SubQ, mocked as the 'AI panacea', releases version 1.1: Pulling in third-party reviews for self verification, but accused of mixing AI jargon According to Beating monitoring, the controversial big model SubQ, which claimed to reduce computing consumption by a thousand times, has released a technical report on version 1.1 Small. In response to accusations that early preview versions were mocked by the community as "AI panacea" (meaning false advertising) due to a lack of papers and independent verification, the research and development company Subquadric, in conjunction with the evaluator Appen, conducted a tripartite evaluation, claiming that the model achieved 98% retrieval accuracy under the limit length of 12 million tokens and achieved results close to mainstream cutting-edge models in practical programming testing. The technical report also revealed that the model was not trained from scratch, but was modified based on open-source cutting-edge models, replacing attention computing mechanisms and incrementally training 1 trillion tokens. Even with third-party reviews and self verification, the developer community is still full of doubts about this update. Some researchers have pointed out that the so-called black technology does not actually have a breakthrough in underlying technology, but is essentially an existing technology that cuts long text into small pieces and dynamically filters them (i.e. block sparse attention mechanism); Some readers roast that AI generated text jargon was mixed into the technical report (especially in Chapter 5.7.1). The system engineer warned that the filtering mechanism would incur additional scheduling overhead when used concurrently by multiple people, resulting in severe lag for the slowest 1% of users. Due to the lack of publicly available core parameters and API interfaces for everyone to download, the so-called promises of reduced computing power and ultra-low pricing are still on paper. [Original link]