The best-selling Android phone that's going to beSolid Large ModelsWhat's going on?
neverthelessPublished and ready to useThat kind - the new version of the phone system comes straight off the shelf, not a whole lot of fluff.
It is important to realize that while domesticLarge ModelHundreds of flowers blossomed, but the cell phone side of the "hundred model war", can be said to have just entered the warm-up stage.
There have been a number of cell phone manufacturers that have developed their own big models, but the ones that actually fit into the cell phone system thatAlmost nothing..
The big news is that for the second year in a row, the country's share ofFirstblue plantvivo.
Some time ago, Vivo quietly developed its ownCell phone AI macromodelingThe news of the rumor is spreading, and now the boots are on the ground, the official announcement is scheduled for November 1 Vivo developer conference, on board the new version of the systemOriginOS4.
Three key points can be glimpsed from the Vivo chief's dramatization: the
- Releasing 5 big models in one go
- The big model is embedded in the cell phone, as an assistant and can also paint.
- 7 billion versions of large models, open and available to the industry
Vivo vice president and director of the Global AI Research Institute around, but also two weeks in advance on microblogging "leaked" some of the capabilities of the big model:.
After a set of coherent operations, Netflix got excited first.
There are also netizens who are already fantasizing about a sudden drop in workload.
So, this Vivo official announcement of the big model, what signals have been released? Specifically on the cell phone side, what features will bring changes?
What signals have been dramatized in the big Vivo model?
Let's start with the reveal of the big Vivo model.
Vivo's new version of its own big model, namedvivoLM.
There are a total of VivoLMs to be released this time around5 versionsThe number of senators is divided intoBillions (1B/7B),Billion (66B)and100 billion (130B/175B)Three levels.
Whether it's the 7 billion large model version or the version with a larger number of participants, VivoLM is already ahead of the curve on the two major Chinese large model review listsC-EvalandCMMLUThe "top of the charts" is on the list.
On the C-Eval list, VivoLM achieved an average score of 82.3 out of 100 on the top one list, especially in STEM, humanities;...
On the CMMLU list, the VivoLM-7B version occupies the Top 1 in both Five-shot (only 5 samples are given) and Zero-shot (0 samples), and also "rides high" in the humanities.
It is worth mentioning that the vivoLM-7B that topped the CMMLU list, i.e., the 7 billion version of the large model, is exactly what the vivoLM will beopen to the outside worldThe version of the
Of course, no matter which version, VivoLM's purpose for this release is very clear - to
Put it in Vivo.up to dateversion of the cell phone system OriginOS4, as a cell phone assistantSensing user needs, recognizing intent and assisting in decision making.
However, it is well known that compared to other application areasMobile is actually one of the hardest scenarios to land on for big models.
On the one hand, by the size, power consumption limitations, the cell phone side of the arithmetic power compared to the cloud computing power "less than poor".
Big models, which are known for being "big", are often difficult to use if deployed directly on the end-side, and even if they do work, the reasoning is not as fast as expected, in the words of the surrounding.
If you do the input method's out-of-the-box word recommendation.2 seconds.It takes a word to come out.
But reducing the size of the model will definitely reduce the effect.
On the other hand, the advantage of end-side deployment is lost if the large model is directly uploaded to the cloud for networking use.
For example, the big model can originally be based on user information, personalized cell phone assistant on the phone, and to ensure that the information is not uploaded to the cloud; but if the big model is loaded in the cloud, it is inevitable that personal information will be uploaded through the network, and privacy and security can not be guaranteed.
On top of that, the cost of running large models in the cloud is very high. The surrounding presentation states.
Vivo has 300 million users in mainland China, if you use it 10 times a day, the computing cost is about 30 million yuan a day, and you need to spend about 9-10 billion yuan a year.
How exactly did Vivo deploy the big model to the mobile side?
From the VivoLM has dramatized the large model parameter settings, as well as some of the existing large model operation research, may be able to speculate on the use of some of the "tricks".
One of the tricks lies in the parameterization of the large model, with different parameter orders of magnitude, each used to handle different tasks.
The smallest of these large models, including those with 1 billion and 7 billion parameters, can be deployed directly on the end-side to ensure that power consumption is not high.
When it comes to user information and other personalized task requirements, it can be done with this type of large model, such as adding memo content to the calendar with a single click and setting an alarm reminder.
Larger macromodels, such as those with 66 billion, 130 billion, and 175 billion parameters, are called depending on the difficulty of the task.
After all, according to the current academic consensus, a few tens of billions of parameters are sufficient for a large model of "intellectual emergence".
Behaviors like summarizing emails with oversized models such as the GPT-4 have been teased as "delivering pizzas in a Lamborghini." (hand dog head)
Therefore, when encountering tasks with "lower difficulty coefficients", we can switch to smaller models, and for more complex tasks, such as those requiring higher context length and output effect, we can call the large model with hundreds of billions of parameters to complete the task.
The second trick lies in the way the big models are run, not relying on a single arithmetic power, but rather a collaborative cloud-based approach that balances running speed and experience.
Currently, hundreds of billions of large models are still difficult to deploy to cell phones, and even if they could be deployed, the speed of operation and power consumption would be unacceptable.
It is not impossible to run large models by end-cloud collaboration, at least from the research available in the industry today.
Taking speculative sampling/decoding, which was introduced by Google and DeepMind at the same time, as an example, this technique would be able to improve the speed of large model inference while ensuring the generation effect.
Specifically, it combines a large model with a smaller large model (draft model) to solve the "memory limitation" problem of large model reasoning.
However, this smaller model is not "easy to find", it must be "compatible" with the larger model, e.g., the interface should be uniform, the probability distribution should be close, etc.
vivoLM the five big models if each other "supporting", can use similar technology to achieve the effect of end-cloud synergy: large models in the cloud for computing, smaller models are placed on the end side of the operation, can save considerable reasoning costs.
With this size combo down, it seems to be a good idea to take a peek at what new features the new Vivo phone system with the big model will have.
How to useOn, vivoLM is expected to be in the form of voice assistant as the entrance, as the whole machine's "intelligent assistant" on call.
Specific functionsOn the other hand, it may be divided into three main categories.
Generative tasks, such as email intelligent composing, AI avatar generation, etc.; and
Complex task scheduling, such as summarizing call content in one click, setting up specific usage scenarios, etc.; and
Intentional understanding, e.g., customizing travel based on fuzzy needs, etc.
As the big model continues to "evolve", more features will emerge. Perceptions.
There are two application phases of big models on cell phones: assisted intelligence and active intelligence. In the active intelligence stage, there are also relatively high order presentations.
So, is Vivo's new system with the big model actually worth the wave of anticipation?
Why is it worth waiting for?
The most immediate reason, of course, is that playing on your cell phone can be more enjoyable.
And that, by the way, pushes more people into theup to dateA round of tech wave, embracing the big models and practically feeling the change of human-computer interaction under the influence of AI.
The deeper reason lies in the fact that Vivo, as the head phone maker, develops its own big model, which is expected in itself.
For one thing, Vivo's self-research AI big model must be user-oriented.
It has a clear application landing scene - cell phone; clear goal - to reshape the operating system with AI.
This makes Vivo's self-developed big model will be inherently more compatible with the cell phone operating system, which can inspire more potential for the big model to be applied on the end side.
Secondly, Vivo itself has a solid self-research background, and has been deeply cultivated in the field of AI for many years.
Now one day launched self-research big model, not only is the R & D strength of another self-evidence, more can be peeped out from this achievement Vivo for theup to dateUnderstanding of technology trends.
Vivo's AI strategy first started in 2017, when Shen Wei, founder and president of Vivo, proposed.
Artificial Intelligence is all about technological innovationBottomLayer technology is something we should be "oversaturated" with for a long time to come.
The Vivo Artificial Intelligence Global Research Institute also began preparations that year and was officially established in 2018.
As a result, Vivo has become one of the first Chinese cell phone companies to set up an institute specializing in AI.
The data shows that since its inception until now, Vivo AI Research Institute has maintained ascale of 1,000The R&D team. By 2021, Vivo has over 600 people focused on developing large models.
Since 2018, Vivo has accumulated a total ofTopOver 70 high level papers have been published in journals and continue to bring thetop notchAcademic results are translated into algorithms and engineering applications.
It has also set up the Institute of Graphics, which builds knowledge graphs through a team of professionals.Currently 2,800 terabytes of data have been accumulated.
△Vivo Vice President and President of the Global AI Research Institute around the
Outside of AI, Vivo's self-research strength in chips, system software, and imaging algorithms is also not bad.
Since 2021, three generations of Vivo's self-developed imaging chips have been launched.
up to dateThe first generation V3 uses a 6nm process to enable Android phones in thefirst4K cinematic portrait video and 4K-level post-shoot editing.
And it also enables a large number of self-developed imaging algorithms to be turned on at the same time.
Things like AI noise reduction, MEMC frame interpolation algorithm, HDR technology, and image zoom algorithms can all be handled simultaneously by the V3 and keep latency within user-acceptable limits.
In terms of system software, Vivo has come up with its own understanding of Android:OriginOS, which makes system interaction and design more humanized.
For example, Atomic Notifications and Atomic Components, which have received a lot of praise, are a good example of OriginOS' more intuitive interaction logic, where components can be used directly by tapping, touching, or swiping.
The sidebar in OriginOS3, with its scene recognition capability, matches the desired apps in the sidebar based on the interface the user is browsing.
In addition, Vivo has injected many innovative technologies into OriginOS, bringing a more silky smooth operating experience from the bottom of the system.
OriginOS's Smart Computing Hub is based on a three-layer framework of innovation, with the bottom layer achieving a breakthrough in basic capabilities, the middle layer optimizing resource scheduling, and the top layer upgrading visual and manipulation perceptions.
It has also developed its own "unfair scheduling", which prioritizes the scheduling of system resources to the foreground and suspends the demand of the background to improve the smoothness of operation.
In terms of more cutting-edge technological exploration, Vivo's layout is also quite ahead of its time.
For example, Vivo has started the 6G project since 2019, released a 6G series white paper in October 2020, and in 2022, it will show a series of 6G core technology prototypes to the public. Including communication and perception integration of wireless breath monitoring and target ranging and speed measurement, very low power communication based on backscattering, and AI communication four prototypes.
And these R&D investments in 6G are also claimed to have "no budget cap, invest as much as you need to."
△Communication Sensing Integration - Target Ranging and Velocimetry Prototype Prototype
To sum up, Vivo has developed its own big model that is really worth waiting for.
Not only because it is another head cell phone manufacturer that quickly follows the trend of big models, but also because Vivo has a deep understanding of the cell phone scene, as well as considerable technical reserves and self-research capabilities.
Its big model answer sheet can directly affect ordinary users, and also brings a new understanding and reference for the industry, and will even contribute to pushing the big model trend to the end side.
What does it mean for Vivo to develop its own big AI model?
First of all for the average person, a cellular operating system with a large model underpinning theAlmost the easiest entry point to embrace the big models.
up to dateA round of technology trends are sweeping in. Who doesn't want to get a quick taste, keep up with the trend, or even change their life with technology?
But there does remain a certain threshold for the use of large models. And for many people, theup to dateHow and where the technology should be used is still a question.
And the cell phone can just be a platform to connect the big models and ordinary people, and directly provide the use of the scene, the operation of the threshold is not high.
As a result, the underlying logic of the new generation of human-computer interaction has unconsciously permeated the mass population.
Secondly, for the industry, Vivo's entry into the big model is not unexpected, but an inevitable mission.
After ChatGPT brought up the trend in the industry, people are interested in thesuperAI assistants have been imagined more concretely, and Iron Man's Jarvis has been discussed over and over again for some time now.
Coming back to reality, isn't the closest thing to that role the voice assistant in a cell phone? So for the cell phone industry, injecting voice assistants with big modeling capabilities is almost a matter of course.
But that doesn't mean the matter is simple.
How can the average user get a deep sense of what the big models are capable of? How to clearly differentiate from the previous generation of voice assistants? How to realize the real meaning of more convenient and efficient? Behind these vague requirements, a clear action goal and specific technology is needed to support.
Who's got what it takes to fix these problems?
Of course it's the one that understands both AI and cell phones, but also the one that can reach the vast majority of ordinary consumers on a first line, and Vivo is one of the names that will come to mind.
Canalys report shows that in the second quarter of this year, Vivo topped the domestic smartphone market with 18% market share.
As a head manufacturer, Vivo's more thorough and comprehensive access to big model capabilities is not only a major upgrade of its own products, but also a positive push for the development of the industry.
The big model trend to the current development of the mature cell phone industry has brought new growth points, how to seize the trend, quickly follow up, but also to see the specific actions of each manufacturer.
From Vivo'sup to dateIt's easy to see in the movement.
- FirstVivo is sure of the big model trend and has decided to grovel into the game; the
- Second, Vivo believes that this will be an industry-wide action, from which they will release the big model at the same time, but also to contribute the ability of the 7 billion parameter version of the model, the ability to train, the ability to tune, the ability to hotplug data from various industries, all integrated and contributed.
Going a step further, Vivo also brings its own understanding to the industry.
Around previously revealed in an interview that.
vivoFrom a 3-5 year perspectiveTo look at the industry's potential conflicts and challenges: large model applications need to be on the top end if they want to solve the cost problem, and if they want to be on the top end they need to solve the problem of minimum system privileges and arithmetic and most importantly, memory resources.
In addition, to solve the problem of everyone's privatization and core competitiveness self-protection, Vivo will come up with solutions and contribute to the industry in the future.
And through this understanding and planning, it is not difficult to conclude that.
Vivo's entry into the big modeling trend is more than enough to create new growth for itself.
Combined with the industry's recent situation, it is not difficult to find that the big model of the end side of the application competition has revealed the first signs. Major cell phone manufacturers have bet on the big model, self-research or cooperation, in short, to first seize the advantage of the position in the track.
But getting in is still just completing the start.
After all, the application of large models to the ground, how to continue to iterate, constantly optimize and improve the experience is more critical, the slightest carelessness may be bent backward. This undoubtedly puts higher requirements on the team's technology development capability, product understanding and trend insight ability.
Therefore, the arrival of the big model end-side application trend not only brings a new growth cycle to the cell phone industry, but also brings a technology upgrade cycle covering the whole industry.
It is no wonder that some people have suggested that this year, the big model to reshape the search engine, next year may have to reshape the end side. This also makes people look forward to the big model researched by cell phone manufacturers, what new changes will be brought?
On November 1, Vivo's self-developed large model will officially meet everyone.