news

Recently, the development of AI has presented a contradictory picture: on the surface, there is a flourishing of all kinds of things, but in reality, there is a hidden sign of fatigue. The field of artificial intelligence seems to have entered a delicate node, with many technology giants and startups launching their own AI products, but it is difficult to hide the dilemma of homogeneity.

The Character.ai incident is a case in point. As a star AI product, Character.ai’s inference request volume during peak hours has reached one-fifth of Google’s search traffic. Are five Character.ai equal to one Google? However, the real situation is not optimistic. According to The Information, Character.ai has less than 100,000 subscribers and has been working hard to cut costs recently. Due to financing difficulties, Character.ai is considering selling.

Most AI social products are limited to chatting and role-playing functions. Heavy users consume a large amount of reasoning resources, while light users have poor retention, and their commercial prospects are worrying. For example, Character.ai's annual revenue last year was only US$15.2 million.

Character.ai’s dilemma is not an isolated case. Looking at the entire AI industry, from tech giants to startups, they all seem to be stuck in a “carving” dilemma, focusing too much on piling up superficial functions, failing to make breakthroughs at the bottom level, and giving way to homogenized functional fine-tuning.

At WWDC24, Apple launched a new personal AI system, Apple Intelligence, for iPhone and other devices, trying to make up for its lagging AI layout by redefining "AI". However, apart from the usual emphasis on privacy protection, the actual function display was lackluster and the integrationChatGPTFunctions include identifying notifications that are important to the user's context, AI text processing, image and text generation, and Siri sensing screen content and performing simple tasks on behalf of the user.

Apple Intelligence is also not immune to "futures". Testing just started a few days ago, and the official launch of the first batch of features may need to be postponed to iOS 18.1 in October, but they are only basic features. Using device-side data to answer questions and understand the content on the user's screen will still take until spring 2025.

What’s even more regrettable is that Apple Intelligence only supports iPhones equipped with A17 Pro chips or newer, which means that most iPhone users cannot experience it.

Microsoft is trying to innovate on the PC, launching Copilot+ PC on May 20, which includes the artificial intelligence tool Recall, which regularly takes screenshots to create activity records so that users can search their previous operations.

In fact, before Recall, a similar app Rewind had been online for more than a year, and had received investments from Sam Altman, a16z, etc. In essence, Recall did not have any innovation, but just copied Rewind. Although Microsoft had high hopes for Recall, it received negative reviews. Many users believed that there were huge privacy risks, and some hackers even demonstrated how to extract all information from Recall's main database. For security reasons, Microsoft shelved the release of Recall.

As the locomotive, OpenAI has also begun to spare no effort in "carving". The release date of its highly anticipated AI video generation model Sora is still undecided. The official explanation is that there is still a lot of security work to be done, but no clear timetable has been given.

Also disappointing is the so-called "Her into reality"GPT-4oVoice mode, which was described by many as a small innovative product that was built by piling up existing functions, was originally scheduled for release in May, but was repeatedly delayed. It was only occasionally given to a small group of people for trial use to maintain everyone's expectations. Even the newly launched SearchGPT cannot hide its "futures" nature: not only is the internal testing mechanism a black box, but even the demo has low-level errors.

In the view of some netizens, instead of focusing on substantial technological breakthroughs,OpenAIIt seems that he is more interested in creating hot topics. The series of "hot topics" released against Google are indeed quite eye-catching. Altman makes great use of social media to influence public opinion and build momentum for his own products. However, beautiful PPT and lively press conferences cannot cover up the reality of slow iteration of model capabilities.

When GPT-5 will be launched has become a complete mystery. The market has predicted that GPT-5 may be released at the end of 2023 or the summer of 2024. OpenAI CTO Mira Murati said it is expected to be launched at the end of 2025 or early 2026, and Sam Altman has emphasized that there is currently no fixed release schedule.

The launch of the small parameter model GPT-4o mini basically explains everything. OpenAI is also going to compete with everyone in "carving".

While the leading companies are stuck in the "carving" dilemma, the entrepreneurial teams are not immune. As the investment window of "model is product" has passed, manufacturers are facing huge pressure to realize their value and are in urgent need of finding landing scenarios. AI search seems to have become another track that developers are rushing to enter after chatbots and Character.ai-like products, becoming a new hot spot.

AI search is considered an important landing direction for large model applications, and many companies and developers at home and abroad have made plans in the hope of getting a piece of the pie. However, the current development trend seems to be unable to escape the dilemma of "carving", and everyone seems to be reinventing the same wheel.

The AI ​​search products launched by many players are similar, lacking substantial differences and innovations. Although the function names are slightly different, the core functions are almost the same, which actually reflects that the development of the current large model has encountered a bottleneck.

Browser plug-ins have become the object of competition among companies due to their lightweight and low-threshold features.Dark Side of the MoonofKimiBrowser plug-ins emphasize "lightweight and small search", and Byte's Doubao browser plug-in is also rapidly iterating new functions. AI plug-ins made by various developers are also emerging in an endless stream.

On the surface, these plug-ins are diverse and have different functions. Some focus on content generation, some focus on information aggregation, and some are committed to improving productivity. But if you look closely, you will find that most of them are just "carving" work. Many plug-ins are just simple encapsulations of existing AI functions, or surface calls to large model APIs, and there are few truly groundbreaking applications.

Everyone wants to get user traffic. But this approach is more like a stopgap measure and does not fundamentally solve the commercialization problem. The number of PC browser users is relatively small compared to mobile terminals, and this reality itself limits the potential user base of plug-ins. It once again confirms the current AI industry's dilemma of "carving patterns", making minor improvements and optimizations within the existing technology framework and homogenized competition.

Looking back at the last waveComputer Vision (CV)The rise and fall of AI companies that started with technology have not really gotten out of the "carving" dilemma. Despite significant progress in technology, scene fragmentation and commercial difficulties are still the main reasons for their lack of follow-up. These companies often remain at the stage of technology suppliers, and it is difficult to ensure long-term market position. At the same time, they have to fight price wars with technology giants to compete for market share.

Today, large language models seem to be repeating this history.GPTAlthough models such as these have made breakthrough progress in technology, commercial applications that can truly create sustainable value are still scarce, and the improvement of model capabilities seems to have stagnated.

The five-level capability assessment system for general artificial intelligence proposed by OpenAI provides a development path for the industry. When AI reaches the second stage (the "reasoner" level), it may be ready for large-scale popularity in the consumer market. This means that large model companies need to continuously improve the general generalization capabilities of AI in order to truly break through the bottleneck of commercialization and productization. However, before that, OpenAI may need to raise tens of billions of dollars to cover its costs.

Unlike CV companies, some large model companies have begun to launch C-side applications in an attempt to approach product-market fit (PMF). This strategy may help them avoid the trap of relying solely on the non-standardized market on the B-side. However, if they simply package existing AI capabilities into consumer-grade products without truly breaking through the model's reasoning capabilities, enhancing cross-domain knowledge integration, and interactive experience, they will soon fall into homogenization.

To prepare for the AI ​​revolution, companies are still investing heavily in building infrastructure such as data centers, but even Wall Street is beginning to rethink its tune, shifting from enthusiastic support to more cautious support.

Does AI entering the carving stage mean a repeat of the Internet bubble? A series of recent research reports have poured cold water on the AI ​​boom, warning that generative AI technology still faces a long and questionable road to development.

When Goldman Sachs says “maybe it’s just a bubble,” you know the industry is in real trouble. In a report titled “Gen AI: too much spend, too little benefit?”, Goldman Sachs analysts discuss whether AI can solve the complex problems it’s been given and express doubts about its still-undetermined “killer app.”

A research report from Barclays Bank has an even more vivid title, "Cloud AI Capex: FOMO or Field-Of-Dreams?" The report points out whether data center investment is creating a bubble that could be like the telecommunications collapse after the Internet bubble in the 1990s. The analysts' conclusion is that it is leaning towards FOMO.

David Kahn, a partner at Sequoia Capital, recently called whether the technology can recoup huge data center investments “AI’s $600 billion question.”

This is obviously not the first time doubts have been raised. Questions about the ultimate revenue or potential of AI chatbots and other tools have become a common topic on any tech company's earnings call.

The huge gains for Microsoft and Nvidia belong to what Goldman Sachs analysts call the "selling shovels and picks" phase of AI investment: providing the semiconductors,cloud computingIt is worth noting that the growth brought by AI to leading cloud vendors such as Google, Microsoft, and Amazon has begun to slow down.

Jim Coviello, head of global equity research at Goldman Sachs, is skeptical that the cost of AI will drop enough to justify its value. “To justify these costs, the technology must be able to solve complex problems.”

Goldman Sachs also spoke to MIT economist Daron Acemoglu, who believes the technology is still far from ready for the mainstream: "Given the focus and architecture of generative AI technology today, these truly transformative changes won't happen quickly, and probably not in the next 10 years, if at all."

Barclays analysts described the thinking of AI supporters as the logic of "as long as we develop AI technology, users and profits will follow naturally", and pointed out that current capital expenditure forecasts are sufficient to support 12,000 ChatGPT-scale AI products by 2026. Given the amount of investment, analysts said the industry may be heading towards "overbuilding" similar to the telecommunications collapse after the Internet bubble, and "next year may see a period of adjustment in AI investment."

David Kahn, a partner at Sequoia Capital, also noted that AI still has a long way to go, warning investors to stay "calm" amid the "speculative frenzy." "We need to make sure not to buy into the illusion that is now spreading from Silicon Valley to the country and the world," Kahn wrote. "The illusion is that AGI will arrive tomorrow and we all need to hoard the only resource of value - GPUs. In reality, the road ahead will be long."

Artificial intelligence will change everything about business forever. This has become a mantra in business and academia, and AI has become a household term. But this has been a zigzag rise. Since the 1950s, AI has experienced several winters, when interest in AI waned due to unfulfilled expectations. Behind this ups and downs is the complex relationship between expectations and reality of the technology. Looking back at these developments can provide us with a historical perspective to understand the current wave of generative AI - there seems to be a pattern here, that is, optimism about AI technology is often followed by a period of frustration and decline in AI investment. In a favorite phrase, it is the AI ​​winter.

Just as AI is probably still stuck in the dial-up era, with the internet not realizing its true potential until dial-up faded away, we will one day have a new form of AI that makes ChatGPT as obsolete as Web Portal was in 1998. We just probably can’t count on it happening in the next year or two.