tencent technology news according to foreign media reports, google has made huge investments in large language models, but gemini still faces the dual challenges of complex operations and developers tending to choose competitor products, which may affect the growth of its cloud services and artificial intelligence business. gemini has made some progress since its release, but there is still a significant gap in reliability and market share compared with competitors such as openai.

some developers and google employees have complained that gemini, google's conversational artificial intelligence, is too complicated to use compared to rival technologies. among the developers complaining about gemini is aidan mclaughlin, founder of a startup called topology, which makes software to improve the reasoning capabilities of existing artificial intelligence models.

he said that the first time he tried to use openai's application programming interface, it took only 30 seconds, while using gemini took four hours because google required him to set up a cloud account and do a lot of configuration. in addition, bugs in google's system sometimes forced him to backtrack. mclaughlin said: "if the model provided by google is clearly superior to the models provided by openai and anthropic in terms of capabilities, i am willing to overcome all these difficulties. the most important thing in the world is the quality of the model, but except for some special cases, google's model seems to be ranked only third."

the difficulties developers encounter when using gemini could have an impact on google's cloud business, as google hopes to use gemini to attract more customers to its server rental business. in addition, some google employees said it could also cause potential problems for google's own artificial intelligence development. by attracting millions of paying users to purchase its artificial intelligence models or use chatgpt, openai has obtained a large amount of implicit and explicit feedback on the performance of artificial intelligence, thereby further improving the model. if google's gemini does not receive the same level of use, the company may use a more vague roadmap to make the same improvements.

gemini is also important for the development of google's consumer and advertising products, which need to provide conversational answers to search and voice assistants. "having developers use gemini to solve complex real-world problems is critical feedback for the next generation of gemini," said bindu reddy, ceo of abacus ai, whose software helps developers build artificial intelligence applications. "the more feedback you have, the better," she said.

although gemini has improved a lot this year, it is still not as reliable as competing models in supporting ai business applications. a google spokesperson said that developer feedback is very important for improving the gemini model, but it is not the only source of data. some ai researchers also said that it is possible to achieve significant improvements by teaching ai to learn from human experts in various fields.

small developers and large companies adopt large language models like openai’s gpt-4 to build applications that automate everything from customer service and data analysis to fraud detection and generating computer code. openai and microsoft each earn about $100 million per month by providing large language models as a service to developers. these developers include quantitative trading firms like jane street, banks like morgan stanley, and fintech companies like stripe.

google also listed several gemini customers, including mercedes-benz, goldman sachs and uber. in statements, the companies said they use ai to power everything from customer service to coding assistance, though the extent of their purchases is unclear (google has previously set a goal of generating $1 billion in revenue from ai cloud services in north america this year, but it sells more ai services than just big language models).

sundar pichai, ceo of google parent alphabet, said on the company’s second-quarter earnings call in july that more than 2 million developers have used gemini since the beginning of the year, but it’s unclear whether that includes developers who have only tried the model once or twice.

compared to openai's model, gemini's unpopularity among developers seems to be an open secret in the industry. nevertheless, google's gemini continues to improve, and the company is also actively taking measures to improve its performance and reliability in order to remain competitive in the field of artificial intelligence.

competitive advantage

for example, a june survey of more than 750 tech workers by enterprise software startup retool found that only 2.6% of respondents said they most frequently used gemini to build ai applications, while 76% said they used openai. gemini narrowly outpaced anthropic's claude, which received 2.3%. but it's worth noting that claude's usage has more than quadrupled since retool's november 2023 report. gemini wasn't available to developers until december of last year.

similarweb, which tracks website traffic, says openai’s app developer page received 82.8 million visits from june to august, while google’s page had just 8.4 million views during the same period. smaller surveys offer similar evidence. late last month, julian saks, founder of finetune, asked 50 ai startup developers at his san francisco co-working space what conversational ai models they used most often. almost all of them said they primarily used models from anthropic or openai; no one primarily used gemini. saks wasn’t surprised by the results. “i don’t think gemini stands out compared to other models in terms of coding,” he said.

it took google nearly a year and a major reorganization to come up with a model that could compete with gpt-4, which openai released in early 2023. google began giving developers access to the gemini models in december 2023, and their performance on popular comparative evaluations was similar to openai's models.

but some google employees say openai’s early lead has earned it brand recognition among developers, and this year, another young ai startup, anthropic, has enjoyed a surge in word-of-mouth marketing as developers have used claude as an alternative to openai, particularly for coding assistance.

developers' troubles

gemini doesn’t seem to have attracted much attention. developers surveyed said gemini models are useful for analyzing long documents or code bases because they can process more information at once than other models. but they also pointed out that google’s confusing options and multiple steps compared to its competitors, as well as its developer system being different from openai, make it more difficult to use these models.

sometimes, the different options google offers for using gemini even compete with each other in google’s own search results. for a time this spring, vertex ai studio — a service google promotes to large businesses that might want to use gemini — appeared as a sponsored result for google ai studio, a simpler tool that lets developers use gemini.

“people get confused when they’re trying to figure out the options,” said alex atallah, the former chief technology officer at opensea, who encountered similar difficulties when trying to use gemini for his startup, openrouter, which makes software for developers that use big language models.

although google launched its ai studio service in december to make it easier for developers to use gemini, they must use a different query-sending method than gpt and claude, two models that share a similar format. mclaughlin said the difference makes it less attractive for developers to switch to gemini.

gemini is often criticized on social media - especially on the social media platform x. brendan dolan-gavitt, an ai researcher at security startup xbow, posted a post on x earlier this month about the many steps he needed to go through to start using gemini through vertex, another ai application development application from google, which attracted widespread attention. other developers have also joined the ranks of complaining and expressed their dissatisfaction with google gemini.

stabilize market share

google is trying to change that image, including responding to gemini's critics at x. it's also holding events for developers to promote gemini. google is also trying other incentives, such as holding a developer competition to build the best gemini-powered app. the first winner will receive a customized delorean-branded electric car with the license plate "g3m1n1." google plans to announce the final winners by the end of this year.

google is also considering merging the features of some overlapping application building products to reduce confusion among developers, according to logan kilpatrick, who works in product development at google. "openai has provided developers with something useful for a longer time, slightly longer than google, in the context of the large language model api," he said. he joined google in april as product leader for ai studio and was previously in charge of developer relations at openai. "we have to fight against their inherently strong market share among developers right now."

this week, openai introduced o1-preview, an inference model that outperforms large language models on some technical or complex tasks. this may put pressure on google to release similar technology, or risk losing more ground in the developer community.

“white glove” service

google is actively developing version 2.0 of gemini in an effort to win the long-term battle in the field of artificial intelligence and catch up with or even surpass openai's technology. nevertheless, gemini has also won the favor of some supporters, especially because it surpasses other models in the amount of information it can handle. jaclyn siu, founder of starcycle, said her company is using gemini 1.5 pro, the most advanced large language model that google provides to developers, mainly for extracting data from documents. thomson reuters is also using gemini to assist with legal tools that can compare long legal documents and identify inconsistencies.

for larger companies, the difficulty of using google's artificial intelligence is not a major obstacle. wayfair's chief technology officer fiona tan revealed that the company has enjoyed "white glove" services from google employees, who are helping the company use gemini to convert traditional databases into a more user-friendly format. the so-called "white glove" service usually refers to a very meticulous and thoughtful service that requires great attention to detail and care.

however, some developers are not planning to try gemini because their existing solutions are already effective enough. martyn strydom, who is developing an app called dots, noted that his company prefers to use claude as a coding assistant and openai's older gpt model to respond to customers because they don't have the time to constantly test different language models.

he said: "at present, almost everyone is using openai, claude or cursor. if the world's top engineers are using these tools, why should we try others?" cursor is an artificial intelligence code assistant driven by openai and anthropic models. (compilation/wu ji)