Machine Heart report
Editor: Zenan
The battle for big models for short video platforms has begun?
Kuaishou's AI big model landed on the ground, and this quietly began.
On the evening of August 18, Kuaishou announced the latest progress based on self-developed large language model applications - the "Kuaishou AI Dialogue" function has been opened for internal testing in the Android version of Kuaishou APP.
There is no need for a demo website, nor a separate classification, and the new feature of Kuaishou is directly placed on the official version of the Android APP. The AI conversation button next to this search box can be clicked by users who have qualified the test.
"Kuaishou AI Dialogue" is the first ChatGPT-like product based on a large language model in the short video live broadcast industry. We've been waiting for it for a long time, how is it experienced?
Be the first to review
Machine Heart obtained the qualification for internal testing for the first time, let's see what level Kuaishou's large model products are at.
Each big model has its own characteristics, let it introduce its own capabilities, so that you can judge where to start.
It says it can answer "any questions," offer advice, explain it, and of course make small talk with you, which seems to be a bit of a search engine revolution. Then let's ask it a headache first:
Not only is there an answer, but what is surprising and surprising is that it can also recommend relevant videos on Kuaishou, and it seems that it will be convenient to find resources with AI.
Searching for more concepts, we found that it can not only have a continuous conversation with people, but also give answers with links to quotes, so that they can be justified. If you query the weather or something, a weather forecast card will be displayed directly.
The news from Kuaishou is that there will be more and more tools in the form of cards.
From here, we can also see Kuaishou's thinking on the application of large model technology: give full play to its own advantages. Smart intelligence can rely on the rich content ecology on the site to help you quickly find short videos, influencers, encyclopedias and other content in the form of interactive dialogues. This is a completely new form of information acquisition.
At the same time, Kuaishou AI Dialogue also breaks the barriers of content on the site for Kuaishou Search, and can provide users with content retrieval services for the whole network. Your next search engine, why a search engine.
It is worth mentioning that when we tested, Kuaishou AI responded quickly to answer every question. Given that the speed at which a single A100 can reason about Llama under highly quantized (llama.cpp) conditions is 30 words per second, am I using an NVIDIA A100 now? It can be directly launched in the application without warning, and presumably engineers have made a lot of optimizations.
Next, the big domestic models all say that their Chinese semantic understanding level has advantages, and we let it do a college entrance examination language question this year:
Answer correctly, try the physics question again:
Science topics can also be done correctly, which shows that Kuaishou AI has read a lot of books.
As a large language model, the level of copywriting must be high. Most companies publish large models that can quickly generate informative articles, suggestions, and responses, but many of the generated text just looks usable but doesn't stand up to scrutiny. For specific tasks, or problems that require a deep understanding of context and human emotions, sometimes AI will output strange results.
On Kuaishou, we will always meet all kinds of masters, but most of them are newbies, can we use the help of AI to quickly gain fans? It's worth trying it out. The hot topic recently is computer installation, I want to turn over profiteers and open an installation shop:
Kuaishou Big Model gives long suggestions, from establishing positioning, understanding the market, training employees, formulating marketing strategies... There are eight counts. From the perspective of previous practice, it is very likely to ignore certain details when doing such a big thing, and the thinking of AI assistants is more meticulous, giving people a reliable feeling.
Let's continue to talk to it, after reading Kuaishou recommends choosing a store name, how to shoot a video to promote my store?
It seems that the content is decent, and it seems that there is no Kuaishou Laotie's unique idea, but it is clear which shots to give.
Since you speak so rigorously, let's test your logic problems:
The question that arises in the mentally retarded bar is not difficult to overcome here, but whether it is wonderful is the art of answering.
What about the problems of the Ministry of Education that Bang Tieli used to deal with civil science?
Although we didn't get it right, we know that the answer to this question is not any of the three options (the correct option 162 will appear in the subsequent dialogue in the game), and the answer given by the AI is relatively close.
Recently, open source Llama 2 is changing the landscape of large models, with an open source product comparable to the early GPT, what will many companies further explore?
There is still a lot to do in terms of technology and landing, the volume is certain, and of course there are many opportunities.
Kuaishou big model, come prepared
Big models are an important track for many technology companies to compete today, and Kuaishou has long been prepared in this regard.
At the results conference call in the first quarter of May this year, Cheng Yixiao, founder and CEO of Kuaishou Technology, revealed that the company has set up a large model research and development team, relying on the past technology accumulation in AIGC algorithms and large language models, it is promoting the development and training of large models as planned, and everything is progressing smoothly.
In fact, there is already information that Kuaishou is making rapid progress in the implementation of large models.
It is revealed that compared with the large model demos displayed by many technology companies, "Kuaishou AI Dialogue" uses efficient indexing of internal resources to partially solve the illusion problem of large language models and improve the accuracy of question answers, on the other hand, it has also begun to explore products and functions in multiple application scenarios very early.
On July 8, Kuaishou announced that it is conducting a closed beta of its Smart Q&A product, planning to bring users new features such as intelligent Q&A and text creation to users in search scenarios, followed by the launch of the "Kuaishou AI Dialogue" closed beta on August 8. With the search scene as the entrance, we have obtained a more convenient search form, and the convenience coverage brought by the new technology is also wider.
This morning, the big Kuaishou models behind these product features finally surfaced. According to media reports, KwaiYii, a self-developed large model from Kuaishou AI's team, has started internal beta testing and provided standard APIs and customized project cooperation solutions for business teams. Since its inception, the model has been at the top of several well-known large model rankings, and it is at the top of the latest CMMLU Chinese rankings.
KwaiYii-13B ranks first in CMMLU Chinese.
According to its GitHub homepage, Kuaiyi is a large-scale language model independently developed by Kuaishou AI team from zero to one, and it is its new version KwaiYii-13B that has attracted attention this time. The model is at the forefront of several benchmarks, proving its excellent performance in natural language processing tasks.
Among them, the KwaiYi-13B-Base pre-training model has excellent general technical base capabilities, and is currently at the leading level of the same model scale on MMLU, CMMLU, C-Eval, HumanEval and other benchmarks, showing that the Kuaiyi large model has very strong capabilities in Chinese and English.
How KwaiYii-13B-Base looks on Benchmark (Source: Kuaiyi GitHub homepage).
The KwaiYii-13B-Chat dialogue model has excellent language understanding and generation capabilities, and supports a wide range of tasks such as content creation, information consultation, mathematical logic, code writing, and multi-round dialogue, and the manual evaluation results show that KwaiYi-13B-Chat surpasses the mainstream open source model, and is close to the same level of ChatGPT (3.5) in content creation, information consultation and mathematical problem solving.
What is more worth looking forward to is that the Heart of the Machine understands that the Kuaishou AI team will continue to iterate the "Kuaiyi" big model, on the one hand, it will continue to optimize model performance and develop multimodal capabilities, on the other hand, it is also promoting the landing of more C-side and B-side business scenarios.
From productivity tools, mobile OSs to short video live broadcast platforms, it only took a few months for large domestic models to quickly cover various platforms, and the technology behind them is becoming more and more reliable. The change has reached this point, and the quantitative change is already producing a qualitative change.
The arrival of big models is bringing us into a whole new world.