Recently, artificial intelligence company Anthropic Highly visible. Not onlyIntroduced the powerful Claude 3.7 Sonnet model!It's still there.Significant progress has been made in financingIn a recent interview, Mike Krieger, Chief Product Officer at Anthropic (and former Instagram co-founder), shared insights into the evolution of the AI industry, product strategy, and future trends.
Balancing Innovation and Trust: The AI Product Launch Approach
Sam Altman has mentioned that one of the great joys of startups is the ability to release products quickly without having to strive for ultimate perfection. However, as the company scales, each release carries a tremendous amount of pressure.
Krieger understands this. He believes Anthropic tries to strike a delicate balance between the more aggressive "move fast, break the mold" strategy of startups and the conservative and slow release pace of large corporations. Especially with millions of users, Anthropic's balance between rapid iteration and user trust has been a key issue.
To accommodate different user groups, Anthropic has explored flexible release methods such as the "opt-in" mechanism. For example, in an API product, where users value predictability and stability, an "opt-in" approach can be used to let users decide whether to try new features. However, this may not be the case for consumer or enterprise products, where users expect a continuously improved and optimized experience.
Krieger admits that Anthropic is still actively exploring the most appropriate release cadence. They want to bring new features to market as quickly as possible to get timely user feedback, but also realize that as the company's profile grows and more people start relying on Anthropic's products to get their jobs done, they can't afford to treat releases as casually as they once did.
Beyond the Model: Building a Moat for AI Products
Krieger emphasized that Anthropic's goal is not just to be a "model provider" but to be an "AI partner" to its customers. This means building deeper, longer-term relationships with customers, not just providing "text-in, text-out" API transactions.
To accomplish this, Anthropic places a high priority on the strategic value of First-Party Product, which Krieger believes not only increases revenue streams, but more importantly accelerates learning, improves modeling capabilities, builds brand loyalty, and creates a stronger competitive moat.
named after Claude Code, as an example, noted that by trying out first-party tools internally, it is possible to get direct feedback on model improvements, thus accelerating the iteration of next-generation models. In addition, first-party products are more likely to build user stickiness and brand loyalty.
However, Krieger also admits that Anthropic still has a lot of room for improvement when it comes to first-party product development. He admits that the company hasn't invested enough in its first-party products, resulting in slower iterations, which has somewhat affected Anthropic's competitiveness in the market.
Competing with Differentiation: Opportunities and Challenges for AI Startups
When it comes to the opportunities and challenges of AI startups, Krieger believes that the most valuable areas will be those with differentiated market strategies and unique knowledge of specific industries or specialized data. He singled out areas such as finance, law, and healthcare as areas whose complexity and specialization provide opportunities for AI startups to build long-term competitive advantage.
The key to AI and product design, Krieger noted, is the delicate balance between demonstrating a vision of the future and leveraging the model's current capabilities. Startups can "overpromise" in order to attract early adopters, but SaaS verticals with mature products and users need to be more careful not to undermine user trust with inadequate AI capabilities.
He also emphasized that startups should "build products for future models". He mentioned that many startups' products didn't really take off until Claude 3.5 Sonnet or similar breakthrough models came along. This means that startups need to actively explore the field, be sensitive to the limitations of current models, and actively experiment with next-generation models.
Insights from DeepSeek: multiple thoughts on technology, marketing and product
Krieger also spoke about his interest in DeepSeek The view of DeepSeek. He believes that the emergence of DeepSeek has made people realize that China's technical strength in the field of AI should not be underestimated.
He noted that underestimating or continuing to underestimate China's capabilities in AI would be a huge mistake. He cited a series of parallel startups that emerged after Instagram was blocked in China as an example, noting that these products are often of high quality, show a lot of creative thinking, and have also achieved large-scale adoption.
As for the rise of DeepSeek, Krieger believes that there are both technical and market-level factors. On the technical level, DeepSeek has done some things that Anthropic should learn from and think about. But in terms of market strategy and market position, DeepSeek's influence on Anthropic is relatively limited.
Krieger singled out DeepSeek's marketing success. He attributed DeepSeek's rise from obscurity to being better known than Claude in many circles to the current world situation and the "DeepSeek is cheaper" narrative. He admits that Anthropic hasn't done enough to tell Claude's story to the public, and hasn't done enough to show Claude's uniqueness.
The rise of DeepSeek also made Krieger realize that Anthropic should be quicker to bring ideas to market rather than focusing too much on how perfect every detail is. He believes that sometimes the novelty of the experience is valuable in and of itself.
Additionally, Krieger observed that DeepSeek's usage and retention rates are high in emerging markets, but not in Western markets. He suggests that this may have something to do with DeepSeek's user profiling and marketing strategy in different markets. He noted that for both DeepSeek and Anthropic, the key to staying competitive will be whoever is the first to realize deeper applications of AI in work and life that are sustainable over time.
Modeling capabilities and user experience: the core of AI product design
Krieger emphasizes that there is a strong correlation between model quality and product user experience (UX). He believes that to be a good UX designer, you must also consider model quality.
He noted that today's designers, product managers, and especially engineers need to think about how to design scaffolding and products around a fundamentally uncertain system. This means that model quality, cue word engineering, and all the other back-end stuff, becomes part of the product design and can have a direct impact on the product.
Krieger believes that in the future, users won't need to choose their own models. He uses the term "abstraction leakage" to describe a design flaw in most current AI products. He points out that users need to choose models, understand how they work, and engineer cues, all of which should not happen. He hopes that future AI products will "make cue engineering completely transparent to the user," allowing models to clarify user needs through dialog, rather than allowing users to distinguish who is a good cue engineer.
Code Generation and Software Development: The Changing Face of AI
When it comes to the use of AI in code generation and software development, Krieger believes that the core value of Claude Code is to improve the efficiency of the development process, not to replace IDEs.
He noted that Claude Code is very good at handling tasks that require intelligent collaboration between different parts of the process, such as updating the back-end, creating the front-end, and submitting translations. He believes that there is an intermediate role between IDEs and fully autonomous AI, namely AI intelligences.
Krieger believes that the role of software developers will change significantly in the future. They will need to acquire interdisciplinary skills and become "multi-faceted," knowing both the product and the technology. In addition, the code review will also change, software developers will change from the main code writer, to the main task delegator and code reviewer.
Generalization and Specialization: The Way Forward for AI Products
When talking about Anthropic's product development direction, Krieger emphasized the importance of versatility. He believes that even if Anthropic chooses a specific target user group or vertical, the products they build should be generic enough to support multiple application scenarios at the underlying architectural level.
However, Krieger also acknowledges that specialized workflow knowledge is critical to building products with a long-term competitive advantage. He cites the example of professional translators, who may need features specific to their translation workflow.
He sees tremendous value in AI for professional use cases and the workflows that are unlocked as a result. But on the consumer or even light professional (prosumer) side, the models are good enough from a basic AI product perspective.
Data, Algorithms, and Evaluation: Key Elements of AI Development
When it comes to key elements of AI development, Krieger believes that improving the environment in which models are trained to better reflect the complex tasks of the real world is one of the biggest challenges today.
He pointed out that even in the field of software engineering, the job of a software engineer is not just to write code, but also to understand requirements, develop schedules, and collaborate with teams. There is no suitable assessment methodology that models these complex workflows well.
Turning to the issue of data, Krieger argued that a mix of human and synthetic data is needed in order to improve models. He noted that good human data can be used to guide models, while synthetic data allows models to explore and learn in a variety of environments.
He also mentioned the importance of the "vibes" of the model. He argued that the "feel" of a model is a very subjective, human-like aspect that is difficult to assess quantitatively. Therefore, it is important to have both data on these soft skills and methods to assess them.
Open source, distillation and commercialization: hot topics in the AI industry
Addressing hot industry topics like open source and distillation, Krieger argued that distillation technology is not necessary to unlock AI capabilities, and that it raises other issues, including national security and terms of service concerns.
He noted that in order for technological advances to continue at the current rate and to be sustainable in the long term, the Laboratory needed to be able to commercialize all of its training and innovations. He believes that finding the right business model is critical.
In response to Llama's release, Krieger argued that this doesn't mean that the model itself has no value; all the value is in the data. He noted that the value is in how good the team is, whether they have the underlying data they need, and how useful the model is in real-world use cases.
AI's future outlook: intelligent guides and human partners
In looking at the future of AI, Krieger makes the thought-provoking observation that AI will evolve beyond being a "tool" or an "assistant" to being an "intelligent guide". Krieger makes the thought-provoking observation that AI will evolve beyond "tool" or "assistant" to "intelligent guide.
In his opinion, AI will proactively insight users' needs, guide their direction, assist in decision-making, and become a key partner for them to achieve their most important goals. The AI products of the future will no longer just be "I can ask questions and occasionally make suggestions", but will be able to provide users with unique value, help them save time, improve efficiency and become a better version of themselves.
Krieger also spoke about the potential of AI to extend life and human longevity. He believes that AI can accelerate the process of drug discovery and clinical trials, bringing new hope for the treatment of various diseases. He is very optimistic about this.
Finally, Krieger emphasized the importance of "discernment" and privacy in the development of AI. He pointed out that as models become more powerful, they also become more knowledgeable and may have access to all kinds of private or sensitive information. He argued that it will be a huge challenge for models to help while at the same time protecting user privacy and data security.
English 
简体中文 
日本語 
Deutsch (Sie) 
Português do Brasil