Is the AI ​​bubble that DeepSeek broken, a blessing or a disaster for Crypto AI?

Author: BUBBLE, BlockBeats

In January 2025, the launch of DeepSeek R1 caused a shock in the AI ​​industry, and it also truly changed the Crypto AI ecosystem. In the past cycle, Crypto AI has mainly revolved around the AI ​​Agent, and DeepSeek R1 and its open source strategy have completely changed the rules of the game: extremely low training costs and breakthrough adaptive training methods, making the vision of decentralization of the AI ​​industry no longer It is empty talk, but a reality within reach. This change has far-reaching impact, the total market value of the Crypto AI market has shrunk significantly, and many AI tokens have experienced a 70% pullback, but is this really a crisis? Or does it mean a complete reshuffle of Crypto AI? Is DeepSeek the "terminator" of Crypto AI narrative or the "breaker" who accelerates its entry into the era of practicality?

DeepSeek growing wildly

DeepSeek's development dates back to 2021. At that time, hedge fund Huanfang, which focused on quantitative trading, began to recruit AI talents on a large scale. It was rare for quantitative companies to switch to AI, and most of them recruited were exploring cutting-edge directions, including fields such as big models (LLM) and literary and graphic models. Although there are rumors that Huanfang has made a transformation to better utilize the idle GPU resources in the company, most of the reason should be the decision to seize the commanding heights of cutting-edge AI technologies such as big models.

By the end of 2022, Huanfang has attracted more and more top AI talents, mainly students from Tsinghua University and Peking University. Stimulated by ChatGPT, Fantasy CEO Liang Wenfeng was determined to enter the field of general artificial intelligence and established DeepSeek in early 2023. However, with the rapid rise of AI companies such as Zhipu, Yuezhi Dark Side, Baichuan Intelligent, DeepSeek, as a pure research institution and lacks star founders, he faces huge difficulties in independent financing. Therefore, the magic party chose to divest DeepSeek from its main body and fund its development wholly. Although this decision is extremely risky, DeepSeek does not need to be under the profit commitment or valuation pressure of the financing party. At the same time, it has relatively sufficient GPU resource reserves, allowing the team to focus on technological breakthroughs, and a group of innovative young people can rush around in a happy land. At this moment, DeepSeek is more like a research institute than a company.

Just like the early days of OpenAI, no one would think of how companies that study robots play Rubik's Cubes and how to finally develop ChatGPT, and no one could think of how the quantification company, Huanfang, used DeepSeek to break through the current AI bubble. The former It took 7 years, and the latter lasted only 2 years. In November 2023, DeepSeek LLM with a parameters of 67 billion and performance close to GPT-4 was launched, DeepSeek-V2 was launched in May 2024, and DeepSeek-V3 released in December of the same year performed with GPT-4o and Claude 3.5 in benchmark tests. Sonnet remained flat. DeepSeek's rapid technological leap is not because of the company's financial resources or high education, but after a technological singularity, "ChatGPT affects the world's AI industry", and the acceleration of singularity of all sizes occurs in any soil that can satisfy imagination. until the next key singularity appears.

Finally, in January 2025, DeepSeek accelerated through the singularity, opening the door with the first generation of reasoning-capable big model DeepSeek-R1 they developed with training costs and excellent performance that were much lower than ChatGPT-O1.

Use open source to distribute the key to the world to open the

interstellar gate

Just one day after DeepSeek R1 released and released its open source model, US President Trump officially announced the start of a "Star Gate" program with a super-large investment of $500 billion. OpenAI, SoftBank, Oracle and investment company MGX jointly formed a joint venture called Stargate to build a new artificial intelligence infrastructure for OpenAI in the United States.

This investment of magnitude is even comparable to the "Manhattan Plan". It is quite necessary to use national strength to use algorithm stacking to push closed-source AI to a climax and monopolize the AI ​​market to ensure the leading position of the US local AI industry. But the current plan was released, it shouldn't have expected that a few days later, the open source model on the other side of the ocean would not open the door. Not only did it bring a hammer to smash the wall by the door, but it also sent hammers to others.

As an open source model that can match the top closed source models, DeepSeek's new training architecture triggered a chain reaction, making closed source AI difficult to move forward. The closed source model of DeepSeek R1 will be directly eliminated by the capital market, even A16z "OpenAI" Marc Andreessen, founder of Investors, has publicly stated that it is necessary to pay more attention to open source AI than to closed source AI. Whether it supports AGI, it may generate or support AI, it can only be used as an upgraded version of the SaaS industry. It is believed that the harm of closed source is far greater than that of open source. No matter it is black box, industrial monopoly, information security, and capital manipulation of attention, any one is a very dangerous development direction.

Although some industry insiders require huge data sets for V3's hybrid expert technology "MoE", it is suspected to be distilled using OpenAI's model. And for the reinforcement learning method based on R1's reinforcement learning "RL", it requires a large amount of hardware resources, and it is suspected that it has made a false statement about the number of training chips used. But it does not affect the industry structural reform it brings in at all.

The open source of DeepSeek R1 breaks the business logic of OpenAI's closed-source big model in the training architecture, and uses the logic that allows the model to evolve itself to avoid a large amount of investment in computing power and data annotations in the traditional paradigm. Although the training model is still open to a blind box, it is blind. The cost of the box is much lower.

At the AI ​​hardware level, DeepSeek's V3 open source directly challenges Nvidia's market dominance. Nvidia's GPU's moat is largely its bottom parallel computing platform and programming model CUDA, its extensive ecosystem and enough developers to make Using non-NVIDIA chips to do training is too expensive, and the high-threshold purchasing conditions and political restrictions have divided the development of global AI.

For us, in the short term, the AI ​​in the US stock market has shrunk sharply, and the total market value of Crypto AI is almost cut, and the market has entered a bear market. But in the long run, the most recognized AI industry is moving towards an open source, transparent and decentralized development path. From any perspective, the combination of Crypto and AI will be more tacit.

Crypto AI’s redemption, move forward! go ahead! Advance by any means

During this round of Crypto AI bubble bursting, many AI concepts tokens have accepted a 70% pullback. The Crypto AI market has shrunk sharply. Some people jokingly say that "A large model can be trained for $5.5 million. If these AIs have a market value of more than they can buy. Crypto AI”. Admittedly, Crypto is a market dominated by capital markets, not product-dominated, and 90% of AI tokens have no practical significance.

But in fact, with the improvement of the crypto market supervision system, the crypto market is still the most suitable soil for small and medium-sized AI companies to start businesses. The 1/100 large model cost and model training method brought by DeepSeek will bring more than ten thousand times the current market growth compared to the current market.

Directly speaking, what DeepSeek brings to crypto is that the decentralized training model makes Depin-type projects more rational, making the training process and information feeding more transparent, and making the value reward mechanism for contributors of the data set more reasonable. , making it easier for both supply and demand parties to settle the model training. The surrounding ecological development of the AI ​​industry of more than ten thousand times has further improved the richness of the downstream industry of Crypto AI. When enough competitive and creative product narratives appear in the market, as long as one of them truly breaks the circle, external funds will Naturally, value will flow back into Crypto. The market has been suffering from PVP for a long time. The series of celebrity coins harvested after TrumpCoin broke the originally abundant liquidity and positive feedback balance of the AI ​​market. Therefore, the bubble burst by DeepSeek is actually a greater benefit.

Currently, many Crypto AIs or have quickly integrated DeepSeek or updated on their architecture, including ElizaOS, Argo, Myshell, Build, Hyperbolic, Nillion Network, infraX, etc. Some of these projects are optimized directly on the product side through DeepSeek.

Myshell

V3 and R1 are added to the production flow of chatbots and application plug-ins, and even the image generation model Janus-Pro is added. Myshell technicians completed the model integration in almost half a day, as one of the rarest blockchains. Projects that insist on polishing products and even make a name for themselves in Web2AI products and are reluctant to issue coins. This time, DeepSeek's open source will bring good news to Myshell users at the cost side, and lower costs will be the original product The already-completed Myshell brings more Agent developers.

Argo

Argo developer Sam Gao DeepSeeked the important functions of Argo in the early stages of designing the product. As a workflow system, Argo built LLM into the standard DeepSeek R1 and handed over the original workflow generation work to DeepSeek. R1 was performed. Because of WorkFlow, the token consumption and context information will be huge. "Average >=10k Token", and Argo also incorporates CoT "Chain-of-Thought" into the WorkFlow thinking process. After the open source of DeepSeek, it not only reduces the cost of workflow products, but also allows LLM to be deployed locally in Argo, so that users' privacy and security can be guaranteed.

Before the launch of DeepSeek R1, Argo had integrated its model preliminary training logic Chain-of-Thought "CoT" into the production process of Argo's Agent Workflow. Especially for tasks such as meme trading and market trend analysis, Argo customizes its workflow using Graph-of-Thought (GoT), a novel approach to constructing reasoning into a graph where nodes represent “LLM ideas” , side indicates the dependence between these ideas.

Given that Argo chose GoT, the only Crypto AI Workflow that currently uses this model, has achieved a more reliable and transparent process. This innovative approach directly affects the security and trust of transactions on the Argo platform. Thinking map ( GoT) is integrated into Web3 AI agents, putting Argo at the forefront of AI crypto transactions. CoT's structured reasoning not only enhances the security of financial transactions, but also ensures transparent and reliable decision-making, which is in decentralized finance ( DeFi) is crucial.

It is worth noting that an article "EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers" written by the core developer of Argo, Sam and Shaw, wrote in collaboration with Shaw on how to diffusion models from large-scale text to image without damaging the overall generation performance of the model. In the paper that removes the concepts that do not want to appear, we received the help of DeepSeek researcher XingchaoLiu.

Hyperbolic

Hyperbolic Labs is also the first to announce hosting the DeepSeek-R1 model on its GPU platform, where users can rent Hyperblic GPU resources to run the DeepSeek-R1 model locally or in a designated data center without sending sensitive data to the servers of DeepSeek. This method not only guarantees data privacy, but also leverages the excellent inference performance of the DeepSeek model. At the same time, through Hyperbolic's decentralized computing network, users can obtain the efficient inference ability of the DeepSeek model at a lower cost, for startups or super individuals. Entrepreneurs and even simple AI efficient users will be very competitive solutions.

This round of bubble burst has indeed caused the Crypto AI market to suffer a heavy blow, and many AI tokens have lost their hype value. But essentially, DeepSeek is not destroying Crypto AI, but is forcing the market to accelerate evolution. After DeepSeek R1, the future of Crypto AI will no longer rely solely on speculation, but will reconstruct the directions of decentralized AI computing, economic incentive mechanisms for model training, fair distribution of AI resources, practical products, etc. The real challenge is whether Crypto can leverage the technological revolution brought by DeepSeek to build a truly valuable AI ecosystem, not just manufacturing concepts and hype.

This is not the end, it is evolution. Crypto AI needs to move forward faster and more radically. / 加速