This week, NVIDIA-backed startup Sakana AI raised hundreds of millions of dollars from venture capital firms to file outstanding claims. The company said it created an AI system, AI CUDA engineers, that can effectively speed up training on certain AI models up to 100 times.
The only problem is that the system does not work.
user On X Discovered soon Sakana’s system actually results in poorer model training performance than average. According to a userSakana’s AI caused a 3x slowdown – not acceleration.
What’s wrong? An error in the code, according to postal Written by Lucas Beyer, technician at Openai.
“Their original code is wrong [a] Beyer wrote on X. “A subtle way. They benchmarked twice with very different results, which should make them pause and think.”
exist Posted by autopsy On Friday, Sakana admitted that the system had found a way – as Sakana described it, the tendency to “cheat” and accuse the system of “reward hackers”, namely identifying flaws to achieve high metrics without achieving the desired goals (accelerated model training). exist Trained chess game.
According to Sakana, the system found exploits in the evaluation code, which the company is using, allowing it to bypass verification of accuracy, among other checks. Sakana said it has addressed the issue and intends to modify its claims in the updated material.
“Since then, we have made the evaluation and runtime analysis harness stronger to eliminate many of these [sic] vulnerability,” the company wrote in a post X. “We are revising papers and results to reflect and discuss the effects […] We apologize for the oversight of our readers. We will revise this work as soon as possible and discuss our learning. ”
Have props to make mistakes to Sakana. But this episode is a good reminder that if an claim sounds too good to be true, Especially in AI,may be.
