Surpassing ChatGPT has long been the obsession of domestic AI large model enterprises. Just like at the StarFire Cognitive Large Model V1.5 upgrade conference held by iFLYTEK in June this year, there were multiple mentions of statements such as ‘iFLYTEK StarFire is just a step away from ChatGPT’ and ‘It will surpass ChatGPT before October this year’.
Although iFlytek has the suspicion of ‘staged marketing’, given the current technological capabilities demonstrated by domestic AI large models, it is not impossible to surpass ChatGPT.
From a product perspective, ChatGPT is essentially just a ‘chatbot’, an artificial intelligence based on NLP logic for natural language processing. It can also be understood as ‘weak AI’. After all, without data training, it is difficult for it to perform tasks other than engaging in conversations with humans.
ChatGPT still has a long way to go in terms of possessing independent thinking and learning abilities, being able to independently complete multi-threaded tasks in an environment without human intervention. Furthermore, beneath its high-tech facade, ChatGPT’s data core still relies on human interference. For example, there have been reports that OpenAI employed ‘digital laborers’ at low wages in a ‘sweatshop’ fashion to meet the data training needs of ChatGPT through repetitive manual data annotation and information correction. Strictly speaking, its technological barriers are not as difficult to overcome as imagined.
According to Morgan Research, the success of ChatGPT lies in its ability to fulfill people’s partial imagination of AI tool attributes. The key to surpassing ChatGPT can be summarized in two directions: 1. Achieving faster and more accurate computational efficiency at the technical level. 2. Seeking breakthroughs in specialization to unleash more imagination for AI tools.
So, how much gap does domestic AI large models need to bridge in order to achieve surpassing at the technical level? The answer may be much more severe than imagined.
According to the Chinese large-scale model ranking released by SuperCLUE, a domestic research institution, in June 2023, GPT-4 ranked first with a score of 78.76 under the standard of human average ability score being determined as 93.09 points, leading the top tier. In the second tier within the range of 60 points, both 360 Brain and Baidu ERNIE Bot achieved relatively similar scores and ranked third and fourth respectively. However, iFlytek fell into the third tier with a regrettable difference of 0.2 points.
Although there are still controversies surrounding the data evaluation of SuperCLUE, further verification is needed for the scores and rankings involved in the leaderboard. However, it is enough to indicate the gap between domestic AI large models and ChatGPT.
Before October 2023, iFlytek may have a chance to catch up with 360 Brain and ERNIE Bot from Baidu, Inc. However, in order to bridge the nearly 20-point gap with GPT-4, it may only be possible to rely on the emergence of some kind of ‘black technology’.
In terms of distinctive innovation, does domestic AI large models have the opportunity to achieve a ‘breakthrough’ in overtaking on curves? Unfortunately, at this stage, most mature generative AI products are mostly based on imitating ChatGPT, and their functional blocks can be said to be similar. Although large models have not yet emerged from the period of shaping new concepts, they have already shown a certain tendency towards homogeneity.
In order to compete for the reputation of ‘China’s version of ChatGPT,’ most domestic AI large-scale model companies have chosen the safest and most robust business competition logic. They have directly copied the proven product design ideas of ChatGPT, focusing their main competitive direction on parameters, computing power, and other core technological elements. In this highly intense competition, AI large-scale models are being pushed from the first half of research and development into the second half of implementation stage prematurely, further suppressing product innovation.
However, among the AI players in China, there are also many small-scale generative AI products. Due to a lack of financial and technological strength, they choose to enter niche markets. For example, the recently popular Miaoyaxiangji precisely targets the market demand for AI-generated ID photos and portrait photos for women.
Even OpenAI has started to change course and set its sights on the niche market of large models. Recent overseas reports suggest that OpenAI plans to launch multiple low-cost, small-scale GPT-4 models for various specialized fields, targeting specific tasks and themes.
This may also mean that the logic of ‘using a cannon to shoot mosquitoes’ style AI products will soon take a back seat, replaced by a ‘surgical knife’ style lightweight design concept. Led by iFlytek, the big models chasing after ChatGPT may be getting left further and further behind.
OpenAI’s shift towards vertical integration is primarily driven by cost pressures. According to a research report previously released by OpenAI, the computational power required for developing large-scale AI models doubled approximately every four months from 2012 to early 2018. Nowadays, just training the AIGC model alone consumes nearly 300,000 times more computational power.
With the increasing usage of ChatGPT, the demand for computing power has also exponentially increased. In fact, in April this year, OpenAI had to temporarily suspend the Plus subscription service for ChatGPT due to its inability to meet the continuously growing computing power requirements.
Behind the surge in computing power is an extremely large expenditure. According to incomplete statistics, the cost of a single training session for GPT-3 is approximately 2 million to 14 million US dollars, equivalent to about 14.37 million yuan to 101 million yuan. Considering that ChatGPT’s training sessions are likely to exceed 1000 times, the total training cost is almost astronomical, and this is only the software cost.
In terms of hardware, it is understood that ChatGPT’s AI computing cluster consists of tens of thousands of NVIDIA A100 and H100 high-performance GPU chips. Rough estimates suggest that the hardware equipment costs approximately 200 million US dollars, equivalent to about 1.4 billion yuan.
Indeed, such high costs have resulted in the development of generative AI products that are considered groundbreaking throughout human history. However, from a business perspective, it is important not to overlook the limited profit potential and challenging cost maintenance of ChatGPT. If achieving a balanced budget is desired, seeking lightweight breakthroughs in niche markets has become almost inevitable.
After all, according to OpenAI’s public data, the price for each information query is 2.5 cents in US dollars, which is approximately 0.18 yuan in Chinese currency. And the computational cost for every 1000 tokens is 0.02 US dollars, equivalent to about 0.14 yuan.
According to public information, currently GPT-4 allows a maximum of 200 requests per account per minute, with a limit of 40,000 tokens for questions and answers. For GPT-3.5, the maximum number of requests per minute is 3,500, with a limit of 90,000 tokens for questions and answers. Roughly calculated, the highest potential revenue for GPT-4 is 36 yuan per minute with a maximum cost of 5,600 yuan; while for GPT-3.5 it is 630 yuan per minute with a maximum cost of 12,600 yuan.
At least before fully opening up the usage permissions, it remains difficult to improve the situation where ChatGPT is not generating enough revenue under the current billing model. Judging from a limited perspective, domestic AI large models may also remain in a state of burning money on a large scale for a long time.
iFlytek is a typical case. According to the financial reports in recent years, iFlytek’s overall performance has been steadily improving, except for the stagnant revenue growth and significant profit decline in 2022 due to the impact of the pandemic and extreme pressure from the United States. It was during December 2022 that iFlytek began preparing for AI large-scale model research and development.
However, in the first quarter of 2023, iFLYTEK’s total revenue decreased by 17.64% year-on-year, and net profit decreased by 152.26% year-on-year, resulting in a loss instead of profit. The financial report also explicitly pointed out that the additional investment in the “1+N Cognitive Intelligence Large Model Special Project” launched on December 15th, 2022 and the official release of Xinghuo Large Model on May 6th, 2023 have had an impact on current profits.
The loss in the first quarter also affected the performance of the first half of 2023. In the recently announced performance forecast, iFLYTEK’s attributable net profit for the first half of 2023 is expected to decrease by 71%-80% compared to the same period last year. Although revenue and profits are expected to increase in the second quarter, iFLYTEK’s growth potential in future quarters still faces risks due to difficulties in profitability and high investment caused by AI large models.
According to the financial indicators displayed by Tianyancha, the total asset turnover rate of iFlytek has declined significantly. It is also important to be aware of the possibility of deteriorating financial conditions.
Nowadays, it has become a trend for AI large models to shift towards vertical competitions. We hope that iFLYTEK can strengthen its focus on lightweight generative AI products and actively seek demand directions that can leverage profit growth, in order to establish a profitable business model as soon as possible.
For all domestic AI big model enterprises, how to explore breakthrough opportunities in the ‘overtime game’ of vertical markets will be a question that must be taken seriously.