Subscribe to leaf
Subscribe to leaf
<100 subscribers
<100 subscribers
Share Dialog
Share Dialog
The fire of the big model has been burning in this land for half a year. With Huawei, Jingdong, Ctrip three launch catching up with the late set, according to the Internet's usual paradigm, the domestic large model of this "new thing" also ushered in their own half-yearly examination.
Just with other business half-yearly test is different, such as new energy vehicles, cell phones, e-commerce platforms and other business forms of the half-yearly test, there is enough public data and information to support, easy to analyze the evidence, while the big model is still in a "black box" state, did not run out of a clear business model, the so-called data and information and other arguments are also The so-called data information and other arguments are also unavailable.
It is quite a joke that even from the perspective of product function, the big model has never given birth to a universal evaluation tool. For the ultimate goal of AGI, there are naturally a variety of evaluation methods, such as the "Squirrel and Mandarin Fish Method", which is the standard for C-suite users in China to "evaluate" big models.
For this reason, most of the major domestic manufacturers have not been able to open up the use of their own large models like OpenAI, but have instead implemented internal testing mechanisms.
Big model more landing exploration to the B end and G end tilt, for example, Tencent preemptive industry big model, as well as Huawei's Pangu 3.0, Jingdong Rhinoceros and so on. As the current head players focus on the track, its big model favors to show as much as possible the mature product form, with commercialization landing as the basic goal. For example, in order to rapidly promote commercialization of this type of large model downward popularity, in addition to the business landing-oriented, the localization and deployment capabilities have also become important reference indicators.
Even so, in the view of industry insiders has been "sent to the bowl in front of the" industry model is still a lack of enterprises to buy, industry model of the wind since June has blown a month, so far there has not been a large-scale commercial cooperation.
Therefore, it is not difficult to see that in today's investment market, the investment related to large models is concentrated in the secondary market rather than the primary market. Even if Wang Huiwen this level of bull entry, public news said its A round of financing is much higher than 230 million U.S. dollars, and its financing ability compared to Microsoft from time to time to receive tens of billions of U.S. dollars to feed the OpenAI is not comparable.
The fire of the big model has been burning in this land for half a year. With Huawei, Jingdong, Ctrip three launch catching up with the late set, according to the Internet's usual paradigm, the domestic large model of this "new thing" also ushered in their own half-yearly examination.
Just with other business half-yearly test is different, such as new energy vehicles, cell phones, e-commerce platforms and other business forms of the half-yearly test, there is enough public data and information to support, easy to analyze the evidence, while the big model is still in a "black box" state, did not run out of a clear business model, the so-called data and information and other arguments are also The so-called data information and other arguments are also unavailable.
It is quite a joke that even from the perspective of product function, the big model has never given birth to a universal evaluation tool. For the ultimate goal of AGI, there are naturally a variety of evaluation methods, such as the "Squirrel and Mandarin Fish Method", which is the standard for C-suite users in China to "evaluate" big models.
For this reason, most of the major domestic manufacturers have not been able to open up the use of their own large models like OpenAI, but have instead implemented internal testing mechanisms.
Big model more landing exploration to the B end and G end tilt, for example, Tencent preemptive industry big model, as well as Huawei's Pangu 3.0, Jingdong Rhinoceros and so on. As the current head players focus on the track, its big model favors to show as much as possible the mature product form, with commercialization landing as the basic goal. For example, in order to rapidly promote commercialization of this type of large model downward popularity, in addition to the business landing-oriented, the localization and deployment capabilities have also become important reference indicators.
Even so, in the view of industry insiders has been "sent to the bowl in front of the" industry model is still a lack of enterprises to buy, industry model of the wind since June has blown a month, so far there has not been a large-scale commercial cooperation.
Therefore, it is not difficult to see that in today's investment market, the investment related to large models is concentrated in the secondary market rather than the primary market. Even if Wang Huiwen this level of bull entry, public news said its A round of financing is much higher than 230 million U.S. dollars, and its financing ability compared to Microsoft from time to time to receive tens of billions of U.S. dollars to feed the OpenAI is not comparable.
The investment market is a qualified barometer. Obviously, the domestic big model in the half-year examination of the time node submitted by the answer sheet is not satisfactory, but also need a period of dormancy and polishing, in order to let the "story" come true.
Big model has no business model?
The business model should be at the forefront of the need for the country's big models to respond to the market's skepticism.
ChatGPT this has long occupied the user's mind of the head of the chair there is a significant decline in heat, the earliest release of the domestic general large model Baidu and Ali two also in a number of players to follow up after the "silence". The reason for this is that the business model of the generic large model failed to run through. Even in the court of public opinion has been recognized by users, but the commercial closed loop has never appeared.
To test the wider range of Baidu big model, for example, its commercialization application Wenxin Qianfan payment model is to call the number of token generated charges, the standard is 0.012 yuan / thousand tokens, the output of a thousand words of text to spend 0.12 yuan.
Leaving aside the speed of its cost recovery, 0.012 yuan / thousand tokens of charges seem cheap, but text generation often requires multiple interactions to obtain the desired results, multiple interactions prompt will increase the hidden cost of unlimited, after all, Wenxin Chifan is not a wave of employees.
Similar to the scene is the question and answer community, academic Sun Quan (a pseudonym) told Photon Planet, the use of the model application experience is similar to the search for high-quality answers in the question and answer community, its user thinking is the problem of granularity, and the willingness to pay is often only generated after finding high-quality answers. Therefore, Baidu has chosen to reason the number of texts as a payment criterion, only at present it is not possible to cover the commercial hidden costs.
If the B-side of the popular monthly payment, that is only the cost of the party from the user to their own, obviously not a long-term solution. ChatGPT face C-side users under the pricing of $ 20 / month, there is still a suspicion of jerry-building is the best evidence.
At present, the commercialization of general large models, whether B-end or C-end, is difficult to achieve break-even, but also likely to encounter compliance risks such as AI ethics, regulation and so on. Therefore, the industrialization and verticalization of large models have become a paradigm shift under the demand for landing.
On the contrary, although the industry model, although its product form began to land demand, but in the actual landing of the problem has yet to be solved.
A kind of reference case is the vertical to C model built on the basis of its own product ecosystem, for example, Zhihu announced early in the product of the internal testing of Zhihaitu AI and Ctrip asked released not long ago.
The advantages of these two companies in entering the large model track are the same, which lie in their own community ecosystems and the high-quality community content derived from these ecosystems. The content, as industry data, can become the training corpus for big models after simple cleaning. The subtle difference between the two is that Zhihu has been a content community since its inception, while Ctrip has only begun to make efforts to do content in recent years.
However, both Zhihu and Ctrip seem to have failed to address the user's pain points and improve their existing functions in the form of big models.
Zhihaitu AI has announced its product "Hot List Summary", which is to capture high-quality Q&A through AI and rewrite the synopsis to present to the user, while another application "Search Aggregation" is to aggregate views from the answers to improve the efficiency of users in obtaining information and forming decisions. efficiency of users' access to information and decision-making.
Recommendation, hot list, a kind of aggregation function is Zhihu "traditional arts", the performance of the big model empowerment in the user level did not set off a splash. Moreover, the process of AI rewriting and embellishment also covers the personalized features of popular answers, and for users, the function of this application only lies in the quick understanding of information, which is contrary to the differentiation and personalized communication advocated by the content community.
The OTA-based Ctrip Ask, in the opinion of Ctrip's Chairman of the Board, Liang Jianzhang, is a "reliable answer bank" for the tourism industry. The effectiveness of its products need time to test, but since the positioning of the view, there is the same "to the end" suspicion.
Tourism in the eyes of young users in the eyes of there is no standard answer, "special forces", "punch card", "immersion" and other diversified forms of tourism has proved this point. If we assume that a large number of users make travel route plans through AI, the uniform route plans will affect the community communication and atmosphere, and even lead to a decline in the user's stay time.
Generally speaking, it seems that the attempt to realize vertical modeling in the C-suite is not smooth, and even has the possibility of becoming a "sunk cost". Perhaps influenced by the myth of "improving efficiency", the product positioning is mostly limited to the word "efficiency", which is not a core dimension of user experience.
The same paradigm has been demonstrated in the field of to B, and in the pursuit of efficiency of the B-side, the business model of the industry big model and the implementation of the problem has been more profoundly demonstrated.
The black box that can't be understood
"AI is not physics, there are rarely any theoretical major technological breakthroughs, it is more about fine-tuning and small optimizations in dimensions such as model structure, data quality, etc., and even many times the model output is better and the team can't find out why."
In the view of an industry insider, the big model in the industry and outside there is a huge cognitive bias, and the reason is that the big model training and the AI industry for the outside world is a no compromise "black box", it is difficult to scrutinize the big model to produce the output results of the reasoning process, which can not be seen and cannot be touched.
This leads to the outside world in the ChatGPT to bring the frenzy period, once calmed down, will be on the big model of the "black box" attitude of caution. This will lead to the big model in the landing of the dilemma, and this phenomenon is more obvious in today's to B route to change the process.
To today's clear to B route of the big factory products, for example, including Tencent Cloud launched MaaS technology solutions, Huawei Cloud launched Pangu big model, relying on its own cloud computing ecosystem, are said to support the deployment of its big model services diversified deployment, including cloud deployment, localization and rapid deployment and so on. There are also achievements in interaction, operation, and subsequent addition of new industry data iterative optimization, etc. It can be said that the threshold of the big model has been reduced to a very low level in order to land.
However, the cognitive wall brought about by "prudence" has not been broken, and even though the wind of ChatGPT has been blowing for half a year, many enterprises do not have the motivation or interest to study how to import large models.
A few years ago, the cloud computing industry can be seen following a similar logic. Cloud computing is in the recognition of the value of data, as a basis for services and derivatives, as for the value of the big model in the enterprise, relatively speaking, it is a leap in the value of data. The same is the lack of technical capabilities of enterprise customers, even the popularization of cloud computing in the domestic enterprise is still far from the end of the road, the big model needless to say.
Industry model is good to use or not, in fact, has been unimportant, after all, the use of commodity value ultimately need to be tapped by the user. What's more, outsiders will roughly measure the level of the model through certain tests and performances, such as the "Squirrel and Mandarin Fish Method" or the Huawei Pangu Weather Model, which has been challenged recently because of errors in predicting the landfall location and intensity of the mega typhoon "Dusu Rui".
Perhaps this is why the recently released Jingdong Rhinoceros Big Model has chosen to prioritize its own business scenarios and is expected to open up to "external serious business scenarios" early next year.
What is worth mentioning is that, "the industry into the wind" under the commercialization of the so-called industry-oriented model to replace the original big model of the "general" narrative at the same time, but also suffered a lot of people's "lost". "The
The definition of the so-called industry model is ambiguous. A Foundation Model is not about the number of participants, but about the generic capabilities that emerge from training with generic data. If the same model architecture is used, but a single domain data is used for the data, not only the generic capability is lost, but even the domain problem cannot be solved due to the discount of emergence.
If the use of industry data on the basis of the original model to do the second pre-training, the equivalent of fine-tuning the original model, then the product itself is still in the model layer, can be called the industry model; such as through the Prompt or plug-in database to join the field of knowledge, it is only on the original model to stimulate the ability of the product should be attributed to the model of the application layer above the industry model is overstating the case.
At present, the vast majority of large manufacturers in the development of industry model are the former, such as Tencent, Jingdong, Huawei and so on. The latter is due to lighter investment and rapid improvement in the performance of the model ability, more will appear in the open source community, such as a period of time ago sparked a heated debate on the legal big model ChatLaw.
"Compared with the former, the latter is more mature in product form, which facilitates the rapid construction of modeling capabilities, but the latter tends to have a higher ceiling after completing the process of instilling domain knowledge", said an industry insider.
Open source threat
Recently, Meta made its latest open source big model, Llama2, freely available under an open commercial license and introduced it to Microsoft's Azure platform, a move that has been hailed as a major milestone for open source LLM and has even begun to threaten the position of closed source head honcho OpenAI.
Through Microsoft, the big model gold owner, Meta challenges OpenAI with a more open stance.
In fact, the "open source faction" has quietly risen as a third party before. "We don't have a moat, and neither does OpenAI." This sentence came from an internal document accidentally leaked by Google in May. Its content is to the effect that on the surface, OpenAI and Google in the big model you catch up with me, but the real winner may not come from the two, the reason for this judgment lies in the increasingly rich open source ecosystem.
The open source ecosystem has become more and more active, and even the emergence of Llama2, a representative of the modeling ability, and LORA, a representative technology of the Finetune (model fine-tuning) paradigm, all of which have made the closed-source giant vendors striving for "vigorously producing miracles" feel a clear chill.
Open source technology sharing and talent flow and other factors, but also in the big model of the black box more and more "glass", the lack of barriers to the inevitable result is a large factory in a huge amount of money, time investment in the Konw How easy for the open source community to be overthrown.
Domestic head of the big manufacturers to respond to most of the "two-handed approach". The left hand "shut the door to build a car", in the form of small-scale internal testing to continuously polish the product form and capabilities, the right hand "brainstorming", based on the cloud developer ecosystem to create an open source community within the ecosystem, but this just requires vendors from the computing power layer, model layer to the application layer of the full stack of the layout. Layout. Ali cloud launched a large model open source community magic ride GPT, Huawei cloud, Baidu cloud, Tencent cloud also have layout.
Overall, whether it is the industry or general, to C or to B, the big model of the six-month test gives us a direct feeling: landing difficulties, profitability expectations continue to move back; the risk of gradual strengthening, it is difficult to say that the technical barriers. So, where is the road to break the current situation?
For now, there are two interesting directions. One is the vector database known as the "Memory of the AI era", and the other is the intelligent hardware empowered by model intelligence.
The so-called vectors are multidimensional data that can represent anything, including text, which is most valued for LLM training today, as well as images, videos, audio and sound. These forms of content are clearly represented in a database and support semantic retrieval, i.e., retrieval by similarity, e.g., man vs. boy. In other words, vector retrieval is the SEO of large models.
As mentioned above, domain knowledge can be fine-tuned or externally tuned to improve the construction and use of industry models through vector database capabilities, which is naturally where the next phase of power lies for the big players. Since May, capital has been pouring into vector data-related tracks, and as an application layer product with a more certain outlook, vector data has also gained the close attention of a number of VCs.
As for the built-in model of intelligent hardware, it is a leap in the ability of intelligent assistants such as "siri" and "Ai", as well as an outreach to real intelligent devices (cell phones and computers). The open source community has long been the big parameter model built-in MAC attempts, while the big manufacturers in the past mobile Internet era has accumulated a certain amount of hardware production capacity, relatively speaking, its first-mover advantage is more obvious.
Less PR-style spring and autumn brushwork, landing has become the core needs of the big model is no longer mysterious, the story is also less and less, began to "deep dive" track players are still making efforts. The industry needs the next "ChatGPT" moment, so that we can see the divers surface and confront each other head-on.
The investment market is a qualified barometer. Obviously, the domestic big model in the half-year examination of the time node submitted by the answer sheet is not satisfactory, but also need a period of dormancy and polishing, in order to let the "story" come true.
Big model has no business model?
The business model should be at the forefront of the need for the country's big models to respond to the market's skepticism.
ChatGPT this has long occupied the user's mind of the head of the chair there is a significant decline in heat, the earliest release of the domestic general large model Baidu and Ali two also in a number of players to follow up after the "silence". The reason for this is that the business model of the generic large model failed to run through. Even in the court of public opinion has been recognized by users, but the commercial closed loop has never appeared.
To test the wider range of Baidu big model, for example, its commercialization application Wenxin Qianfan payment model is to call the number of token generated charges, the standard is 0.012 yuan / thousand tokens, the output of a thousand words of text to spend 0.12 yuan.
Leaving aside the speed of its cost recovery, 0.012 yuan / thousand tokens of charges seem cheap, but text generation often requires multiple interactions to obtain the desired results, multiple interactions prompt will increase the hidden cost of unlimited, after all, Wenxin Chifan is not a wave of employees.
Similar to the scene is the question and answer community, academic Sun Quan (a pseudonym) told Photon Planet, the use of the model application experience is similar to the search for high-quality answers in the question and answer community, its user thinking is the problem of granularity, and the willingness to pay is often only generated after finding high-quality answers. Therefore, Baidu has chosen to reason the number of texts as a payment criterion, only at present it is not possible to cover the commercial hidden costs.
If the B-side of the popular monthly payment, that is only the cost of the party from the user to their own, obviously not a long-term solution. ChatGPT face C-side users under the pricing of $ 20 / month, there is still a suspicion of jerry-building is the best evidence.
At present, the commercialization of general large models, whether B-end or C-end, is difficult to achieve break-even, but also likely to encounter compliance risks such as AI ethics, regulation and so on. Therefore, the industrialization and verticalization of large models have become a paradigm shift under the demand for landing.
On the contrary, although the industry model, although its product form began to land demand, but in the actual landing of the problem has yet to be solved.
A kind of reference case is the vertical to C model built on the basis of its own product ecosystem, for example, Zhihu announced early in the product of the internal testing of Zhihaitu AI and Ctrip asked released not long ago.
The advantages of these two companies in entering the large model track are the same, which lie in their own community ecosystems and the high-quality community content derived from these ecosystems. The content, as industry data, can become the training corpus for big models after simple cleaning. The subtle difference between the two is that Zhihu has been a content community since its inception, while Ctrip has only begun to make efforts to do content in recent years.
However, both Zhihu and Ctrip seem to have failed to address the user's pain points and improve their existing functions in the form of big models.
Zhihaitu AI has announced its product "Hot List Summary", which is to capture high-quality Q&A through AI and rewrite the synopsis to present to the user, while another application "Search Aggregation" is to aggregate views from the answers to improve the efficiency of users in obtaining information and forming decisions. efficiency of users' access to information and decision-making.
Recommendation, hot list, a kind of aggregation function is Zhihu "traditional arts", the performance of the big model empowerment in the user level did not set off a splash. Moreover, the process of AI rewriting and embellishment also covers the personalized features of popular answers, and for users, the function of this application only lies in the quick understanding of information, which is contrary to the differentiation and personalized communication advocated by the content community.
The OTA-based Ctrip Ask, in the opinion of Ctrip's Chairman of the Board, Liang Jianzhang, is a "reliable answer bank" for the tourism industry. The effectiveness of its products need time to test, but since the positioning of the view, there is the same "to the end" suspicion.
Tourism in the eyes of young users in the eyes of there is no standard answer, "special forces", "punch card", "immersion" and other diversified forms of tourism has proved this point. If we assume that a large number of users make travel route plans through AI, the uniform route plans will affect the community communication and atmosphere, and even lead to a decline in the user's stay time.
Generally speaking, it seems that the attempt to realize vertical modeling in the C-suite is not smooth, and even has the possibility of becoming a "sunk cost". Perhaps influenced by the myth of "improving efficiency", the product positioning is mostly limited to the word "efficiency", which is not a core dimension of user experience.
The same paradigm has been demonstrated in the field of to B, and in the pursuit of efficiency of the B-side, the business model of the industry big model and the implementation of the problem has been more profoundly demonstrated.
The black box that can't be understood
"AI is not physics, there are rarely any theoretical major technological breakthroughs, it is more about fine-tuning and small optimizations in dimensions such as model structure, data quality, etc., and even many times the model output is better and the team can't find out why."
In the view of an industry insider, the big model in the industry and outside there is a huge cognitive bias, and the reason is that the big model training and the AI industry for the outside world is a no compromise "black box", it is difficult to scrutinize the big model to produce the output results of the reasoning process, which can not be seen and cannot be touched.
This leads to the outside world in the ChatGPT to bring the frenzy period, once calmed down, will be on the big model of the "black box" attitude of caution. This will lead to the big model in the landing of the dilemma, and this phenomenon is more obvious in today's to B route to change the process.
To today's clear to B route of the big factory products, for example, including Tencent Cloud launched MaaS technology solutions, Huawei Cloud launched Pangu big model, relying on its own cloud computing ecosystem, are said to support the deployment of its big model services diversified deployment, including cloud deployment, localization and rapid deployment and so on. There are also achievements in interaction, operation, and subsequent addition of new industry data iterative optimization, etc. It can be said that the threshold of the big model has been reduced to a very low level in order to land.
However, the cognitive wall brought about by "prudence" has not been broken, and even though the wind of ChatGPT has been blowing for half a year, many enterprises do not have the motivation or interest to study how to import large models.
A few years ago, the cloud computing industry can be seen following a similar logic. Cloud computing is in the recognition of the value of data, as a basis for services and derivatives, as for the value of the big model in the enterprise, relatively speaking, it is a leap in the value of data. The same is the lack of technical capabilities of enterprise customers, even the popularization of cloud computing in the domestic enterprise is still far from the end of the road, the big model needless to say.
Industry model is good to use or not, in fact, has been unimportant, after all, the use of commodity value ultimately need to be tapped by the user. What's more, outsiders will roughly measure the level of the model through certain tests and performances, such as the "Squirrel and Mandarin Fish Method" or the Huawei Pangu Weather Model, which has been challenged recently because of errors in predicting the landfall location and intensity of the mega typhoon "Dusu Rui".
Perhaps this is why the recently released Jingdong Rhinoceros Big Model has chosen to prioritize its own business scenarios and is expected to open up to "external serious business scenarios" early next year.
What is worth mentioning is that, "the industry into the wind" under the commercialization of the so-called industry-oriented model to replace the original big model of the "general" narrative at the same time, but also suffered a lot of people's "lost". "The
The definition of the so-called industry model is ambiguous. A Foundation Model is not about the number of participants, but about the generic capabilities that emerge from training with generic data. If the same model architecture is used, but a single domain data is used for the data, not only the generic capability is lost, but even the domain problem cannot be solved due to the discount of emergence.
If the use of industry data on the basis of the original model to do the second pre-training, the equivalent of fine-tuning the original model, then the product itself is still in the model layer, can be called the industry model; such as through the Prompt or plug-in database to join the field of knowledge, it is only on the original model to stimulate the ability of the product should be attributed to the model of the application layer above the industry model is overstating the case.
At present, the vast majority of large manufacturers in the development of industry model are the former, such as Tencent, Jingdong, Huawei and so on. The latter is due to lighter investment and rapid improvement in the performance of the model ability, more will appear in the open source community, such as a period of time ago sparked a heated debate on the legal big model ChatLaw.
"Compared with the former, the latter is more mature in product form, which facilitates the rapid construction of modeling capabilities, but the latter tends to have a higher ceiling after completing the process of instilling domain knowledge", said an industry insider.
Open source threat
Recently, Meta made its latest open source big model, Llama2, freely available under an open commercial license and introduced it to Microsoft's Azure platform, a move that has been hailed as a major milestone for open source LLM and has even begun to threaten the position of closed source head honcho OpenAI.
Through Microsoft, the big model gold owner, Meta challenges OpenAI with a more open stance.
In fact, the "open source faction" has quietly risen as a third party before. "We don't have a moat, and neither does OpenAI." This sentence came from an internal document accidentally leaked by Google in May. Its content is to the effect that on the surface, OpenAI and Google in the big model you catch up with me, but the real winner may not come from the two, the reason for this judgment lies in the increasingly rich open source ecosystem.
The open source ecosystem has become more and more active, and even the emergence of Llama2, a representative of the modeling ability, and LORA, a representative technology of the Finetune (model fine-tuning) paradigm, all of which have made the closed-source giant vendors striving for "vigorously producing miracles" feel a clear chill.
Open source technology sharing and talent flow and other factors, but also in the big model of the black box more and more "glass", the lack of barriers to the inevitable result is a large factory in a huge amount of money, time investment in the Konw How easy for the open source community to be overthrown.
Domestic head of the big manufacturers to respond to most of the "two-handed approach". The left hand "shut the door to build a car", in the form of small-scale internal testing to continuously polish the product form and capabilities, the right hand "brainstorming", based on the cloud developer ecosystem to create an open source community within the ecosystem, but this just requires vendors from the computing power layer, model layer to the application layer of the full stack of the layout. Layout. Ali cloud launched a large model open source community magic ride GPT, Huawei cloud, Baidu cloud, Tencent cloud also have layout.
Overall, whether it is the industry or general, to C or to B, the big model of the six-month test gives us a direct feeling: landing difficulties, profitability expectations continue to move back; the risk of gradual strengthening, it is difficult to say that the technical barriers. So, where is the road to break the current situation?
For now, there are two interesting directions. One is the vector database known as the "Memory of the AI era", and the other is the intelligent hardware empowered by model intelligence.
The so-called vectors are multidimensional data that can represent anything, including text, which is most valued for LLM training today, as well as images, videos, audio and sound. These forms of content are clearly represented in a database and support semantic retrieval, i.e., retrieval by similarity, e.g., man vs. boy. In other words, vector retrieval is the SEO of large models.
As mentioned above, domain knowledge can be fine-tuned or externally tuned to improve the construction and use of industry models through vector database capabilities, which is naturally where the next phase of power lies for the big players. Since May, capital has been pouring into vector data-related tracks, and as an application layer product with a more certain outlook, vector data has also gained the close attention of a number of VCs.
As for the built-in model of intelligent hardware, it is a leap in the ability of intelligent assistants such as "siri" and "Ai", as well as an outreach to real intelligent devices (cell phones and computers). The open source community has long been the big parameter model built-in MAC attempts, while the big manufacturers in the past mobile Internet era has accumulated a certain amount of hardware production capacity, relatively speaking, its first-mover advantage is more obvious.
Less PR-style spring and autumn brushwork, landing has become the core needs of the big model is no longer mysterious, the story is also less and less, began to "deep dive" track players are still making efforts. The industry needs the next "ChatGPT" moment, so that we can see the divers surface and confront each other head-on.
No activity yet