Introducing the latest innovation, Spark v3.5, proudly touted to outperform OpenAI’s GPT-4 in various language tasks and beyond. According to its creators, Spark v3.5 takes language capabilities to new heights, excelling not only in linguistic workloads but also in the art of crafting human-like speech.
The creators assert that this cutting-edge AI model is adept at synthesizing speech with a remarkable range of emotions, tones, and speech patterns, showcasing its versatility in communication.
Chinese scientists claim that they have developed an artificial intelligence (AI) system capable of surpassing one of the world’s most extensively used large language models (LLMs).
As reported by the Chinese-government-affiliated media outlet Shine, iFlytek’s Spark v3.5 excels beyond OpenAI’s GPT-4 Turbo in language capabilities, math, and coding. It also closely competes with the American AI system in various other domains.
According to Liu Qingfeng, iFlytek’s chairman, Spark v3.5 surpasses GPT-4 Turbo slightly in multimodal tasks. This indicates its enhanced proficiency in comprehending one type of input and delivering a different form of output, like interpreting a text prompt and generating an image.
Chairman Liu Qingfeng of iFlytek stated on Jan. 29 at a company conference that Spark v3.5 surpasses GPT-4 Turbo slightly in multimodal tasks.
This indicates its enhanced proficiency in comprehending one type of input and delivering a different form of output, such as interpreting a text prompt and generating an image.
GPT-4 Turbo is an enhanced iteration of GPT-4, the engine behind ChatGPT. Launched in November 2023, it is widely acknowledged as one of the most potent AI tools available.
There is no universally accepted method for directly comparing Large Language Models (LLMs) with one another, and a publicly accessible database for comparing various proprietary AI systems is currently unavailable.
Instead, companies rely on numerous benchmarks to assess performance across different domains. These benchmarks serve as tools for AI companies to evaluate their models against industry-leading counterparts.
As an illustration, in December 2023, Google unveiled that its latest Gemini Large Language Model (LLM) surpassed the standard version of GPT-4 and other prominent models in 30 out of 32 academic benchmarks employed in AI research and development. These benchmarks encompassed high school exams and assessments on morality.
According to reports from the state-owned China Global Television Network (CGTN), Spark v3.5 demonstrated the ability to synthesize speech conveying diverse emotions, tones, and speech patterns. Additionally, CGTN highlighted that its voice recognition capabilities surpassed OpenAI’s Whisper in 37 languages, including English, Chinese, French, and Russian.
iFlytek has incorporated Spark into various devices, including smart devices, school blackboards, and tablets, as reported by Shine. Additionally, the company introduced a voice-to-text mobile app in collaboration with China Mobile on January 29.
This app utilizes Spark v3.5 to transcribe phone calls and emphasize key information conveyed during the conversation.
The AI tool underwent a 90-day training period on the “Feixing No. 1” computing platform. Due to U.S. government restrictions on AI-related exports to Chinese companies, the AI company faced limitations in utilizing state-of-the-art components, particularly graphics processing units (GPUs) manufactured by Nvidia. This includes the A100, employed in training ChatGPT, along with the H100 and H200 chips.