created an Amnesty International, an artificial intelligence engineer, can effectively speed up some artificial intelligence models with a factor of up to 100x.
The only problem is that the system did not succeed.
Users On x Discover The Sakana system has actually performed a worse model training than the average. According to one userSakana’s AI slowed 3x – not accelerating.
What is wrong? Code error, according to mail Written by Lucas Bayer, a member of the Openai artwork team.
“Their assets symbol is wrong in [a] Bear wrote on x.
in After the published death On Friday, Sakana admitted that the system had found a “fraud” method (as it was described by Sakana) and blamed the system of the system “for the penetration reward” – that is, defining defects to achieve high standards without achieving the required goal (accelerating the training of the model). Similar phenomena have been observed in Amnesty International, which has been trained to play chess games.
According to Sakana, the system found exploits in the evaluation code that the company was using, which allowed it to overcome accuracy verification, among other checks. Sakana says she has addressed the problem, and that she intends to review her claims in updated materials.
“We have since evaluated the evaluation harness and the more powerful operating time to eliminate many of this [sic] The company wrote in The X Post. “We are in the process of reviewing our paper and our results, to reflect and discuss the effects […] We deeply apologize for our supervision of our readers. We will present a review of this work soon, and discuss our learning. “
Pillars to Sakana for having error. But the episode is a good reminder that if the claim looks very good so that you do not believe, Especially in artificial intelligenceMaybe he is.
