Opportunities for iFlytek Hidden in Vertical Large Models?

Surpassing ChatGPT has lengthy been the obsession of home AI massive mannequin enterprises. Just like on the StarFire Cognitive Large Model V1.5 improve convention held by iFLYTEK in June this yr, there have been a number of mentions of statements resembling ‘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 present technological capabilities demonstrated by home AI massive fashions, it isn’t unimaginable to surpass ChatGPT.

From a product perspective, ChatGPT is actually only a ‘chatbot’, a synthetic intelligence primarily based on NLP logic for pure language processing. It may also be understood as ‘weak AI’. After all, with out knowledge coaching, it’s tough for it to carry out duties apart from partaking in conversations with people.

ChatGPT nonetheless has a protracted method to go in phrases of possessing unbiased pondering and studying skills, having the ability to independently full multi-threaded duties in an surroundings with out human intervention. Furthermore, beneath its high-tech facade, ChatGPT’s knowledge core nonetheless depends on human interference. For instance, there have been reviews that OpenAI employed ‘digital laborers’ at low wages in a ‘sweatshop’ style to satisfy the info coaching wants of ChatGPT by means of repetitive guide knowledge annotation and knowledge correction. Strictly talking, its technological limitations aren’t as tough to beat as imagined.

According to Morgan Research, the success of ChatGPT lies in its capacity to meet individuals’s partial creativeness of AI instrument attributes. The key to surpassing ChatGPT will be summarized in two instructions: 1. Achieving sooner and extra correct computational effectivity on the technical degree. 2. Seeking breakthroughs in specialization to unleash extra creativeness for AI instruments.

So, how a lot hole does home AI massive fashions must bridge in order to realize surpassing on the technical degree? The reply could also be rather more extreme than imagined.

According to the Chinese large-scale mannequin rating launched by SuperCLUE, a home analysis establishment, in June 2023, GPT-4 ranked first with a rating of 78.76 underneath the usual of human common capacity rating being decided as 93.09 factors, main the highest tier. In the second tier inside the vary of 60 factors, each 360 Brain and Baidu ERNIE Bot achieved comparatively comparable scores and ranked third and fourth respectively. However, iFlytek fell into the third tier with a regrettable distinction of 0.2 factors.

Although there are nonetheless controversies surrounding the info analysis of SuperCLUE, additional verification is required for the scores and rankings concerned in the leaderboard. However, it is sufficient to point out the hole between home AI massive fashions and ChatGPT.

Before October 2023, iFlytek could have an opportunity to meet up with 360 Brain and ERNIE Bot from Baidu, Inc. However, in order to bridge the almost 20-point hole with GPT-4, it might solely be potential to depend on the emergence of some form of ‘black technology’.

In phrases of distinctive innovation, does home AI massive fashions have the chance to realize a ‘breakthrough’ in overtaking on curves? Unfortunately, at this stage, most mature generative AI merchandise are largely primarily based on imitating ChatGPT, and their useful blocks will be stated to be comparable. Although massive fashions haven’t but emerged from the interval of shaping new ideas, they’ve already proven a sure tendency in direction of homogeneity.

In order to compete for the status of ‘China’s model of ChatGPT,’ most home AI large-scale mannequin firms have chosen the most secure and most sturdy enterprise competitors logic. They have straight copied the confirmed product design concepts of ChatGPT, focusing their most important aggressive course on parameters, computing energy, and different core technological parts. In this extremely intense competitors, AI large-scale fashions are being pushed from the primary half of analysis and improvement into the second half of implementation stage prematurely, additional suppressing product innovation.

However, among the many AI gamers in China, there are additionally many small-scale generative AI merchandise. Due to a scarcity of economic and technological power, they select to enter area of interest markets. For instance, the lately widespread Miaoyaxiangji exactly targets the market demand for AI-generated ID photographs and portrait photographs for ladies.

Even OpenAI has began to alter course and set its sights on the area of interest market of huge fashions. Recent abroad reviews counsel that OpenAI plans to launch a number of low-cost, small-scale GPT-4 fashions for numerous specialised fields, focusing on particular duties and themes.

This may imply that the logic of ‘using a cannon to shoot mosquitoes’ type AI merchandise will quickly take a again seat, changed by a ‘surgical knife’ type light-weight design idea. Led by iFlytek, the large fashions chasing after ChatGPT could also be getting left additional and additional behind.

OpenAI’s shift in direction of vertical integration is primarily pushed by value pressures. According to a analysis report beforehand launched by OpenAI, the computational energy required for creating large-scale AI fashions doubled roughly each 4 months from 2012 to early 2018. Nowadays, simply coaching the AIGC mannequin alone consumes almost 300,000 occasions extra computational energy.

SEE ALSO: iFlytek Denies Layoffs Rumor, Plans to Recruit Over 1,000 Employees in 2022

With the growing utilization of ChatGPT, the demand for computing energy has additionally exponentially elevated. In truth, in April this yr, OpenAI needed to quickly droop the Plus subscription service for ChatGPT as a consequence of its lack of ability to satisfy the repeatedly rising computing energy necessities.

Behind the surge in computing energy is a particularly massive expenditure. According to incomplete statistics, the price of a single coaching session for GPT-3 is roughly 2 million to 14 million US {dollars}, equal to about 14.37 million yuan to 101 million yuan. Considering that ChatGPT’s coaching periods are more likely to exceed 1000 occasions, the full coaching value is sort of astronomical, and that is solely the software program value.

In phrases of {hardware}, it’s understood that ChatGPT’s AI computing cluster consists of tens of hundreds of NVIDIA A100 and H100 high-performance GPU chips. Rough estimates counsel that the {hardware} tools prices roughly 200 million US {dollars}, equal to about 1.4 billion yuan.

Indeed, such excessive prices have resulted in the event of generative AI merchandise which can be thought-about groundbreaking all through human historical past. However, from a enterprise perspective, it will be important to not overlook the restricted revenue potential and difficult value upkeep of ChatGPT. If attaining a balanced finances is desired, searching for light-weight breakthroughs in area of interest markets has turn into virtually inevitable.

After all, in keeping with OpenAI’s public knowledge, the value for every data question is 2.5 cents in US {dollars}, which is roughly 0.18 yuan in Chinese foreign money. And the computational value for each 1000 tokens is 0.02 US {dollars}, equal to about 0.14 yuan.

According to public data, at the moment GPT-4 permits a most of 200 requests per account per minute, with a restrict of 40,000 tokens for questions and solutions. For GPT-3.5, the utmost variety of requests per minute is 3,500, with a restrict of 90,000 tokens for questions and solutions. Roughly calculated, the very best potential income for GPT-4 is 36 yuan per minute with a most value of 5,600 yuan; whereas for GPT-3.5 it’s 630 yuan per minute with a most value of 12,600 yuan.

At least earlier than absolutely opening up the utilization permissions, it stays tough to enhance the scenario the place ChatGPT isn’t producing sufficient income underneath the present billing mannequin. Judging from a restricted perspective, home AI massive fashions may stay in a state of burning cash on a big scale for a very long time.

SEE ALSO: iFlytek Claims Its Large Language Model Outperforms ChatGPT in Three Key Areas

iFlytek is a typical case. According to the monetary reviews in current years, iFlytek’s general efficiency has been steadily enhancing, besides for the stagnant income progress and important revenue decline in 2022 because of the impression of the pandemic and excessive stress from the United States. It was throughout December 2022 that iFlytek started making ready for AI large-scale mannequin analysis and improvement.

However, in the primary quarter of 2023, iFLYTEK’s whole income decreased by 17.64% year-on-year, and web revenue decreased by 152.26% year-on-year, ensuing in a loss as an alternative of revenue. The monetary report additionally explicitly identified that the extra funding in the “1+N Cognitive Intelligence Large Model Special Project” launched on December fifteenth, 2022 and the official launch of Xinghuo Large Model on May sixth, 2023 have had an impression on present earnings.

The loss in the primary quarter additionally affected the efficiency of the primary half of 2023. In the lately introduced efficiency forecast, iFLYTEK’s attributable web revenue for the primary half of 2023 is predicted to lower by 71%-80% in comparison with the identical interval final yr. Although income and earnings are anticipated to extend in the second quarter, iFLYTEK’s progress potential in future quarters nonetheless faces dangers as a consequence of difficulties in profitability and excessive funding attributable to AI massive fashions.

According to the monetary indicators displayed by Tianyancha, the full asset turnover fee of iFlytek has declined considerably. It can be necessary to concentrate on the potential of deteriorating monetary circumstances.

Nowadays, it has turn into a pattern for AI massive fashions to shift in direction of vertical competitions. We hope that iFLYTEK can strengthen its give attention to light-weight generative AI merchandise and actively search demand instructions that may leverage revenue progress, in order to determine a worthwhile enterprise mannequin as quickly as potential.

For all home AI massive mannequin enterprises, the right way to discover breakthrough alternatives in the ‘overtime game’ of vertical markets will likely be a query that should be taken significantly.