Written by: Techub News Compilation

Introduction

Recently, during a dialogue organized by the cloud computing and data platform company Databricks, Dario Amodei, co-founder and CEO of the AI frontier company Anthropic, engaged in an in-depth conversation with Ali Ghodsi, co-founder and CEO of Databricks. They announced a strategic partnership to deeply integrate the Claude series models into the Databricks platform, discussing how enterprises can leverage AI to build competitive advantages, the future forms of AI agents, the synergy between data and models, as well as security and governance issues of joint concern across industries. Amodei is not only at the helm of Anthropic but also a deep thinker in the field of AI safety and capability expansion, and his views provide an important perspective for understanding the next generation of enterprise-level AI applications.

Summary

• The future form of AI will be agents, which can proactively operate tools, access data, and complete tasks.
• Proprietary data is the core asset for enterprises to build "irreplaceable value" and competitive moats in the AI era.
• Data governance, security, and privacy are key bottlenecks for the adoption of AI in enterprises, especially in regulated industries like healthcare.
• Open-source and closed-source models will coexist in the long term; the capabilities of the model itself and the accompanying risks are more important than its open-source characteristics.
• The Model Context Protocol (MCP) is intended to become a universal "adhesive" or "USB-C" connecting AI models with external tools/data.

The Grand Blueprint for AI: From Curing Diseases to Empowering the Economy

At the start of the dialogue, Ali Ghodsi asked Dario Amodei to share his vision for the future of AI. Amodei mentioned his previous article "Machines of Love and Grace" and stated that the writing came from a sense of "frustration"—he believes that many people's imaginations for the positive aspects of AI are not concrete and grand enough.

Amodei used his biological background as an example, emphasizing that biomedical innovation will be one of the areas where AI has the greatest impact. He believes that many complex diseases facing humanity today (such as cancer and neurodegenerative diseases) are essentially "complex system diseases," and AI's unique advantage in analyzing vast amounts of data and simulating biological processes holds promise for helping us tackle these challenges. He mentioned that the data and unsolved problems of a pharmaceutical company are areas where AI can truly shine.

Of course, Amodei believes the impact of AI will span the entire economy. Through collaboration with Databricks, they will reach companies across various industries, and AI will be widely applied. The future he envisions is: society will not only change due to increased productivity, but also undergo fundamental transformations due to leaps in human health and enhanced capabilities to solve long-standing issues.

Regarding the timeline for this future, Amodei takes a cautiously optimistic stance. He believes that foundational technologies could be ready "in the next few years, or even in a year or two." The real challenge lies in how to rapidly promote these technologies and ideas across society, especially enabling sectors that are far from technological innovation to keep pace and apply them. This is precisely where business collaboration and market promotion are needed, and it is also where the value of the collaboration between Anthropic and Databricks lies.

Data: The Ultimate Moat for AI Enterprises

When the topic turned to how enterprises can prepare for the AI future, both CEOs reached a strong consensus: data is the most critical strategic asset for enterprises.

Amodei distinguished between two types of data: one is large datasets used to train foundational models, which is clearly vital; the other is proprietary data specific to certain industries, companies, or customers. He believes the latter is equally, if not more, critical. Whether in pharmaceuticals, finance, or technology companies, they all hold an abundance of proprietary data that cannot be accessed externally. AI models cannot know facts they have not been exposed to, thus this exclusive data is the source of unique value that models can create for enterprises.

Amodei listed various ways to utilize this data: fine-tuning models; allowing models to operate and act directly on the data; incorporating data into context through techniques such as retrieval-augmented generation (RAG); or building agents that can run on the data. He emphasized that it is this proprietary data that establishes "irreplaceable value" between AI companies and enterprises that possess data, a value that cannot be replicated by either party alone. One core aspect of the collaboration between Anthropic and Databricks is to more efficiently combine powerful models with enterprise data assets.

Ali Ghodsi wholeheartedly agreed, adding that he has observed that the most interesting enterprise AI use cases often revolve around a client's unique, long-accumulated special data. This data forms the protective layer for the enterprise and is where they should innovate—using models provided by Anthropic, combined with their own data, to foster unique innovation. He summarized, "This is where AI will have a real impact."

Agents as the Mainstream Future of AI

Amodei clearly expressed his viewpoint: "The future of AI is primarily agents." He believes models will increasingly operate in the form of agents. Anthropic has already explored this direction, such as the "Computer Use" demonstration released last year, where models can receive screenshots from computers and take action; and the recently released Claude Code, an early agent focused on programming.

Agents will use various tools, one of which is accessing data in multiple ways—through databases, searches, etc. Amodei envisions that this collaboration will produce enormous synergies. Ali Ghodsi shared two specific use cases: the telecom giant AT&T uses their technology to analyze vast amounts of wireless user data to detect fraud; the payment company Block (formerly Square) utilizes generative AI, allowing merchants to configure their stores through dialogue, greatly simplifying the process.

Amodei further introduced Claude Code. It is essentially a specific way of using the Claude 3.7 Sonnet model, reverting to the concept of command-line tools, making it easy to integrate with Git and other tools. Users can request Claude to write code, create pull requests (PRs), etc. While it will ask for user permission before executing actions on the computer, it has already demonstrated the powerful capabilities of early agents. Within just a few days of release, nearly 100,000 people attempted to use it, and many enterprises expressed strong interest, fully showcasing the potential of coding models and agents.

Model Context Protocol (MCP): AI's "Universal Connector"

When discussing how to better enable models to gain context and data, Amodei highlighted the Model Context Protocol (MCP) launched by Anthropic. They found that there seems to be a lack of a standard "connector" between the models themselves and the common tasks they need to accomplish (like integrating tools, accessing data). These tasks are fundamentally similar, yet their implementations differ each time. Thus, they pondered: isn’t this a good problem for Claude to solve?

As a result, the MCP was born, functioning like an "adhesive" that connects models with the things they need to access and use (like data, tools). Amodei revealed an interesting phenomenon: within a few months after the release of MCP, interest surged explosively in recent weeks, making it nearly a hot topic on social media. He believes this indicates that people have recognized its practicality.

Ali Ghodsi confirmed that Databricks is adopting MCP, exposing all information within its platform to large language models (LLMs) through this protocol. Amodei likened MCP to "the USB-C of the AI era," a flexible, malleable universal connector. He believes the MCP will play an important role in their collaboration, as it seamlessly connects the technology stacks of both companies and hopes it will be widely adopted in the industry.

Governance, Security, and Trust: Non-Technical Bottlenecks That Cannot Be Ignored

As AI penetrates the core business of enterprises, especially in sensitive data sectors such as healthcare and finance, data governance, security, and privacy become unavoidable topics. Amodei pointed out that while achieving a sufficient level of intelligence in models to assist drug development and analyze clinical trial data is a significant technical challenge, often obstructing progress are those issues that "seem a bit 'stupid' to technicians but are crucial from a business perspective."

"Can we ensure that data does not leak? Can we ensure compliance with all necessary regulations?"—Amodei emphasized that if these issues are not addressed, they could stall the entire process for years, delaying the emergence of therapies that could save lives. Therefore, it is essential to govern data correctly.

Ali Ghodsi shared a similar experience from Databricks: their data governance product Unity Catalog has now become the company's top R&D focus and the area with the most personnel investment. He lamented that any business involving data must also confront privacy and governance, making Databricks, in a sense, a privacy and security company.

Amodei added a critically important point: trust. Customers must trust that their partners will do the right thing and are worthy of their data. Both Anthropic and Databricks are committed to building and maintaining a "trustworthy" reputation. The joint offering from the two companies is not just a product portfolio but a solution that customers can trust.

Openness and Open Source: A World Where Capability and Risk Coexist

Addressing the long-standing industry debate over open-source versus closed-source models, Amodei offered his insights. He believes that whether from a business or security perspective, the distinction between open and closed is "somewhat exaggerated."

He pointed out that current models do not yet pose the major risks he worries about (even though he believes these risks are approaching rapidly). At the current stage, both closed-source and open-source models can thrive and promote scientific and industrial advancement in different ways. As models become more powerful, risks (such as national security risks) will emerge. But the key is that the models themselves must be very powerful, so we must address these risks in some way; the approach to managing risks for open-source and closed-source models might differ slightly, but whether a model is open-source has never been the factor he prioritizes most.

Ali Ghodsi followed up, stating that since open-source models may continue to exist (e.g., DeepSeek), it is best if these models are developed "here" (implying in a healthy regulatory and competitive environment). Amodei agreed, saying, "Under all conditions being equal, I think that may be correct."

The Evolution of Reasoning Models and Scaling Laws

Finally, the conversation touched on cutting-edge technology topics in AI. Regarding the relationship between the reasoning models recently emphasized by Anthropic and the classic pre-training scaling laws, Amodei stated that both are focal points for the company.

He mentioned that both reasoning models and regular pre-training show strong returns that follow scaling laws. Anthropic is not the first to enter the market with reasoning models, but they believe they are "doing it correctly." They introduced "hybrid reasoning models," where users can tell the model how long to think, and can toggle reasoning mode on or off using the same set of model weights. "Just like humans," Amodei metaphorically said, "I do not have one brain for simple questions and another brain for difficult ones. I have one brain; I just decide how carefully to think and how much time to spend based on the question."

Regarding whether the scaling laws for reasoning models are different, Amodei believes the basic dynamics of scaling laws remain unchanged, but what is "scaled" may slightly change over time. Whether it is fine-tuning the Transformer architecture or transitioning from human feedback reinforcement learning (RLHF) to the current "reasoning training" (which is essentially still reinforcement learning with different goals), the scaling law itself has not been "defeated"; it continues to be effective. Nowadays, companies often scale multiple aspects simultaneously and stack them together.

On the generalization of reasoning abilities, Amodei pointed out that it depends on what is being scaled. When releasing Claude Sonnet 3.7, Anthropic intentionally reduced investment in narrow, specialized tasks like math competitions, as these tasks lack strong generalization. They focused on broader tasks and found that while generalization is not perfect, it leads to better overall generalization than focusing on math competitions.

At the end of the dialogue, Ali Ghodsi summarized the significance of this collaboration: enterprise users can now directly use the Claude model in Databricks' model services, agent framework, vector search, and RAG features. Both parties are excited about this partnership aimed at accelerating enterprise AI innovation and implementation.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。