Home About us

In pursuit of TC-PMF, will zero one become AI first?

Internet things 2024/07/08 10:32

Determined to be comparable to GPT-5's zero and one things, since the Yi open source version, it has released a closed-source large model Yi-Large.

It is reported that the Yi-Large closed-source model outperforms the top five international models in the evaluation set of authoritative instruction compliance at home and abroad.

In fact, in May last year, six months after its establishment, Zero One Everything released its first Chinese-English bilingual large-scale model Yi series. And from the very beginning, Kai-Fu Lee set a grand and difficult goal: "to become the world's No. 1".

Born as a "popular spicy chicken", 010000 also relied on the background team of a large factory and excellent model performance, reaching a valuation of 1 billion US dollars.

At the YI-Large press conference, Kai-Fu Lee also announced that the next generation of Yi-XLarge MoE model training has been launched, which will impact the performance and innovation of GPT-5.

But as a believer in AGI, we need to think more about catching up with GPT-5 without affecting the development path of AGI?

What about the Internet, the pursuit of TC-PMF, will zero one become AI first?

1. In the fierce market, how far is Zero One Everything from AGI?

In an interview with APPSO, Kai-Fu Lee said: We are pragmatic AGI believers, and we must use the least chips and the lowest cost to train the best model we can train. At the same time, we will continue to explore and find TC-PMF.

It is reported that at present, the total number of users of overseas productivity applications is close to 10 million, and the revenue of a single ToC product is expected to reach 100 million yuan this year.

However, the receivable of a single product of 100 million yuan can only represent TC-PMF, but cannot represent that 010000000 things have achieved TC-PMF at the AGI level.

It is important to know that unlike weak models such as NLP and VC, AGI after concretization is an omnipotent agent that integrates the knowledge of industry experts in various fields, and is a general artificial intelligence that can help the demander complete a set of requirements and has adaptability and initiative in the process.

The realization of AGI requires a high cost to a certain extent, and the cost of this may not be 10 billion or 100 billion US dollars.

But Kai-Fu Lee also said that zero and one thing will not use pure strength to produce miracles, and the pursuit of whether it can use 10 billion US dollars and 100 billion US dollars to realize the landing of large models in the scene.

What about the Internet, the pursuit of TC-PMF, will zero one become AI first?

Compared with the irrational ofo-style bloodshed and money-burning style in the industry, Zero One Thousand Things is more inclined to allow large models to use healthy and benign ROI to store energy for long-distance running.

But what we need to think about is that even if we find TC-PMF to develop, will we be able to get closer and closer to AGI?

In the development of AGI, cognitive artificial intelligence is the clearest, most definite and most direct way to AGI.

While there are several cognitive architecture projects that have been active for decades, none have shown sufficient commercial prospect to date, been widely adopted, or received particularly adequate funding.

The reasons are manifold, complex, but one common feature is that they are operating in a modular and inefficient manner, and lack in-depth learning feedback and cognitive theory.

Looking back at the development of AI hardware in recent years, we will find that in fact, the process of AGI in hardware has been constantly breaking through, so what really affects the realization of AGI is not the obstacle to the emergence of software and hardware, but the biggest obstacle is actually a sufficiently accurate development project and a large amount of financial support.

What about the Internet, the pursuit of TC-PMF, will zero one become AI first?

"Accurate development projects" may also be difficult for the current development of weak models. However, it is actually more difficult to achieve AGI and TC-PMF through the existing open-closed source large model and one-stop AI platform Know.

Because at present, Wanzhi, a one-stop AI workbench that can do meeting minutes, weekly reports, writing assistants, speed reading documents, and PPT, is positioned as a 2C productivity tool, but it is also more of a text generation model in the process of application.

This is still far from being able to realize an agent that can help the demander complete a complete set of requirements landing.

Whether it is zero one thing or other large model players, it seems that they are more committed to a narrow sense of artificial intelligence, in order to quickly land a large model in a specific scene.

For example, the Yi series of large models of zero and one things involve AI writing, AI programming, medical treatment, consumer 3C, biochemical environmental materials and other fields.

But an objective criterion for the development of AGI is whether the AI work done in the process of achieving AGI has clearly defined steps or an overall detailed plan, and very few AI jobs meet this standard, including zero and ten things.

For Zero One Things, what can be seen at present is actually the core methodology of Zero One Things to make a large model, such as the integration of the model base - the parallel development of the model and AI Infra; Simulation integration - the model and application are developed in parallel.

2. How can AGI believers be vigilant against the "narrow AI trap"?

However, in the integration of basic models, Zero One Everything not only develops its own AI Infra, but also sets AI Infra as an important direction, and highly co-builds the model team and the AI Infra team, with a ratio of 1 to 1 people.

Of course, for everything that has its own ROI requirements, pragmatic tactical development is often more prudent. At the same time, we should pay attention to the participation of talents, which may help Zero One Everything to better develop towards AGI.

In order to get closer to AGI in the true sense, the market needs to move from the second wave of AI to the third wave of AI, and from statistical generative AI to cognitive AI.

That is, from a large model characterized by statistics and reinforcement learning to a large model with autonomous, real-time learning, adaptation, and advanced reasoning as the core.

However, if we are eager to verify the zero and one things of TC-PMF through Everything, how can we determine the path of taking the application layer, which is more conducive to the realization of AGI?

It is important to know that the process of the transformation of the AI wave is both simple and complex, simple is the simplicity of the whole transformation cognition, and complexity is not only a "turn around" to change from the development benchmark of large models, but also need to be vigilant against the emergence of "narrow AI traps".

What about the Internet, the pursuit of TC-PMF, will zero one become AI first?

In layman's terms, the "narrow AI trap" means that even if everything goes smoothly towards the predefined goals of AGI - with a good theoretical foundation and development plan, an excellent development team and abundant funds, as well as the correct target benchmarks and development standards, there are still hidden dangers of the "narrow AI trap".

Because the market-wide urgency to achieve AGI will also lead to the majority of enterprises eventually using external human intelligence to achieve a specific outcome, or make progress on a given benchmark, rather than in a way that integrates intelligence (adaptive, autonomous problem-solving) into the system.

To put it bluntly, the upgrade of large models that is biased towards specific tasks is actually contrary to the adaptability and initiative pursued by AGI.

If learning is reinforced, task-specific, the end result is likely to be: only narrow AI jobs that are nominally AGI. What's more, at present, the large models that can realize the application of specific scenarios and solve specific problems are collectively referred to as narrow artificial intelligence.

However, the results of zero and one things on the multimodal large model are obvious.

As the only way to develop AGI, the Yi-VL-34B version of the zero-100000 multimodal large model surpasses a series of multi-modal large models with an accuracy of 41.6% on the MMMU test set, second only to GPT-4V (55.7%).

If you put aside the comparison with GPT, the accuracy rate of 41.6% is not outstanding for a multimodal large model.

It is important to know that the integrity of information, the adaptability of the environment, the naturalness of interaction, and the universal application are the four most important aspects of multimodal development.

The completeness of the information requires the large model system to consider more information dimensions to determine the quality and accuracy of the final decision. Conversely, it can be said that the backward inference of 41.6% accuracy is the finite data type of the multimodal large model, which leads to the system not taking into account more information dimensions when making decisions.

In addition, the 01000000 multimodal team is exploring multimodal pre-training from scratch to approach and surpass GPT-4V faster and reach the world's first echelon level.

GPT is not only the most wanted competitor of 0100000 to catch up, but also the opponent that hundreds of large model companies want to surpass, but when AGI has not yet taken shape, the target vision of 0100000 should be broader.

What about the Internet, the pursuit of TC-PMF, will zero one become AI first?

Image source: AI Frontline

After all, in the past seventy or eighty years of artificial intelligence development, we can see that every new wave of artificial intelligence is a major development of artificial intelligence brought about by the number of model parameters, the number of training samples and the jump in computing power.

Overseas countries are better at improving the number of model parameters by continuously investing in Scaling Law, while domestic countries can only take one step at a time.

For example, we will develop more cost-effective AI chips, more energy-efficient intelligent computing centers, and AI model acceleration technologies, and accelerate innovation in multimodal model architecture, data synthesis, and multimedia data annotation.

3. The focus of the third wave of AI may be monetization?

In the past two years, the market discussion on model performance improvement has focused on the training and algorithm improvement of multi-modal large models. For the initial scenario-based landing, all of them are free.

Perhaps, in a market with limited financing, the current focus of large model companies is not only to find commercialization, but also to pay more attention to the growth of large model parameters, in order to obtain more financing possibilities by expanding the imagination space.

After all, in the domestic capital market, the direction of funds is often more in the direction of certainty rather than betting on uncertainty.

Therefore, in the same way, in the period when it is difficult to obtain financing in the future in 2023, large models are more inclined to move towards 2B2C to make money and survive. On the C-side, such as OpenAI and Midjourney, Perplexity provides productivity liberation tools to individual users and monetizes them in a monthly subscription model.

On the B side, such as Microsoft and Salesforce, AI technology is integrated into traditional products and provides vertical customization services, which can be monetized with monthly subscriptions or usage models.

What about the Internet, the pursuit of TC-PMF, will zero one become AI first?

Domestic enterprises that have not yet been realized are constantly in the process of looking for commercialization.

For example, Baidu launched the Wenxin Yiyan subscription model on the C-side, and the B-side provides the underlying architecture and solutions; 360 focuses on AI office on the C-side, and focuses on scenarios such as AI security and knowledge management on the B-side. iFLYTEK tries to combine large models with its own hardware products.

At present, although the Wanzhi AI assistant is completely free and open to users, it is reported that the follow-up Wanzhi will launch a charging model based on product development and user feedback.

He believes that the development of large-scale C-end products can be divided into six stages: the initial stage is to use it as a productivity tool, and then gradually expand to the fields of entertainment, music and games.

Then we move on to search, then to the e-commerce market; Later, it further expanded into social and short video fields; Eventually, it will develop into the O2O product stage.

However, in the domestic market, C-end users do not seem to lack productivity tools, nor entertainment and social tools. After opening the fee, can Wanzhi really go as smoothly as overseas?

At present, according to the form of the left-handed C-end and right-hand B-end of the head large model, the large model mainly charges to the B-end, and the C-end is less charged, and the user's willingness to pay is low.

This is also destined to be among the many large models that position productivity, and most C-end users will be more inclined to low-pricers.

What's more, judging from Similar's data, the top three AI products with the largest domestic visits on the web in May were Kimi, Wenxin Yiyan, and Tongyi Qianwen, with 22.5 million, 17.8 million and 8 million visits, respectively, while Wanzhi ranked tenth, with only 320,000 visits.

Facing overseas, the achievements of 0100000 things may be very eye-catching, but at home, not necessarily.

Reference:

Intelligently connected everything: AI unicorn "01 Everything" has accelerated the layout of the overseas matrix, and Kai-Fu Lee, who "only does To C", is gradually entering the AI 2.0 era

APPSO: 010000000 Parametric Model Yi-Large, Kai-Fu Lee: China's large model catches up with the United States and aspires to compete with GPT-5

Z Finance: Deep丨AI unicorns with heads and faces are stepping up to go overseas, and Yi-Large has landed on Fireworks.ai, the world's leading model hosting platform

Z Finance: Depth | The "New AI Four Little Tigers" with a valuation of over 10 billion yuan was born, and zero and one thing was left behind, and it was out of the game light years away!

AI Frontier: Zero One Everything released the Yi-VL multimodal language model and opened it up, second only to GPT-4V in evaluation

This article is from Xinzhi self-media and does not represent the views and positions of Business Xinzhi.If there is any suspicion of infringement, please contact the administrator of the Business News Platform.Contact: system@shangyexinzhi.com