2025-10-20 08:00:00

When I worked at Google, I was lucky to collaborate with some of the brightest machine-learning (ML) engineers. They worked on feature engineering. By picking the factors to guide the ML model, their advances could generate tens to hundreds of millions of additional revenue.

Imagine an Excel spreadsheet with hundreds of columns of data. Add two columns, multiply two, divide by another, and subtract a fourth. Each of these is a feature. ML models used features to predict the best ad to show.

It started as a craft, reflecting the vibes of the era. Over time, we’ve mechanized this art into a machine called AutoML that massively accelerates the discovery of the right features.

Today, reinforcement learning (RL) is in the same place as feature engineering 15 years ago.

What is RL? It’s a technique of teaching AI to accomplish goals.

Consider a brave Roomba. It presses into a dirty room.

Then it must make a cleaning plan and execute it. Creating the plan is step 1. To complete the plan, like any good worker, it will reward itself, not with a foosball break, but with some points.

Its reward function might be: +0.1 for each new square foot cleaned, -5 for bumping into a wall, and +100 for returning to its dock with a full dustbin. The tireless vacuum’s behavior is shaped by this simple arithmetic. (NB : I’m simplifying quite a bit here.)

Today, AI can create the plan, but isn’t yet able to develop the reward functions. People do this, much as we developed features 15 years ago.

Will we see an AutoRL? Not for a while. The techniques for RL are still up for debate. Andrej Karpathy highlighted the debate in a recent podcast.

This current wave of AI improvement could hinge on RL success. Today, it’s very much a craft. The potential to automate it—to a degree or fully—will transform the way we build agentic systems.

2025-10-17 08:00:00

It’s 9:30 AM. Do you know where your agent is?

As we enter the era of agentic AI, this is an increasingly important question. ChatGPT launched the consumer use of AI, compressing all of human knowledge into a single model. Now we are asking AI to act on our behalf at work.

Analyze this mortgage statement. Figure out if this person has medical insurance. Answer this inbound call about a 2026 Honda Odyssey minivan. AI agents are doing all of this today. Plus, AI is coding millions of lines every day.

My coding agent, Claude Code, helps me write software. It runs on my computer in my terminal, orange cursor blinking. When I use Notion AI to summarize a document, it’s running in Notion’s cloud. A large enterprise handling customer support queries runs their AI in their cloud, whether it’s Azure or Google or Amazon or other.

Which kind of agents should run in which location : on-device, on my cloud, or on a software-vendor’s cloud?

For a vendor : A vendor must offer all three, but for different reasons. Client-side agents handle privacy-sensitive tasks, including accessing data only available locally, like browser history, local files & clipboard content. They may also take advantage of local resources. Own cloud deployments meet enterprise control & compliance requirements. Hosted cloud provides scalability & convenience for most use cases.

For an enterprise : An enterprise will need to manage all three. Coding agents will very likely remain running on employee devices. Certain high-velocity agents will likely be on hosted clouds. Core software & infrastructure will run on their own cloud. This presents the most complex topology, but it’s a business reality.

For an individual user : Individual users will use predominantly client-side & hosted clouds. Aside from software developers, not many people run their own infrastructure.

These hybrid deployments are then multiplexed by all the communication protocols across agents, whether it’s MCP or A2A or ones coming in the future. The net of it is we are creating new agentic networks within enterprises.

So next time it’s 9:30 AM & your agent goes running, you’ll know whether to check your laptop, your cloud, or your vendor’s data center.

2025-10-16 08:00:00

I asked Claude this morning about the most important news in tech. A few follow-up questions about Salesforce’s Dreamforce, Samsung’s XR headset, & TSMC earnings later, I popped over to ChatGPT Pulse. It suggested some updates about the Fed’s rates policy. Then, I listened to podcasts through Gemini-powered AI transcriptions.

All of my news was filtered through AI.

Fifty years ago, most Americans received their news through one of three major television stations : CBS, ABC & NBC. Edward R. Murrow ended his CBS broadcast with “Good night, & good luck.”

Then cable exploded the channel count from 10 to 100 to 1000, serving every possible interest. Social media fragmented conversations further. But AI is a countervailing force.

What if AI is the new CBS, ABC, & NBC, the new mass-media?

The Economist argues AI is killing the web. Who will visit web pages in a few years? Instead of visiting many links, we ask AIs to summarize & prioritize. The traffic drop is real & immediate.

I see it in my behavior. I’d hazard a guess I visit 10% of the websites I did a year ago.

Why sift through pages and pages, when I can have an answer to my question, with all the relevant context across hundreds of sites? Web browsing feels like the card catalog and Google.com like the Dewey Decimal system - anachronisms of legacy information retrieval taught to schoolchildren.

The networks are back. This time, they are probabilistic.

Good morning. Your briefing is ready.

2025-10-14 08:00:00

Why was the Fivetran-dbt merger all but inevitable?

Fivetran & dbt Labs announced their merger yesterday. The all-stock deal combines two companies into an entity approaching $600 million in ARR.

The beauty of the modern data stack was the explosion in choice. As the cloud exploded onto the scene, the legacy data warehouse was replaced by a collection of fast-moving platforms. In that era, specialization won. The pendulum is now swinging back towards consolidation.

Why? The answer lies in compute economics & revenue scale asymmetry. The table below shows why.

There are three different categories of software within this subset of the ecosystem:

1. Ingestion takes data from software & moves it into a cloud data warehouse. Snowflake acquired Datavolo, which commercializes the open source product Apache NiFi, calling it Openflow. Databricks acquired Arcion for ingestion through change data capture, calling it LakeFlow Connect. Fivetran focuses exclusively on this layer.

2. Transformation means reformatting the data within the cloud data warehouse. Snowflake launched native dbt Projects on Snowflake. Databricks offers Delta Live Tables, native SQL, & Python, plus supports hosted dbt through Databricks Workflows. dbt Core/Cloud is the leading independent transformation tool.

3. Compute revenue is generated when we ask questions of our data. Snowflake remains one of the leaders in structured data analysis with their cloud data warehouse. Databricks’ compute is their own as well.

Here’s the asymmetry in one number. Compute represents 72% of the overall market ($7.6B of $10.6B). As a result of their massive operations, Snowflake & Databricks exert significant gravity within the ecosystem. They have expanded beyond the compute market to impose their presence & capture marginal revenue within customers, pressuring the competitive ecosystem.

That’s not to say these components are independent. George Fraser analyzed Snowflake workloads in September 2024, finding transformation represents 40-45% of total Snowflake compute, which means even at smaller scales, startups can have significant impact on these behemoth businesses.

The Fivetran-dbt merger is an inevitable evolution of a maturing market. Two unicorns must partner to compete against two decacorns. They solve two of the three customer problems. But not yet compute.

One could surmise this consolidation signals the end of the modern data stack. I view it differently. The MDS has succeeded beyond our expectations. The stakes are higher now. Broad platforms, fast growth, & AI-native architectures define the next phase. Expect more consolidation.

2025-10-10 08:00:00

Philip Schmid dropped an astounding figure¹ yesterday about Google’s AI scale : 1,300 trillion tokens per month (1.3 quadrillion - first time I’ve ever used that unit!).

Now that we have three data points on Google’s token processing, we can chart the progress.

In May, Google announced at I/O² they were processing 480 trillion monthly tokens across their surfaces. Two months later in July, they announced³ that number had doubled to 980 trillion. Now, it’s up to 1300 trillion.

The absolute numbers are staggering. But could growth be decelerating?

Between May & July, Google added 250T tokens per month. In the more recent period, that number fell to 107T tokens per month.

This raises more questions than it answers. What could be driving the decreased growth? Some hypotheses :

1. Google may be rate-limiting AI for free users because of unit economics.

2. Google may be limited by data center availability. There may not be enough GPUs to continue to grow at these rates. The company has said it would be capacity constrained through Q4 2025 in earnings calls this year.

3. Google combines internal & external AI token processing. The ratio might have changed.

4. Google may be driving significant efficiencies with algorithmic improvements, better caching, or other advances that reduce the total amount of tokens.

I wasn’t able to find any other comparable time series from neoclouds or hyperscalers to draw broader conclusions. These data points from Google are among the few we can track.

Data center investment is scaling towards $400 billion this year.⁴ Meanwhile, incumbents are striking strategic deals in the tens of billions, raising questions about circular financing & demand sustainability.

This is one of the metrics to track!

2025-10-09 08:00:00

A Founder’s Guide to Product Management

Great products don’t happen by accident. They emerge from disciplined product management practices that balance customer needs, business objectives, & technical constraints. This guide distills lessons from hundreds of successful product organizations.

Table of Contents

1. Finding Product-Market Fit
2. Product Strategy & Vision
3. Roadmap Prioritization
4. User Research & Discovery
5. Product-Led Growth
6. Building Product Teams
7. Product Operations
8. Moving Up-Market

Finding Product-Market Fit

What Product-Market Fit Feels Like

Marc Andreessen’s definition : “Being in a good market with a product that can satisfy that market.”

Signals you have PMF:

• Users get visibly upset when the product is down
• Sales cycles shorten & close rates increase
• Word-of-mouth drives >40% of new customer acquisition
• Retention cohorts flatten (users stick around)
• You’re struggling to keep up with demand

Signals you don’t have PMF:

• High churn rates (>5% monthly for SMB, >20% annually for enterprise)
• Long, unpredictable sales cycles
• Feature requests pull in many different directions
• Customers view product as “nice to have” not “must have”

Product-market fit isn’t binary,it exists on a spectrum & varies by customer segment

The PMF Journey

Stage 1 : Problem Validation

• Do customers have the problem you think they have?
• Is it painful enough they’ll pay to solve it?
• Are existing solutions inadequate?

Stage 2 : Solution Validation

• Does your solution actually solve the problem?
• Is it meaningfully better than alternatives?
• Can you deliver it reliably?

Stage 3 : Market Validation

• Is the market large enough?
• Can you reach customers economically?
• Will they pay enough to build a business?

Measuring Product-Market Fit

Quantitative indicators :

• Sean Ellis Test: Survey users “How would you feel if you could no longer use [product]?” , >40% saying “very disappointed” suggests PMF
• Retention curves: Cohorts should flatten after initial drop-off
• Net Promoter Score (NPS): >50 indicates strong product satisfaction
• Time to value: Users achieving “aha moment” quickly

Qualitative indicators :

• Customer testimonials & case studies
• Unsolicited feature requests (they’re invested enough to suggest improvements)
• Champions advocating for your product internally
• Customers renewing & expanding without sales effort

Product Strategy & Vision

Crafting Product Vision

A compelling product vision :

• Aspirational: Where will the product be in 3-5 years?
• Customer-centric: What will customers be able to achieve?
• Differentiated: How is it uniquely positioned?
• Believable: Ambitious but achievable

Example (Figma): “Make design accessible to everyone” Example (Slack): “Make work life simpler, more pleasant, & more productive”

Product Strategy Framework

Strategy connects vision to execution :

1. Target customers: Who are you building for? (ICP definition)
2. Value proposition: What problem are you solving? Why choose you?
3. Key capabilities: What must the product do exceptionally well?
4. Competitive positioning: Where do you play? How do you win?
5. Success metrics: How will you measure progress?

When to Pivot vs Persevere

Pivot if:

• Core assumptions about customer needs proved wrong
• Market size is smaller than estimated
• Unable to reach customers economically
• Competitive dynamics make winning unlikely

Persevere if:

• Early positive signals even if metrics lag
• Core hypothesis still valid, execution needs refinement
• Market timing is the issue, not product-market fit
• Passionate users exist even if small in number

Roadmap Prioritization

Prioritization Frameworks

RICE Scoring

• Reach: How many users will this impact?
• Impact: How much will it improve their experience?
• Confidence: How sure are you of reach & impact estimates?
• Effort: How many person-months will it take?
• Score: (Reach × Impact × Confidence) / Effort

Value vs Complexity Matrix

• High value, low complexity : Do first
• High value, high complexity : Plan & sequence
• Low value, low complexity : Fill gaps in capacity
• Low value, high complexity : Avoid

Kano Model

• Basic needs: Must-haves that don’t delight but cause dissatisfaction if missing
• Performance needs: More is better (speed, reliability)
• Delight features: Unexpected capabilities that wow users

Saying No

The hardest part of product management :

• Opportunity cost framing: “Yes to X means no to Y”
• Data-driven decisions: Prioritize based on usage, retention impact
• Strategic alignment: Does it advance the vision?
• Customer segmentation: Right for which customers? (Might be wrong for your ICP)

The best product managers aren’t measured by features shipped,they’re measured by features they successfully avoided building

Roadmap Communication

Internal roadmap (detailed):

• Quarterly themes & objectives
• Feature specs with success criteria
• Engineering estimates & dependencies
• Release timeline

External roadmap (directional):

• Strategic themes (not specific features)
• Timeframes in quarters/halves (not dates)
• Subject to change disclaimer
• Customer feedback loops

User Research & Discovery

Continuous Discovery Habits

Product decisions informed by ongoing customer contact :

• Weekly user interviews: PMs talk to 3-5 users per week minimum
• Usage data analysis: Quantitative signals on what users do
• Support ticket review: Qualitative insights on pain points
• Sales call shadowing: Hear customer objections & desires firsthand

Research Methods

Generative research (exploring):

• User interviews : “Tell me about the last time you…”
• Ethnographic studies : Observe users in their environment
• Diary studies : Users log experiences over time

Evaluative research (testing):

• Usability testing : Watch users attempt tasks
• A/B testing : Compare variants quantitatively
• Beta programs : Gather feedback before full launch

Jobs To Be Done Framework

Understand why customers “hire” your product :

• Functional job: What task are they trying to accomplish?
• Emotional job: How do they want to feel?
• Social job: How do they want to be perceived?

Example: Customers don’t buy a CRM to “manage contacts”,they hire it to “close more deals & hit quota so they earn commission & feel successful.”

Avoiding Research Traps

Common mistakes:

• Asking what users want vs observing what they do
• Talking to vocal minority vs representative sample
• Leading questions that bias responses
• Confusing “interesting insights” with actionable insights

Best practices:

• Focus on past behavior (“Tell me about the last time…”) not hypotheticals (“Would you use…”)
• Recruit diverse participant pools (new users, power users, churned users)
• Ask “why” five times to get to root causes
• Synthesize insights into clear themes & decisions

Product-Led Growth

What is Product-Led Growth?

The product itself drives acquisition, activation, & retention :

Core principles:

• Free or trial access: Users experience value before buying
• Self-service onboarding: No sales rep required to get started
• Value before payment: Prove ROI first, monetize later
• Viral loops: Users invite others naturally through product usage

Companies that nailed PLG: Slack, Zoom, Dropbox, Notion, Figma

The PLG Flywheel

1. Awareness: Word-of-mouth, content, community
2. Acquisition: Free tier or trial signup
3. Activation: Reach “aha moment” quickly
4. Engagement: Regular usage & habit formation
5. Monetization: Upgrade to paid tier
6. Advocacy: Users recommend to others (back to awareness)

Designing for PLG

Onboarding:

• Time to value <5 minutes for simple products, <30 minutes for complex
• Progressive disclosure : Don’t overwhelm with all features upfront
• Templates & examples : Help users start with best practices
• Empty state design : First-time user experience is different from power user

Freemium model:

• Free tier useful enough to drive habit formation
• Paid tier unlocks collaboration, scale, or advanced features
• Clear upgrade triggers (usage limits, team size, feature gates)

Viral mechanics:

• Collaborative features : Users invite teammates naturally
• Content sharing : Output shared outside product (e.g., public Notion pages)
• Integrations : Product becomes embedded in workflows

When PLG Works (and Doesn’t)

PLG thrives when:

• Individual users can trial & adopt without approval
• Time to value is fast (minutes to hours, not weeks)
• Product solves clear, frequent pain points
• Low price point (<$100/month initial tier)

PLG struggles when:

• Complex enterprise requirements (security, compliance)
• Long implementation cycles
• Unclear value proposition without sales education
• High-touch customization needed

Building Product Teams

When to Hire Your First PM

Indicators you need dedicated product management :

• Founder spending >20 hours/week on product decisions
• Engineering team unclear on priorities
• Features shipped but not used
• Customer requests overwhelming the team
• Roadmap decisions made ad-hoc, not strategically

Typical timing: 15-30 employees, post product-market fit, scaling phase

Product Team Structure

Early stage (1-2 PMs):

• PMs own entire product surface area
• Close collaboration with founders on strategy
• Generalists handling research, roadmap, GTM

Growth stage (3-10 PMs):

• PMs assigned to product areas (e.g., onboarding, core product, integrations)
• Head of Product or VP Product role emerges
• Specialized roles : Growth PM, Platform PM, Data PM

Scale stage (10+ PMs):

• Product groups aligned to business objectives
• PM levels : APM, PM, Senior PM, Group PM, Director
• Product operations team for process & tools

PM Skills & Competencies

Core competencies:

• Customer empathy: Deep understanding of user needs
• Strategic thinking: Connect features to business outcomes
• Technical fluency: Understand feasibility & tradeoffs
• Communication: Align cross-functional teams
• Data literacy: Make evidence-based decisions

Specialized skills:

• Growth PM: Experimentation, funnel optimization, virality
• Platform PM: APIs, developer experience, ecosystem
• Enterprise PM: Complex workflows, security, compliance
• AI/ML PM: Model performance, training data, bias mitigation

For comprehensive AI implementation strategies & team building, see our AI Implementation Guide.

PM-Engineering Collaboration

Healthy dynamics:

• PMs define problems & success criteria, engineers own solutions
• Engineers have input on roadmap (they see technical opportunities)
• Joint sprint planning & retrospectives
• Clear DRI (directly responsible individual) for each decision

Red flags:

• PMs writing detailed specs without engineering input
• Engineers building without understanding customer context
• Constant scope creep mid-sprint
• Finger-pointing when features don’t succeed

Product Operations

Streamlining Product Processes

As product teams scale, operations become critical :

Product ops responsibilities:

• Tools & systems: Roadmap software, analytics platforms, research tools
• Process design: Sprint planning, launch checklists, PRD templates
• Data & insights: Usage dashboards, metric definitions, experiment infrastructure
• Training & onboarding: New PM ramp-up, skill development

When to hire: 10+ PMs, 100+ employees, or significant process pain

Experimentation Infrastructure

Running effective A/B tests requires :

• Statistical rigor: Proper sample sizes, significance thresholds
• Velocity: Deploy & measure experiments quickly
• Learning culture: Share results, build on insights
• Avoid local maxima: Test big swings, not just button colors

Product Reviews & Operating Cadence

Weekly product reviews:

• Roadmap progress check-ins
• Blockers & cross-team dependencies
• Key metrics review (usage, retention, NPS)
• Upcoming launch readiness

Quarterly planning:

• Reflect on past quarter OKRs
• Set next quarter themes & objectives
• Resource allocation across teams
• Alignment with company strategy

Moving Up-Market

Why Companies Move Up-Market

Drivers:

• Higher revenue per customer
• More predictable, larger contracts
• Lower churn rates
• Better gross margins (fewer customers to support)

Challenges:

• Longer sales cycles
• More complex product requirements
• Higher CAC (field sales teams)
• Slower velocity

Product Changes for Enterprise

Must-haves for enterprise buyers:

• Security & compliance: SOC 2, GDPR, HIPAA certifications
• Single Sign-On (SSO): SAML, Okta integration
• Role-based access control (RBAC): Granular permissions
• Audit logs: Track all user actions
• SLAs & uptime guarantees: 99.9% availability commitments
• Dedicated support: CSM, Slack channels, priority tickets

Advanced capabilities:

• Custom integrations: APIs, webhooks, data export
• Deployment flexibility: On-premise, private cloud options
• Admin controls: Usage dashboards, user provisioning
• Professional services: Onboarding, training, customization

Balancing SMB & Enterprise

Common tension:

• SMB wants simplicity, self-service, fast iteration
• Enterprise wants customization, stability, compliance

Resolution strategies:

• Separate tiers: Different feature sets & service levels
• Platform approach: Core product + enterprise add-ons
• Good/better/best packaging: Gradual capability increase
• Dedicated teams: SMB product team vs enterprise product team

Frequently Asked Questions

How do I find product-market fit?

Start by validating the problem (do customers have this pain?), then validate your solution (does it actually solve the problem?), & finally validate the market (is it large enough?). Key signals of PMF include users getting upset when the product is down, sales cycles shortening, >40% word-of-mouth acquisition, & flat retention cohorts. PMF exists on a spectrum & varies by customer segment.

What are the best product prioritization frameworks?

RICE (Reach × Impact × Confidence / Effort) works well for data-driven teams requiring quantifiable decisions. Value vs Complexity matrix is best for visual prioritization & quick communication. Kano Model excels when focusing on user satisfaction & feature categorization. Choose based on your team’s needs : RICE for rigor, Value/Complexity for speed, Kano for customer-centricity.

How should I structure my product team?

Early stage (1-2 PMs): Generalists owning entire product surface area. Growth stage (3-10 PMs): Assign PMs to product areas with specialized roles emerging (Growth PM, Platform PM). Scale stage (10+ PMs): Product groups aligned to business objectives with clear levels (APM, PM, Senior PM, Group PM). Hire your first PM at 15-30 employees post-PMF when founders spend >20 hours/week on product decisions.

What is product-led growth?

PLG is when the product itself drives acquisition, activation, & retention without sales reps. Users experience value before buying through free trials or freemium tiers. It works best when time to value is fast (<30 minutes), individual users can adopt without approval, & pricing is low (<$100/month). Companies like Slack, Zoom, & Notion exemplify PLG.

When should I hire my first product manager?

Hire when founders spend >20 hours/week on product decisions, engineering is unclear on priorities, features ship but aren’t used, customer requests overwhelm the team, or roadmap decisions happen ad-hoc. Typical timing is 15-30 employees post product-market fit during scaling phase. First PM should be a strong generalist who can handle research, roadmap, & GTM.

How do I prioritize feature requests?

Use frameworks like RICE scoring for objective quantification, Value vs Complexity for visual prioritization, or Kano Model for customer satisfaction focus. Always frame decisions with opportunity costs (“yes to X means no to Y”), use data on usage & retention impact, ensure strategic alignment with vision, & consider customer segmentation (right for which ICP?). The best PMs are measured by features they successfully avoided building.

What is good product strategy?

Good strategy connects vision to execution with five elements : target customers (who you’re building for), value proposition (what problem you’re solving & why choose you), key capabilities (what the product must do exceptionally well), competitive positioning (where you play & how you win), & success metrics (how you measure progress). It should be aspirational yet believable, customer-centric, & differentiated.

How do I move up-market?

Moving up-market requires product changes (SSO, RBAC, audit logs, SLAs, compliance certifications), organizational shifts (field sales team, customer success, professional services), & balancing multiple customer segments. Common tension : SMB wants simplicity while enterprise wants customization. Resolve through separate tiers, platform approaches (core + enterprise add-ons), or dedicated teams for each segment.

Related Guides

Data Strategy Guide

Build data-driven product organizations. Product analytics, experimentation frameworks, metrics selection, & using data to accelerate product-market fit.

AI Implementation Guide

Integrate AI into your product strategy. AI product features, ML implementation, AI team structure, & building AI-powered products users love.

SaaS Strategy Guide

Master SaaS product strategy. Product-led growth, pricing strategies, freemium models, & building viral SaaS products.

Go-to-Market Strategy Guide

Bridge product & GTM. Product-led growth strategies, positioning, messaging, & collaborating with sales & marketing teams.