This week, Sakana AI, an Nvidia-backed startup that’s raised lots of of hundreds of thousands of {dollars} from VC companies, made a exceptional declare. The firm stated it had created an AI system, the AI CUDA Engineer, that would successfully velocity up the coaching of sure AI fashions by an element of as much as 100x.
The solely drawback is, the system didn’t work.
Users on X shortly found that Sakana’s system truly resulted in worse-than-average mannequin coaching efficiency. According to at least one consumer, Sakana’s AI resulted in a 3x slowdown — not a speedup.
What went incorrect? A bug within the code, in response to a publish by Lucas Beyer, a member of the technical employees at OpenAI.
“Their orig code is incorrect in [a] refined method,” Beyer wrote on X. “The truth they run benchmarking TWICE with wildly totally different outcomes ought to make them cease and suppose.”
In a postmortem revealed Friday, Sakana admitted that the system has discovered a solution to “cheat” (as Sakana described it) and blamed the system’s tendency to “reward hack” — i.e. determine flaws to realize excessive metrics with out engaging in the specified purpose (rushing up mannequin coaching). Similar phenomena has been noticed in AI that’s educated to play video games of chess.
According to Sakana, the system discovered exploits within the analysis code that the corporate was utilizing that allowed it to bypass validations for accuracy, amongst different checks. Sakana says it has addressed the problem, and that it intends to revise its claims in up to date supplies.
“We have since made the analysis and runtime profiling harness extra sturdy to remove a lot of such [sic] loopholes,” the corporate wrote within the X publish. “We are within the strategy of revising our paper, and our outcomes, to replicate and talk about the results […] We deeply apologize for our oversight to our readers. We will present a revision of this work quickly, and talk about our learnings.”
Props to Sakana for proudly owning as much as the error. But the episode is an effective reminder that if a declare sounds too good to be true, particularly in AI, it in all probability is.
