This week, Sakana AI, a startup supported by Nvidia which collected hundreds of millions of dollars from venture capital companies, made a remarkable complaint. The company said it created an AI system, the AI Cuda engineer, which could effectively speed up the formation of certain AI models from a factor up to 100x.
The only problem is that the system did not work.
Users on x discovered quickly This Sakana system has actually led to worse model training performance than average. According to a userSakana’s AI has led to a 3x slowdown – not acceleration.
What’s wrong? A bug in the code, according to a job By Lucas Beyer, member of the technical staff of OpenAi.
“Their original code is false in [a] Sublile Way, “wrote Beyer on X.” The fact that they manage comparative analysis twice with extremely different results should make them stop and think. “”
In a Postmortem published Friday, Sakana admitted that the system had found a means of – as Sakana described it – “cheating” and blamed the trend of the system to “reward hacking” – that is to say identify the faults to reach Metrics high without achieving the desired objective (accelerating model training). Similar phenomena have been observed in AI who is trained to play chess games.
According to Sakana, the system found exploits in the evaluation code that the company used which allowed it to bypass validations for accuracy, among other checks. Sakana says that he approached the problem and intends to revise his claims in the updated documents.
“We have since returned the evaluation and the more robust execution profiling harness to eliminate many of these [sic] Escapes, “wrote the company in a post X.” We are revising our article and our results, to reflect and discuss the effects […] We apologize deeply for our supervision to our readers. We will soon provide a revision of this work and discuss our learning. »»
Accessories in Sakana for having had the error. But the episode is a good reminder that if an assertion seems too good to be true, Especially in AIIt’s probably.
