How AI Is Changing What Work Means

Context

This episode uses a strange but revealing case, Citrini Research working with Anthropic on scorpion venom protein design, to explore a broader shift in what AI changes about work. The point is not just that AI makes existing tasks faster. The deeper claim is that AI lowers the cost of entering adjacent domains, which means companies can test opportunities that used to be far outside their normal expertise. For PMs, that matters because it changes how you think about product boundaries, platform risk, user behavior, and where the real bottlenecks still live.

The Big Idea

AI is not only compressing the time it takes to do familiar work. It is making credible adjacent work cheap enough that teams can now test opportunities far outside their traditional lane.

The Citrini experiment matters because a financial research firm did not just analyze biotech from a distance. It ran a real AI-assisted scientific experiment. But the episode also shows the limit of that shift: digital reasoning can speed up design, while physical validation, regulation, and real-world deployment still move much more slowly.

Key Insights

Adjacent work got cheaper

• What happened: Citrini Research, a firm known for thematic equity investing and macro research, used Claude with Anthropic to design proteins aimed at scorpion neurotoxins.
• Why it matters: The important shift is not that Citrini became a biotech company. The important shift is that AI made a serious first-pass experiment in another field cheap enough to justify doing, instead of only interviewing experts.
• Example: Citrini started from a 35-page prompt, ran 300 sub-agents, generated 318,000 candidates, selected 211 for testing, and found 17 proteins that bound to four different scorpion toxins after outside lab validation.
• PM takeaway: When expertise becomes cheaper to access, teams can validate an adjacent market thesis by doing limited real work, not only by researching it.

Physical bottlenecks still dominate

The Citrini case is also a reminder that faster thinking does not equal faster delivery. The AI system compressed dry-lab design into roughly 64 hours, but the experimental results still took 39 days, and a real treatment would still take 4 to 7 years to reach the world. Citrini's own conclusion was that AI can create a burst of digital progress, then hand a large queue of work to human scientists, labs, and institutions. For PMs, this is a useful pattern well beyond biotech. AI often accelerates the design and planning stages first, while the true constraint stays in validation, compliance, hardware, operations, or change management.

Platforms may become operators

• OpenAI's direction: OpenAI bought Orchestra Bio and is building GPT Rosalind plus Rosalind Workbench to help life science teams explore ideas, design experiments, interpret results, and connect to scientific instruments.
• Anthropic's direction: Anthropic has also set up its own biology lab, which is a different move. It is not only selling intelligence to labs. It is doing some biological work directly.
• Why it matters: The host's view is that AI labs competing with customers is inevitable, but it will not happen all at once or across entire industries in one move.
• PM takeaway: If your product depends on a foundation model provider, treat vendor overlap as a spectrum. A vendor can stay a toolmaker, become a workflow layer, or move into direct execution.

Downloads are not durable usage

Meta's Muse reached 10 million US downloads in 40 days, which puts it in a very small group of apps to grow that quickly. Most of the comparable apps were mobile games, with Disney Plus and Threads as the only major non-gaming examples mentioned. Sensor Tower data also showed 30% retention after 7 days, 23% after 14 days, and 2.5 million daily active users last week, up 15% week over week. Those are strong early signals, but the host argues that the real question is not whether Muse had a big launch. It is whether personal agents become either emotionally rewarding enough to open often, or operationally reliable enough to become infrastructure.

Personal context changes interfaces

Gemini in Android Auto used to behave mostly like voice control for maps and music, with some basic messaging and search. The newer version can pull from personal Google context, such as checking email for flight details while also using Maps traffic data, or finding a place based on a photo. The host is explicitly uncertain about how important this will be. That uncertainty is the insight. AI features matter less because they sound magical, and more because they beat the easier fallback behavior that users already have, which in this case is picking up the phone sitting next to them.

Mental Models & Frameworks

Efficiency versus opportunity

This episode is built around a distinction the host often makes.

• Efficiency: AI helps you do the same work faster, such as drafting emails or summarizing information.
• Opportunity: AI lets you do work that was previously too slow, too expensive, or too far outside your team's skill set.

Use this lens when evaluating an AI initiative. A feature that only saves minutes should be judged on workflow ROI. A feature that opens a new category of work should be judged on whether it changes what the team can credibly attempt.

Orthogonal work test

Citrini's experiment suggests a practical test for adjacent opportunities.

• Start with your normal domain expertise.
• Identify a market thesis you cannot confidently resolve through expert opinion alone.
• Ask whether AI makes a limited, real-world experiment cheap enough to run.
• Use the experiment to improve your core decision, not to permanently become a different company.

This is useful when a PM team is exploring a new segment, workflow, or problem area that sits near the existing business but would previously have required a large up-front commitment.

Toolmaker or operator

The episode presents two distinct AI lab strategies.

• Toolmaker: build models and agents that help domain experts do their job better. OpenAI's biotech push is framed this way.
• Operator: use AI plus your own execution layer to do some of the domain work directly. Anthropic's biology lab points in this direction.

Use this model when assessing partners, platforms, or competitors. A company moving from toolmaker to operator changes the risk profile for anyone building on top of it.

Cognitive speed, physical drag

AI often speeds up the cognitive parts of work first.

• Ideation
• Search
• Design
• Synthesis
• Candidate generation

The physical world often stays slow.

• Lab validation
• Hardware build cycles
• Regulatory approval
• Enterprise procurement
• Organizational rollout

Use this split when planning roadmaps. It helps prevent overpromising based on the fastest part of the system while ignoring the slowest part.

Third-space agent problem

The episode cites a useful way to think about personal agents. Many agent products sit in a middle zone where they are not fun enough to be entertainment and not reliable enough to be invisible infrastructure. That makes habit formation fragile. Use this lens when reviewing consumer AI ideas. Ask whether the product is winning because people enjoy spending time with it, or because it removes enough work that one failure does not break the habit.

Decision Principles

Principle: Test adjacencies cheaply

• When: expert opinions are polarized and the cost of a small real-world experiment has dropped enough to be practical.
• Why: Citrini got a clearer signal by running an AI-assisted biology experiment than by collecting one more round of conflicting opinions about whether AI drug discovery is real.

Principle: Map the real constraint

• When: an AI system produces impressive speed gains in design, drafting, or analysis.
• Why: the fastest step in the workflow is often no longer the limiting step. In the Citrini case, design moved in days, but validation and deployment still moved on biological and institutional timelines.

Principle: Assume vendors can expand

• When: a product strategy depends heavily on a frontier model provider or AI platform.
• Why: the episode shows both tool-provider behavior and direct operator behavior from leading labs. A partner today can become a workflow competitor later.

Principle: Measure habit before hype

• When: a new AI feature shows explosive top-of-funnel growth.
• Why: Muse's download pace is impressive, but the durable question is whether users fit the product into daily life in a way that survives friction, failures, and novelty decay.

Trade-offs & Nuance

Speed versus verification

AI can generate candidates at extraordinary speed, but that does not remove the need for verification. In biotech, that means synthesis, purification, controls, and testing. In product work, it means user validation, security review, operational rollout, and support readiness. The trade-off is not whether to move fast or validate carefully. It is where to spend human attention once generation becomes cheap.

Tooling versus competition

OpenAI's biotech work was presented as enabling existing pharma companies such as Novo Nordisk and Moderna. Anthropic's biology work raised a different concern because it looked more like a lab doing some domain work itself. For PMs, the nuance is that both can be true at once. A platform can create value for partners while gradually internalizing the most strategic parts of the stack.

Context versus setup friction

Adding personal context can make an AI assistant far more useful, as the Gemini in Android Auto examples suggest. But every added integration also increases setup effort, trust requirements, and failure modes. The feature only wins if the value of richer context exceeds the convenience of the old default behavior.

Exposure versus convenience

Muse's sharing issue is a good reminder that convenience features can create serious privacy risk. Researchers were able to enumerate 6,800 shared links and found material such as a loan application, an internal performance report, and an internal revenue review. Fast sharing helps growth, but PMs have to treat AI conversations as potentially sensitive content, not casual social objects.

Hype versus durable value

The host is careful not to overclaim either the scorpion work or the public-opinion impact of scientific wins. Citrini found binders, not a proven anti-venom treatment. And even if AI helps create real medical advances, that may not translate into broad public affection for AI. Big outcomes can still become invisible if they arrive gradually through many small improvements.

Common Mistakes

Mistake: Equating speed with completion

The Citrini case would be easy to misread as proof that AI can now do drug discovery end to end. What it actually shows is a strong acceleration in candidate generation and early exploration. The better interpretation is narrower and more useful: AI can collapse some of the research cycle, while leaving the hardest real-world gates intact.

Mistake: Mistaking installs for fit

A launch like Muse can produce huge download numbers, strong ad momentum, and respectable early retention without answering the hardest product question. The better approach is to ask where the product fits in a user's life once novelty fades. Consumer AI products need either delight, reliability, or both.

Mistake: Trusting obscurity for privacy

Muse's sharing issue appears to have come from making share URLs easy to discover. That is a classic product mistake. If a feature can expose chats, images, or searches, the design should assume that the content is sensitive and protect it with strong access controls, not just hard-to-guess links.

Mistake: Ignoring platform encroachment

Teams often treat model providers as neutral infrastructure. This episode is a reminder that the boundary can move. A frontier lab may start by selling capabilities into an industry, then later decide to own more of the workflow itself. PMs should plan for that possibility before it becomes urgent.

Practical Application

Run one adjacent experiment

Pick one product or market thesis that your team currently studies only through customer interviews, analyst reports, or expert calls. Ask whether an AI-assisted prototype, dataset analysis, or workflow simulation could produce a sharper signal in one to two weeks. The goal is not to launch a new business line. The goal is to learn whether direct experimentation now beats secondhand opinion.

Build a bottleneck map

For one major initiative, split the workflow into two columns.

• AI-accelerated steps: drafting, synthesis, candidate generation, segmentation, analysis.
• Still-slow steps: compliance review, hardware delivery, customer rollout, human approval, support readiness.

Then replan the project around the slow column. This prevents an AI speedup in one area from creating false confidence about overall delivery time.

Review vendor overlap risk

List the model providers, agent platforms, or AI infrastructure vendors your product depends on. For each one, write down three possibilities: they stay a toolmaker, they move one layer up into workflow software, or they become a direct operator in your space. Use that review to decide where you need differentiation, proprietary data, or less dependency.

Redefine launch success metrics

If you are shipping a user-facing AI feature, do not stop at installs, trials, or day-1 curiosity. Add measures that show whether the feature becomes either habitual or indispensable, such as repeat use in a core workflow, successful task completion after a failure, or multi-week retention for a clear job.

Threat-model sharing features

Audit any feature that lets users share AI-generated content, prompts, chats, or uploaded files. Check whether access depends only on secret URLs, whether sensitive business or personal information can appear in the shared artifact, and whether the user clearly understands what becomes public. Treat this as a product review, not only a security review.

Questions to Consider

Adjacent bets worth testing

Where could an AI-assisted experiment let your team test an adjacent market, workflow, or customer problem directly, instead of relying only on interviews and expert opinion?

True constraints in delivery

Which step in your product workflow is now cheap cognitive work because of AI, and which step still depends on slow human validation, regulation, procurement, or operations?

Vendor overlap exposure

If a foundation model provider moved from selling tools into running part of your customer's workflow directly, which part of your product would become most vulnerable?

Habit beyond the launch spike

For any AI feature with fast installs or impressive trial growth, what evidence shows that users see it as either rewarding enough to return to or reliable enough to depend on?

Bottom Line

The real shift is not just that AI makes work faster. It makes some kinds of adjacent work cheap enough to attempt, which changes what teams and companies can credibly explore. For PMs, the opportunity is bigger experimentation and broader product scope, but the discipline is knowing where physical bottlenecks, privacy risks, and platform competition still put hard limits on that ambition.

Resources Mentioned

Case Studies Mentioned

Citrini's scorpion binder experiment

Citrini Research started from an investing question, not a biotech mission. The firm believed life sciences might be undervalued in a world where AI could change drug discovery, but it did not trust expert opinion alone because views ranged from AI has created nothing important to AI will cure everything soon. So it used Claude to design candidate proteins for scorpion toxins, paid outside labs to synthesize and test them, and got early binding results for 17 candidates. The lesson is that AI can make cross-domain experiments economically reasonable for teams that would never have attempted them before.

Muse's early consumer traction

Muse reached 10 million US downloads in 40 days, with large ad support and early retention that compared well with other AI chat products. That makes it a meaningful signal, but not proof of a lasting consumer habit. The more durable lesson is about evaluating personal agents: strong launch growth does not settle whether the product becomes a routine part of life or stalls in a middle zone between novelty and infrastructure.

Anthropic's biology expansion

Anthropic's wet lab matters because it suggests an AI lab might do more than supply models to other companies. Even if the lab's remit is limited, it changes the conversation from AI as horizontal infrastructure to AI as a company that may sometimes execute inside a vertical domain itself. The product lesson is to watch for platform providers moving closer to the end application.

Tools & Products