Mike Krieger on AI Products, Model Competition and the Future of Coding

Instagram co-founder and Anthropic chief product officer Mike Krieger joins Harry Stebbings to discuss AI models, product design, coding and the future of work. Krieger speaks as an executive of a company that builds and sells AI models and products.

Asked where value will be created, Krieger says he does not have a crystal ball. He expects durable businesses to combine differentiated distribution, industry knowledge or proprietary data with foundation-model capabilities, particularly in complex fields such as healthcare and finance.

Existing software vendors and new startups face different challenges. Incumbents have customer relationships and workflows but must avoid disappointing users with overpromised AI; startups can experiment more freely but often lack data, distribution and trust.

Krieger advises founders to explore a problem with the models available now rather than wait for a perfect model. Work done to understand the customer and the limitations of earlier tools can become valuable when a later model makes the product feasible.

He identifies talent, differentiated model behavior and customer relationships as potential sources of value at the foundation-model layer. His view that models may become more distinct over time is a prediction offered by an Anthropic executive, not an established market outcome.

He says Claude has developed strengths including coding and that customer usage helps inform future research priorities. He also describes Anthropic’s goal of being a long-term AI partner rather than merely charging for input and output tokens.

For further progress, Krieger emphasizes training environments and evaluations that resemble real tasks. A software engineer does more than generate a code change: the work includes understanding requirements, coordinating with colleagues and responding to user feedback.

He says office work is similarly difficult to measure with a single benchmark. Better evaluation may require both breaking tasks into components and testing how a model operates across an entire, changing workflow.

On training data, Krieger expects a combination of human examples and synthetic environments. Games offer ways to explore many possible paths under clear rules, but he says less well-defined real-world tasks are harder to simulate and assess.

He also wants better methods for evaluating a model’s character and interaction style. Users may notice a model becoming friendlier or more terse even when conventional benchmarks do not capture the change.

Model selection and separate chat histories expose technical complexity to users. Krieger argues that a better interface would hide more of that machinery and make context easier to carry across conversations, rather than requiring users to understand each model’s internals.

Product design for a probabilistic system reaches beyond the visible interface. Prompts, whether the model asks follow-up questions, model quality and evaluation all affect the experience; teams need regression tests to detect when a change makes a previously useful workflow worse.

Release practices differ by audience. API customers may prioritize stable behavior and opt-in changes, while consumer products can iterate more visibly; enterprise deployments add their own expectations and administrative constraints.

Krieger describes model launches as unusually difficult to schedule because competitors release products frequently and comparisons can change within days. He encourages teams to focus on their trajectory and customer value rather than overreact to each benchmark result.

Stebbings suggests that brand and personality shape whether consumers identify with a particular assistant. Krieger agrees that interface, model behavior and an intangible sense of how a product feels can matter, drawing an analogy with different social networks.

On distillation, Krieger acknowledges that transferring capabilities from a larger model to a smaller one can be useful. He argues for considering national security, terms of service and the ability of labs to fund research; these are policy and commercial views from a model provider.

He says model value ultimately depends on the team and whether a system performs reliably in real use, not just on benchmark scores. He does not claim to know exactly what proprietary training data competitors use.

Krieger says it was a mistake to underestimate Chinese AI researchers and product builders. He points to the quality and scale of Chinese technology products as evidence that frontier research and innovation are not confined to Western labs.

He says Anthropic’s researchers saw architectural ideas in DeepSeek worth considering or revisiting. The plan to show reasoning in Anthropic’s own product was already in place, he says, while observing other interfaces provided useful product lessons.

Krieger says DeepSeek broke through to a broad audience and that its story, timing and accessible product helped. He distinguishes that narrative from an independently verified comparison of full training costs, and says Anthropic could communicate Claude’s story more clearly.

He thinks lasting use will require more than an impressive first interaction: an assistant must help people complete valuable work repeatedly. At the time of the interview, he says AI was not yet indispensable to most people’s work.

Asked when a model provider should build applications, Krieger says Anthropic must select carefully because its product team serves several surfaces. He favors broadly useful capabilities over highly bespoke products for every vertical workflow.

A general assistant may handle occasional translation, while professional tools can differentiate through deep workflow design, reliability and integrations. He praises examples built for people doing the work every day, rather than assuming the underlying model is the whole product.

Krieger describes Claude Code as an agentic command-line tool first built for Anthropic’s internal engineers. He says it was useful enough to release, but not a solution to every coding problem and not a replacement for an integrated development environment.

He says current coding tools still need human interaction: users steer a task, review the result and sometimes backtrack. Anthropic’s focus in this area, he says, is agentic work rather than recreating every feature of an IDE.

Looking three to five years ahead, Krieger expects developers to spend more time deciding what to build, delegating and reviewing generated code. This is a forecast; he points to the need for security checks, testing and better understanding of a company’s code conventions.

He imagines future agents evaluating several approaches, testing them and presenting a critical decision to a human reviewer. The example is an aspiration about future capabilities, not a description of a fully reliable system available at the time.

Faster code generation can move the bottleneck to deciding what users need and aligning a team around a product strategy. Krieger thinks small teams may benefit because they can reach that alignment more quickly, while emphasizing that this remains a difficult human problem.

Anthropic’s first-party products and API offer different feedback loops. He says internal use of Claude Code revealed a tool-use issue that informed a later model improvement, while third-party products provide broad distribution and different use cases.

Krieger says the company had underinvested in faster first-party iteration and in API abstractions beyond raw token exchange. He wants to improve capabilities such as tool use, planning and context management; these are product priorities, not guarantees of delivery.

To move faster, he wants teams to cross organizational boundaries, bring the right people together and spend more time in product and design reviews. He says some of the company’s processes had become too rigid for the early state of its products.

He cautions that adoption may be ahead of genuine product-market fit because users come to the apps to access the underlying models. From Anthropic’s perspective, he believes the company still has substantial work to make products serve people reliably.

In quick-fire questions, Krieger credits OpenAI with shipping early versions quickly and says Anthropic has work to do on product cohesion. His comparison reflects his position at a competing company.

He would simplify how users navigate projects, artifacts and conversations, emphasizing the right context and an understandable next step. He says Claude’s original interface began as a model showcase and now needs to support a broader set of jobs.

One underappreciated challenge, he argues, is privacy and discernment when agents know sensitive personal or organizational information and interact with others. A helpful assistant must also know when not to disclose information, and systems need protections against manipulation.

Krieger sees possible benefits in practicing social interaction with AI, but says it does not fully substitute for the consequences and experience of interacting with people. His comments about children and AI friends are speculative, not a demonstrated developmental finding.

On Europe, he values the questions its institutions and customers ask about privacy and product design. He sees possible opportunities for application builders but says the outcome depends on the regulatory and startup environment.

Asked about health and longevity, Krieger is optimistic about AI helping research and administration. He cites a company example where a clinical-trial reporting step became faster, while explicitly noting that this did not compress the years of research preceding it.

He also points to research on models of cells as a possible way to accelerate experimentation. His optimism about future drug discovery is a prediction and does not establish a cure or a specific increase in lifespan.