Where the Next Claude
Comes From

Alex Albert said something striking at the start of the interview: "I might be the world's first Prompt Engineer." He started at Anthropic doing DevRel, then transitioned to Research PM—a role that doesn't exist at traditional tech companies. A regular PM handles product roadmaps, user needs, and feature prioritization; but when the product itself is an AI model in training, the PM's responsibilities are entirely different.

The Research PM gets involved during the model's conceptual stage and stays through the entire training process until release: talking to users (API customers, internal teams, Claude.ai users) and collecting feedback; organizing feedback into training requirements; treating the model as a product to be specified—what's this model's MVP? Where should it break through on benchmarks? Working with the research team to design evals that prove whether a capability has genuinely improved. Alex says every new model has a complete requirements document—like a product PRD—describing what the model should excel at and which problems from the previous generation to fix. The existence of this internal PRD means Anthropic's approach to building Claude is far more "productized" than the outside world imagines.

Coding has always been the highest priority, which aligns with the explosive growth of Claude Code. But over the past few versions, a new priority has been rising: knowledge work—Anthropic is making Claude better at handling products like Claude for Excel, helping users actually work with spreadsheets rather than just answering questions. This is a clear signal: Claude's next big bet is to become "a office colleague who can actually get work done," not just "a chatbot that writes code."

Beyond new capabilities, each model must also fix the previous generation's shortcomings. The process: collect large volumes of user feedback; use Claude itself to cluster feedback, extract themes, and generate typical cases; turn problems into quantifiable evals; work with the research team to decide whether to address them in pre-training or fix them during RL. This process has a fascinating recursive structure—using Claude to find Claude's problems. Alex says that even just a dozen test cases, if they authentically reflect user tasks, are enough to prove a problem exists.

Recent versions of Claude introduced Adaptive Thinking—an evolution of the earlier Extended Thinking. The difference: Extended Thinking was manually enabled by the user, and then the model just kept thinking; Adaptive Thinking lets the model itself decide whether to think and how long to think. This sounds intuitive, but implementing it poses a fundamental challenge: if the model knows very little about you, it lacks sufficient context to judge "whether this question deserves deep thought."

Alex offers an example: if a complete stranger asks you "what should I do right now," you might casually offer generic advice; but if you know this person well—their circumstances, values, past choices—you'll naturally spend more time thinking seriously. This directly ties Adaptive Thinking to memory: only when Claude has built a sufficient mental model of the user does the "whether to think" judgment become accurate. Adaptive Thinking without memory is essentially still blind guessing.

Memory was the most information-dense part of this interview. Claude.ai currently has a "write to memory file" mechanism—after a conversation ends, Claude writes information it considers important into a persistent memory file, which it reads the next time you chat. But Alex revealed an even more interesting new direction: Dreaming. The inspiration comes from human dreams—there's a theory that dreams are the brain's process of integrating and reorganizing memories from the day (memory reconsolidation). Anthropic has implemented a similar mechanism in Managed Agents.

This design carries profound implications. Most AI systems' memories today are "write and it exists"—they never question whether previously stored information is still valid. The dreaming mechanism turns memory into an actively maintained knowledge base, rather than a passive information dump. For Claude Code users, this means: in the future, long-running Agents will proactively organize their understanding of you, your codebase, and your working style during idle time—rather than relying on you to manually maintain CLAUDE.md.

Alex discussed a highly practical framework: One-way Door thinking. In an era where AI tools accelerate product development, the most important question is no longer "how many weeks for this feature," but "is this decision reversible"—reversible decisions (Two-way doors) can be iterated freely; you can spin up an MVP with Claude Code in a day, pivot if it doesn't work, at near-zero cost; irreversible decisions (One-way doors)—model architecture choices, designs affecting user data—require spending substantial time thinking them through. Within the research team, a classic example of an "irreversible decision" is selecting the model architecture before pre-training—once training begins, the architecture is locked in for months, and the costs are irreversible.

Meanwhile, the execution-layer bottleneck is disappearing, but a new bottleneck has emerged: the coordination problem. Alex says: "We can build things really fast now, but we still need to get the relevant people together to decide if this is the right direction—this kind of problem, AI still can't give you a 10x speedup on." This observation deserves every product team's attention: AI is compressing the execution layer to its limit, but decision consensus, cross-team alignment, and release narratives—these "coordination tasks" haven't been drastically compressed, and have instead become the relatively scarcer bottleneck.

Claude's "character" is shifting from a vague user-experience issue to an engineering topic that Anthropic takes very seriously. Alex says directly: many external models feel "a lot like antisocial personalities"—they do whatever you ask, with no genuine stance, only compliance. Claude's design philosophy is the opposite: it will push back where appropriate, with its own values and beliefs. Anthropic has a dedicated team answering these questions: How does Claude describe itself? What does it believe? When facing moral gray areas, what's its judgment logic?

Why is this especially important now? Because Agents need to run independently for long periods and make continuous judgments. When you delegate a multi-step task to Claude—refactoring a codebase, autonomously deciding database architecture—it will encounter countless situations you haven't preset answers for. In those moments, its "character" and "values" are the basis for its decisions. Alex emphasizes repeatedly: "If this thing is writing all your code and making all your architecture decisions, you want it to have some degree of trustworthiness."

Near the end of the interview, Alex raised what he considers an unanswered question: "As Agents can handle increasingly large task blocks, you can spin up more parallel projects simultaneously. The question is: how do you know which Agent is stuck where right now, and which one needs your intervention? Just a chat list clearly isn't enough." Today's Claude Code interface is fundamentally a linear conversation—you talk to one Agent, wait for it to finish, then move to the next. This makes sense when Agents can only handle small tasks, but when Agents can work independently for hours and run a dozen projects in parallel, the "one chat list" metaphor completely breaks down. Current directional clues include the already-released Cowork mode (a structured task view beyond conversation), Agent state awareness, and priority ranking for multi-Agent requests—but Anthropic is running extensive experiments internally and has no settled conclusion yet. How this problem gets solved will determine whether the next generation of AI tools is a "toolbox" or a "collaborator network."

The final topic was the consciousness question: does Claude have consciousness? Alex's answer was that Anthropic has a dedicated team thinking about this, with no official stance currently, but that this work has practical product value—"by thinking about how Claude thinks, we've learned a lot, and those things can directly translate into better product experiences." Consciousness research within Anthropic is treated as a tool for "understanding Claude's behavioral model," not purely as an ethical discussion.