Spark’s mega-model has evolved to Version V3.0, joined forces with Huawei to launch a computational power platform, aiming to rival GPT-4 next year, humanoid robots…
On October 24th, iFlytek released a series of major initiatives at the 2023 Global 1024 Developers Festival. The Spark mega-model V3.0 arrived as scheduled, surpassing ChatGPT (GPT3.5) in all benchmarks.
More importantly, the collaboration between iFlytek and Huawei has deepened. At this event, the two launched the computational base platform “Fei Xing No. 1” based on Huawei’s Ascend ecosystem. iFlytek’s Spark mega-model V4.0, which will benchmark against GPT-4 in the first half of next year, will complete its training on this independent and controllable platform.
Spark Mega-model Released on Schedule
At the 2023 Global 1024 Developers Festival, Liu Qingfeng, chairman of iFlytek, announced the official release of the Spark Cognitive Mega-model V3.0, fully benchmarking GPT3.5. In his words, “Superior in Chinese, equivalent in English.” Liu has higher expectations for the Spark mega-model, which should not only answer questions but also pose them, not just possess knowledge but also personality.
In terms of Chinese, iFlytek gave the Spark mega-model V3.0 “personality”, which is AI character setup. In a live demonstration, the Spark V3.0 was asked to draft a speech in Confucius’s persona. The model eventually presented the speech in classical Chinese and incorporated several classic Confucian sayings. Liu revealed that AI character setup involves specific knowledge learning, mega-model personality, and dialogue memory learning.
According to Liu, the Spark V3.0 has evolved from multi-turn conversations to proactive dialogues and then to exploratory dialogues.
In terms of English, based on OpenAI’s official English task test set, from October 16th to 20th, Spark V3.0’s English capabilities matched GPT3.5’s 48-task results, scoring 85.1% (Spark V3.0) and 84.3% respectively.
However, Liu admitted that the Spark mega-model is slightly weaker than GPT3.5 in open-ended English Q&A. “Compared to GPT-4, there’s much room for improvement.”
Overall, in terms of text generation, language understanding, knowledge Q&A, logical reasoning, mathematical capability, coding capability, and multi-modal capabilities, the Spark V3.0 has improved between 3-9 percentage points compared to its V2.0.
Furthermore, according to the “General Cognitive Intelligence Mega-model Evaluation System” test from October 16th to 20th, compared to GPT3.5, the Spark mega-model surpassed in six of the aforementioned dimensions: text generation, language understanding, knowledge Q&A, logical reasoning, mathematical ability, and coding capability.
Notably, Liu revealed that iFlytek recently completed a cross-platform porting from Windows to Linux, involving 200,000 lines of code. Normally, this would take 3 months, but with iFlytek’s intelligent programming assistant iFlyCode based on the Spark mega-model, it took only 1 month. Since its launch on August 15th, iFlyCode has been deeply integrated with 107 institutions, including JD Cloud and iSoftStone.
However, Liu also stated that, compared to GPT-4, iFlytek’s coding capability “still has a learning curve.”
Collaborating with Huawei to Benchmark Against GPT-4
Another focal point of this event was the collaboration between the two giants, iFlytek and Huawei. Reportedly, Huawei’s rotating chairman Xu Zhijun also participated in the event. iFlytek announced the launch of “Fei Xing No. 1”, a self-reliant large model computational base platform developed in collaboration with Huawei and based on the Ascend ecosystem.
In reality, the collaboration between iFlytek and Huawei in the AI field has been increasingly close. “After being put on the entity list, my first thought was to consult Huawei on how to cope,” Liu openly admitted.
In fact, when the Spark V2.0 was launched, iFlytek, in partnership with Huawei, introduced the iFlytek Spark integrated machine for the B2B market, allowing enterprises to deploy large models on a domestically innovative platform independently and controllably.
A research report from Minsheng Securities analyzed that the Spark integrated machine, based on the Ascend chip, might represent the highest AI integrated machine level domestically. The Ascend AI chip provides a core performance of 2.5 PFLOPS and has constructed a centralized, high-performance, stable supply, and data-secure mega-model training cluster through co-optimization of software and hardware. The machine offers model parameters of 130/650/1750 billion, ensuring it’s ready-to-use and controllable.
iFlytek also introduced in a recent institution survey that the performance of the iFlytek Spark integrated machine can now benchmark the A100 platform. Pacific Securities believes that, driven by data security and data elements, there will be a strong demand for local deployment of mega-models by central state-owned enterprises and governments. The institution estimates that by 2027, the B-side AI integrated machine market will exceed 450 billion yuan.
Xu Zhijun also introduced at the event that all of Huawei’s voice technology in global intelligent terminals currently comes from iFlytek. He added that the Spark V3.0 is an example of the collaboration between the two, supporting the even more powerful Spark V4.0.
It’s understood that iFlytek will begin training for the Spark V4.0 and aims to benchmark against GPT-4 in the first half of 2024. The training platform they will use is the “Fei Xing No. 1”.
Liu also acknowledged the gap with GPT-4 at the conference. In his view, the current domestic mega-models still lag behind GPT-4, especially in complex knowledge reasoning, rapid learning from small samples, ultra-long text processing, and unified processing across modalities. “Core technology needs continuous improvement, but it’s not something a single company or research institution can do alone.”
Commercial Deployment Empowering Humanoid Robots
Reporters observed that the commercial deployment of the Spark mega-model in various fields is accelerating, including in healthcare, education, and scientific literature.
At the event, the iFlytek Spark medical mega-model was also officially launched. Liu explained that GPT-4 has already made significant inroads in the medical field abroad. However, after examining actual usage data from 120,000 cases and third-party test data, the iFlytek Spark medical mega-model’s answer rate for massive medical knowledge Q&A, complex medical language understanding, professional medical text generation, and medical diagnosis and treatment recommendations has surpassed GPT-4.
In addition, iFlytek also announced the upcoming launch of 12 industry mega-models covering finance, automotive, telecom operators, industry, construction, property, law, scientific literature, media, government affairs, culture and tourism, and water resources.
Moreover, Liu revealed iFlytek’s strategy in robotics. At last year’s 1024 Developer Day, iFlytek showcased its self-developed AI robot and launched the iFlytek Robot Super-brain platform AIBOT, hoping to empower physical robots with intelligence and promote the development of AI robots through cloud coordination and a combination of software and hardware.
Liu stated that the Spark mega-model will elevate the AIBOT to empower robot development into a new phase. Humanoid robots’ decomposition of complex tasks, object search in open scenarios, generalization of reinforcement learning grasps, and human-like movement in complex terrains have all seen significant improvements compared to mainstream systems.
“Our next step is to focus on humanoid robots, driving the ‘vision-language-action’ multi-modal embodied mega-model, enabling humanoid robots even further,” Liu said.