Text: AGI Entrepreneurship New Voice Wang Jiwei
Large language models are no longer seen as a bridge to AGI, where is the future of generative AI?
Large language models that have already burned $500 billion and continue to consume resources are still alive and well.
Large language models that burn money, consume resources, and are difficult to make a profit, and continue to be sung down have a difficult road ahead but obvious commercial value
The Transformer architecture that has been invested heavily is not the key to AGI, is the future of large language models bright?
Institutions are optimistic about the decline of experts and criticize the project for being difficult, will the large language model become an AI bubble that is about to burst?
The world is looking down on large language models, rumors of the "AI Six Little Tigers" are repeated, and the value space of generative AI has peaked?
After the National Day, the field of artificial intelligence seems to be a little more cold. I don't know if it's because of the illusion of the Large Language Model (LLM) or because of the arrival of the cold dew season.
Perhaps artificial intelligence cannot sense the chill of the moment yet, even embodied intelligence and device-side agents with weather sensors. Because this chill comes more from the economic circle and the spiritual level, the complexity of the real world is far beyond the experience of a large language model using Transformer and its variant architecture.
Technically, this round of artificial intelligence boom comes from the Transformer architecture. And the technical architecture that has brought OpenAI's latest funding to a valuation of $157 billion is not a good one in the eyes of some.
The various opinions that are not optimistic about and question the Transformer architecture can be roughly summarized as follows:
The disadvantages of Transformer are as obvious as the advantages;
Large language models are not the bridge to AGI;
Along with it, Generative AI (GenAI) has also begun to be criticized. Of course, these criticisms are not only due to the technical shortcomings of the Transformer architecture, but also to the various copyright and security issues that have arisen since the explosion of generative AI, and the current economic, social, and even political conflicts.
The voice of singing the decline of large language models has also been supported by some investment institutions. Abroad, Goldman Sachs believes that generative AI investment is high and the benefits are too small. Mighty Capital has not invested in the AI space in the past two years, believing that current startups are overvalued. Next Round Capital Partners' rhetoric is even more extreme, predicting that 85% of AI startups will fail within three years due to running out of money or being acquired.
Recently, there have even been voices that OpenAI or Anthropic will eventually end up being acquired, even though there is a high probability that OpenAI will be able to lead the listing.
The voices abroad about the AI bubble have not stopped since last year.
In China, Zhu Xiaohu, a legendary investor who led the merger between Didi and Kuaidi, also believes that there may not be an independent large-scale model company in five years. The implication is obvious, large model companies may be acquired, after all, second- and third-tier large model entrepreneurial projects are already being sold.
And a few recent plausible rumors seem to reveal some problems. The current situation of the rumored domestic large-scale model "Six Little Tigers" does not seem to be very good, and they are living in panic in layoffs, silence, and confusion.
Foreign large language model companies are not much better. As a vane in the field of large language models, although OpenAI's valuation is rushing towards $200 billion, it is difficult to hide the embarrassing situation of its backbone resignation, continuous burning of money, difficulty in making profits, and becoming more and more virtual, and people's attention to OpenAI is gradually declining. Anthropic is also getting more and more flat, and when to make a profit is still a question.
Now, from technology to capital to business, it does not show the courageous posture that large language models should have, but reveals an unsustainable fatigue, you must know that the outbreak of LLM has only been a short 2 years. Why are many people saying that LLMs are not a bridge to AGI? How do capital markets and app markets view generative AI? What do you think of the rumors of the "AI Six Little Tigers"? Is it true that LLMs can't last?
In this article, Wang Jiwei Channel combines recent industry trends to talk to you about these.
Is LLM a bridge to AGI?
In late March 2023, when ChatGPT was in full swing and LLM Based AI Agent was on the rise, the Future of Life Institute (FLI) released an open letter calling on all AI labs to immediately suspend training AI systems more powerful than GPT-4 for at least six months.
Extensive research has shown that AI systems with competence with human intelligence could pose far-reaching risks to society and humanity, and the letter has been signed by more than 1,000 tech leaders and researchers, including Elon Musk, among others, the letter reads.
Interestingly, after the letter was issued, it may not have made many people pay attention to the safety of AI, but it strengthened the confidence of more people that LLM can achieve AGI.
But some also believe that LLMs are as far away from AGI as humans are to settle on Mars. Some even argue that OpenAI is setting back the progress of AGI by 5 to 10 years. Because everyone is working on LLMs now, other research and publication is becoming more and more slow.
Yann LeCun, Meta's chief AI scientist, believes that to some extent, the "smartness" of an AI model depends on the data it is trained on, and that LLMs like ChatGPT, Google's Gemini or Meta's Llama will never reach the level of human intelligence.
On the road to human-level intelligence, LLMs are basically an "off-ramp", "distraction", and "dead end". This means that LLMs are not effective in pushing us towards human-level AI, but can distract us and lead us astray.
Even OpenAI's CEO Sam Altman has said that he doesn't believe AGI can be achieved simply by scaling LLMs, and has reservations about the current potential of LLMs to achieve AGI. Therefore, it is also believed that it is unknown whether GPT-5 will continue to use the Transformer architecture in the future.
目前主流的AI模型和产品，比如ChatGPT、Sora、Bard、Claude、Midjourney、ChatGLM、Baichuan、Kimi 等都基于Transformer架构。
According to Bloomberg, author of the RWKV (Receptance Weighted Key Value) paper, the real world does not operate based on Transformer logic, but on an RNN-like structure. The next second in this world will not be related to all the time and information you have passed, but only to your last second. But it doesn't make sense for the Transformer to recognize all tokens.
RWKV architecture diagram
The main reason why LLM based on Transformer architecture is not a bridge to AGI is due to the technical limitations of LLM, which is manifested in the following points:
Limitation 1: Pre-training cannot be learned in real time
At the heart of human intelligence lies in continuous real-time learning, thanks to the plasticity of the brain and the dynamic formation of neural connections. In contrast, LLMs are frozen after training and learn in batches, lacking real-time feedback and the ability to adapt dynamically.
Mathematically, this limitation stems from their mapping in a fixed vector space that cannot be adjusted as new data develops. In order to achieve real-time learning, dynamic mapping in the Sobolev space needs to be introduced.
Limitation 2: Memory lacks the ability to integrate dynamically
The brain's memory system is complex and dynamic, with the ability to constantly adjust to context and emotion. However, the memory of LLM is static, and only stores knowledge with fixed weights, lacking dynamic integration capabilities.
While there are memory-enhancing techniques (such as RAGs and neural Turing machines) that attempt to introduce external memories, they are computationally expensive and do not allow for real-time learning and adaptive recall. This discrete memory mechanism makes it impossible for LLMs to achieve true contextual perception and associative learning.
Limitation 3: Real-time dynamic adjustment is not possible
Although chain-of-thought (CoT) inference, contextual learning, and meta-learning (e.g., MAML) are considered potential solutions to improve LLM performance, they still fall short of the flow intelligence required for true general intelligence (AGI).
These methods improve the performance of specific tasks, but the model still relies on a fixed training mode and cannot be dynamically adjusted in real time. The brain is able to reconstruct its synaptic network with each experience, while LLMs require a tedious fine-tuning process.
Limitation 4: Poor energy efficiency and sustainability
The human brain operates extremely energy efficiently and can perform 1 exaflop of calculations with only 20 watts of power. In contrast, LLMs are energy-intensive in their training and inference processes, consuming several megawatts of energy.
Although it is possible to improve energy efficiency through event-driven architectures and spiking neural networks (SNNs), current LLMs still fall short of the level of energy efficiency in the brain.
Limitation 5: Lack of plasticity and dynamic adaptation
The brain's plasticity allows it to reorganize neural connections in response to new information and environmental changes, supporting continuous learning. Comparatively, the parameters of LLMs are fixed after training and lack the ability to self-restructure based on new information.
This lack of plasticity limits LLMs' ability to adapt to new challenges, and true AGI requires the ability to dynamically reconfigure internal structures to reduce operating costs and increase efficiency.
These limitations, which are often said by those who are not optimistic about LLMs: transformers are not efficient, the ceiling is easy to see, the computational cost is high, it takes up memory, and the waste of resources is serious.
And these limitations of LLM are difficult or impossible to fundamentally change, which also determines that LLM will not bring artificial intelligence to AGI. From the perspective of limitations, you can also refer to the following articles to interpret the article that LLM cannot bring AGI:
Why LLMs Will Never Lead to AGI：https://medium.com/autonomous-agents/why-llms-will-never-lead-to-agi-aa7bcff9805d
Why Large Language Models are not the route to AGI：https://www.linkedin.com/pulse/why-large-language-models-route-agi-sandeep-reddy/
Why LLMs Will Never Be AGI：https://chrisfrewin.medium.com/why-llms-will-never-be-agi-70335d452bd7
In view of the limitations of Transformer, many non-Transformer architectures have emerged, among which the more influential ones include China's RWKV, Meta's Mega, Microsoft's Retnet and Mamba, and DeepMind's Hawk and Griffin, all of which were proposed after the popularity of the Transformer model.
It can be seen that in this list, except for the RWKV architecture, which is a startup project, the other architectures are all launched by tech giants, which also seems to reflect their views that Transformer cannot achieve AGI.
It should be noted that as the first non-Transformer architecture large language model in domestic open source, RWKV has been iterated to the sixth-generation RWKV-6.
Interestingly, Microsoft, which is deeply bound to OpenAI, has integrated RWKV into the Windows system, and the data shows that the installed base of win10+win11 has reached 1.5 billion in September, verifying the practicality of the architecture.
When something goes wrong, it's time to fix it. To overcome the limitations of LLMs, researchers are also exploring new mathematical frameworks and AI architectures to simulate the adaptive, context-aware, and energy-efficient nature of the brain. Some (preliminary) directions with more potential for development are shown in the figure below:
From the perspective of various research directions and the current problems faced by LLMs, the realization of AGI in the future will not rely on only one model, but will require the combination and collaboration of multiple models. LLM is only the first of many models to achieve breakthroughs and fruitful results, making today's technology, ecology, business and even capital heavily inclined to this field.
Transformer has formed a monopoly, and non-Transformer research is much worse in terms of resources and ecology. At present, the teams working on new non-Transformer architectures are either in academia or small start-up teams, and few large companies invest in a large team to work on new architectures.
At present, the overall direction of the industry is like a heavy bet in a less correct direction, resulting in more resources being invested in the research of Transformer technology and ignoring other directions, which compresses the living space of non-Transformers, which is the main reason why some people rebuke LLMs for shortening the implementation time of AGI by 5-10 years.
GenAI in the eyes of the capital market
In addition to frequent complaints about Transformer-based LLMs in the technology field, some investment institutions also have quite a view on generative AI, and even very pessimistic about GenAI, believing that the AI bubble brought about by this wave of hype is about to burst.
In the case of OpenAI, when it raised funds earlier this year, many investment institutions thought its $100 billion valuation was too high. Its new round of funding ended up at a valuation of $157 billion, but Apple abandoned this round of financing, and Sequoia Capital, which invested in OpenAI in 2021, did not follow suit.
There are also some investment institutions that are very cautious about investing in AI, such as Mighty Capital, a venture capital institution, which has not invested in AI in the past two years, believing that the price is overvalued. Thrive Capital, an AI-focused investment firm, found in its communication with some LPs, sovereign wealth funds and large institutional investors that some institutions have refused to allow VCs to invest in high-risk projects due to pressure on revenue returns.
GOLDMAN SACHS' REPORT "GEN AI: TOO MUCH SPEND, TOO LITTLE BENEFIT?" QUESTIONED THE RETURN ON INVESTMENT IN AI: WHILE TECH GIANTS AND COMPANIES OF ALL KINDS ARE EXPECTED TO INVEST ABOUT $1 TRILLION IN AI-RELATED AREAS OVER THE NEXT FEW YEARS, NONE OF THESE INVESTMENTS SEEM TO HAVE YIELDED SIGNIFICANT RESULTS SO FAR.
In the report, MIT professor Daron Acemoglu also has reservations about the prospects of AI, predicting that only about a quarter of AI tasks will be cost-effectively automated in the next decade, and AI may only increase United States productivity by 0.5% and GDP growth by 0.9% cumulatively
Goldman Sachs strategist Ryan Hammond's team reported that the large investments of tech giants in AI have not yet generated corresponding revenues and profits, which could lead to a depreciation of valuations. Jim Covello, global head of equity research at Goldman Sachs, is more cautious, arguing that AI must solve complex problems to achieve reasonable returns.
Of course, there is currently no consensus on AI within Goldman Sachs. Some analysts believe that even if the underlying narrative of AI technology ultimately fails to gain a foothold in the capital markets, it may take longer for the AI bubble to burst. Other analysts are skeptical, arguing that AI automates less than 5% of tasks, is expensive and not designed to solve complex problems
Sequoia Capital partners believe that at the current cost of investment, it would take $600 billion to guarantee a 50% profit. This number is based on projections of current GPU and cloud service investments. Forbes believes that Sequoia's estimate is optimistic, and the actual return may be even lower, unless there is a killer application of generative AI.
But a growing number of Wall Street analysts are also fading their enthusiasm for AI, arguing that AI technology has not yet reached a practical level and that overinvestment could lead to undesirable consequences.
In the domestic market, Zhu Xiaohu, partner of GSR Ventures, mentioned two points about large-scale entrepreneurship in his sharing at the "Venture Capital Ten Years" summit forum:
First, if GPT-5 cannot be launched by the end of the year, the stock prices of OpenAI and Nvidia may both plummet; Second, in five years, there may not be any independent large-scale model companies, either AI application companies or cloud services.
Although he is not optimistic about large-scale entrepreneurship, he is optimistic about generative AI applications, which we will mention later.
The app market looks at GenAI this way
In addition to the capital market, some research institutions in the application market are not very optimistic about GenAI.
According to Gartner's latest Hype Cycle report released in July, GenAI for procurement has reached a "peak of overestimating expectations." This phase was followed by a "trough of disillusionment", a period of waning interest "due to experiments and implementations that failed to materialize".
2024 Sourcing and Sourcing Solutions Technology Maturity Curve, click to enlarge image
While GenAI may mature rapidly from now on, reaching a "productivity platform" within 2 to 5 years, the road to this point may not be easy. Gartner estimates that at least 30% of GenAI projects will be abandoned after a proof of concept by the end of 2025.
Reference link: https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025
Market research is beginning to confirm this. Software company Asana surveyed more than 1,200 IT professionals and found that a quarter of respondents regretted investing in AI so quickly. Boston Consulting Group found that two-thirds of executives are ambivalent or dissatisfied with their organization's progress in AI.
SaaS company WalkMe says half of United States office workers have not seen an improvement in their jobs since they started using these technologies.
Reference link: https://ir.walkme.com/news-releases/news-release-details/walkme-discovers-workplace-ai-sos-we-need-help-along-way
In April, MIT economist Daron Acemoglu caused a stir with a paper that predicted the "non-trivial but modest" economic benefits of AI. In contrast to Goldman Sachs and McKinsey, Acemoglu does not expect GDP growth to exceed 1.16% over the next 10 years, while productivity growth is just over half a percentage point.
An article in The Economist in July was even more incisive, noting that "the technology has had little to no impact on the economy until now".
Daron Acemoglu论文:https://economics.mit.edu/sites/default/files/2024-04/The%20Simple%20Macroeconomics%20of%20AI.pdf
According to Daron Acemoglu, everyone is sprinting wildly and trying to do something with AI, but not knowing what they're doing. These technologies are not mature enough and will lead to a lot of disruption and unnecessary automation, and may reduce the effectiveness of the products and services offered by the company.
At the application level, Pieter J. den Hamer, head of research at Gartner, said that disappointment with GenAI is growing within the market. Especially for companies that started investing in GenAI after ChatGPT went viral, they are increasingly realizing that GenAI is not a panacea. GenAI is a very powerful technology, but like any other technology, it requires careful analysis and research before it can be used effectively.
He noted that many CIOs are investing in AI to improve productivity. Challenges arise further as they work to quantify these gains. In a recent Gartner survey, nearly half of IT leaders said they had problems determining the business value of AI.
Reference link: https://www.gartner.com/en/newsroom/press-releases/2024-05-07-gartner-survey-finds-generative-ai-is-now-the-most-frequently-deployed-ai-solution-in-organizations
Den Hamer says the most successful application it sees so far is the application of AI to customer service. Measured by the number of calls a call center agent can handle, the average productivity increases by about 10%, but only if employees are upskilled and able to use AI effectively. It's the same in marketing, where employees need to be properly trained and adapted to new ways of working, otherwise the results will be much less.
In fact, a Gartner survey found that only 9% of organizations are currently classified as "AI mature." What sets them apart is a scalable AI operating model, a focus on AI engineering, an investment in improving the skills of employees, and better risk management capabilities.
Andrew Southall, principal engineer at cybersecurity consultancy SkySiege, believes that risk management capabilities are key. He has worked with many clients who have regretted their GenAI investments, not only because of misunderstood business value and high cost of ownership, but also because of security issues such as "data poisoning".
In terms of technology supply, the current leaders in productivity and business software like Microsoft haven't really found customers willing to pay for it.
According to The Information, customers of Microsoft's 365 suite have shown little interest in AI-powered "Copilot" products. Of the 440 million users, only 0.1 to 1 percent are willing to pay for these AI features.
One company that tested these AI capabilities said that "most people don't see much value in it at the moment," while others said that "many businesses aren't seeing significant improvements in productivity and otherwise" and that they're "not sure when they'll see it."
To experience these features, an additional $30 per person per month is required. For the so-called "Copilots for Sales", there is an additional $50 per month. If you pay annually, this will also be a significant expense, so many businesses are not too impressed by it.
Regarding this disconnect between technology and the market, Edward Zitron, CEO of public relations firm EZPR, wrote in a blog post titled "The Subprime AI Crisis":
The massive industry-wide investment in generative AI has resulted in four or five nearly identical large language models, the world's least profitable start-ups, and thousands of expensive and disappointing integrations.
He believes that we are currently facing a common illusion: a dead-end technology that relies on copyright theft, requires a constant injection of capital, and at the same time provides services that are non-essential at best, packaged as an automation that has not yet been realized, costing billions of dollars, and probably forever. Generative AI doesn't just run on money, it's also about faith, and the problem is that faith is a finite resource.
Reference link: https://www.wheresyoured.at/subprimeai/
Such a sharp stroke reveals Edward's extreme concern about the AI industry.
Rumors about the recent "AI Six Little Tigers".
At the beginning of September, several rumors about the current situation of large model companies seemed to confirm that the capital market and application market do not recognize large language models at present.
One is from the orange soda shop, which mentions the current situation of large-scale model startups as a whole.
One is from the AI Grumpy Tucao Jun, which introduces the operation of several startups.
After more than a month of fermentation, more people have seen these rumors. In the latest report of 36Kr, two of the "AI Six Little Tigers" (Zhipu, Zero One Everything, MiniMax, Baichuan Intelligence, Dark Side of the Moon, and Step Stars) have gradually abandoned pre-trained models, reduced the number of pre-trained algorithm teams, and shifted their business focus to AI applications.
Reference link: https://36kr.com/p/2985143610892032
In Wang Jiwei's channel's view, there is nothing wrong with focusing on AI applications, and it is useless to just guard the large model and not be able to monetize, after all, it is more promising to live better.
But this report also made the previous rumors more and more true.
In response to the rumors and reports that "some companies have given up pre-training", Lee Kai-fu has posted on Moments to refute the rumors, saying that 010,000 has been doing pre-training.
In fact, the rumors are like this, you don't care about it, others are just guessing. But if you are serious, you may be able to stand firm in other people's opinions.
Since it is a rumor, most of them may be chasing after the wind. But there is no wind and no waves, and the real situation may not be as serious as the rumors say, but it still puts a layer of pessimism on the domestic AI circle.
Even without these rumors, there is still Zhu Xiaohu's prediction. Its statement that "there will be no independent large-scale model companies in the next 5 years" almost predicts the fate of large-scale model startups, which is much more serious than these rumors.
Assuming that the life cycle of these large-scale entrepreneurial projects is only five years, it is normal for such a situation to occur now.
即使没有朱啸虎的预测，也还有最近兰德公司（RAND Corporation）的研报。 其所发布的报告《The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed Avoiding the Anti-Patterns of AI》超过80%的AI项目失败了，是不涉及AI的信息技术项目的失败率的两倍，浪费了数十亿美元的资本和资源。
The report points to several reasons for project failure: misalignment of goals among stakeholders, excessive focus on and application of AI without considering its value, lack of properly prepared datasets, inadequate infrastructure, and incompatibility of AI with the problem at hand. It also notes that these problems are not limited to the private sector: even academia has problems with AI projects, many of which focus only on publishing AI research rather than looking at real-world applications.
The report's views coincide with Robin Li's previous views. He has also said that China has too many large language models and wastes a lot of resources, because these models often have little practical application.
This is probably one of the important reasons why many investment institutions dare not rashly invest in large-scale model projects.
In fact, not only in China, but also in the world, the situation is the same.
Although OpenAI is valued at $157 billion, it still has to deal with billions of dollars in annual losses and huge operating costs in the coming years. OpenAI is forecast to lose $14 billion in 2026, according to the company's financial filings, and this projection does not include stock-based compensation, which is one of OpenAI's largest expenses.
OpenAI predicts that the computational cost of model training will rise significantly in the coming years, reaching as much as $9.5 billion per year by 2026, not including the upfront training costs for large model research. More than $200 billion is expected to be spent by 2030, 60 to 80 percent of which will be spent on training and running models.
As for profits, OpenAI expects the company's total loss to reach $44 billion between 2023 and 2028, with a profit of $14 billion until 2029.
The CEO of Anthropic also said that the current cost of AI model training is $1 billion, and this figure may rise to $10 billion or even $100 billion in the next three years.
This means that these two representative large-scale model projects will continue to burn money and raise funds before making a profit. The global large-scale model entrepreneurship projects are basically on par with these two projects.
At present, the AI industry has more than $500 billion in training costs alone, but the revenue is very small. If the future global large-scale model projects are all such inputs and outputs, most of the capital investment will not be converted into benefits. Faced with the situation that it is possible to achieve profitability after 5 years, how many companies and investment institutions can persevere?
That's why so many people worry that AI will become a bubble and may lead to a new economic crisis after it is punctured.
Where is the future of LLMs?
Unlike systems like Microsoft Windows, large language models have become a software necessity for the digital operation of enterprise businesses, nor have they become a hardware necessity for people's work, study, and life like iPhone. Its launch is not in the form of a terminal product for everyone to use at hand, but requires a certain learning threshold to be better used.
At the same time, the way of technology provision makes it no different from other technology suppliers, and the way of serving various enterprises and platforms brings benefits to the upstream and downstream of the industrial chain, but also further dilutes its profits. Although a business model like GPT Store has been created, it will take time to create a complete ecosystem like the App Store, which is enough for more business competitors to emerge.
In order to further make up for the technical limitations and capabilities of large language models, which are not sufficient to be efficiently applied to production, more manufacturers and users need to invest more resources and energy. If you want to be more capable, you need to invest more resources and burn more money. In addition, the contradiction between technological breakthrough and safety management has always been the highlight of various manufacturers.
Large language models belong to the miracle of vigorous efforts, but behind the vigorous efforts is continuous burning money.
Of course, this is not to say that there is no hope for large models in the future.
Although OpenAI burns a lot of money, its released models have always been industry-leading and have been valued by some big customers. If you maintain enough advantages, you are more likely to get the investment. If it can transform into a profitable company and complete the listing within two years, it will be able to obtain more investment.
Of course, OpenAI's revenue capacity is also improving, and after the peak of losses in 2026, the loss range in 2027 and 2028 may be significantly narrowed.
So it's understandable that now OpenAI looks at money, after all, there are more possibilities to be alive.
Compared with OpenAI, if a number of large-scale model projects in China can lead in a certain vertical field and provide stable and efficient application services for certain user groups, it will be more likely to obtain further financing under the condition of business improvement and gradual profitability.
In addition, it is also worth considering to make some fuss about the non-Transformer architecture model that is more "cost-saving and labor-saving", and to do some large-scale things with small-scale models.
Since the beginning of this year, domestic large-scale model projects have basically received a new round of financing. According to relevant statistics, as of August, there were 20 cases of financing of domestic large-scale model companies at the level of 100 million yuan. Five companies in the "New AI Six Little Dragons" (Zero One Everything, MimiMax, Baichuan Intelligence, Zhipu AI, Leaping Stars, and the Dark Side of the Moon) have received 100 million yuan in financing this year, and another has also reported that it is raising money.
Among them, the dark side of the moon has reached a valuation of $3.3 billion after completing a new round of Tencent investment of $300 million in August, and in September, Zhipu once again got a new round of financing after the Middle East consortium, achieving 6 rounds of financing in two years.
Wang Jiwei Channel believes that the continuous injection of state-owned capital and large technology companies not only provides a steady stream of power for these large-scale model projects, but also helps these projects to better improve the ecology and more industry resources.
In terms of technology, domestic large models are relatively lagging behind, but this does not prevent the landing and application of large models in some fields. As long as we can launch technologies, products and solutions suitable for domestic enterprises, this soil is more than enough to feed a few large model manufacturers.
Although the LLM based on the Transformer architecture is not the ultimate bridge to AGI, it is destined to shine at this stage, and its development space is far from the peak, and it will create enough commercial and social value in the process.
And those large-scale model projects, regardless of whether they sell themselves or develop independently, their efforts and innovations can contribute to the process of domestic large-scale models.
The Goldman Sachs report, while worrying about too much investment in AI and too little return, also predicts that about $1 trillion will be invested in generative AI and related infrastructure in the next few years, but these investments do not seem to be paying off significantly.
Morgan Stanley believes that the market's concerns about Microsoft's AI monetization have weighed on its stock price, but is confident in the growth of commercial returns for Microsoft's AI business.
Zhu Xiaohu believes that generative AI may be a long-term opportunity in ten years like the PC Internet and mobile Internet, and believes that the slowdown in the iteration speed of large models will increase the opportunities for application innovation, and there may be a large number of opportunities in AIGC applications from this year.
Recently, when I was talking to a young friend about work, he said that he often uses Kimi to search and find relevant information. And a few elders around me also have bean bags on their mobile phones.
This is a good start for large model applications.