Sunil Venkataram
Digital Garden

A collection of notes and ideas at various stages of development. Unlike blog posts, these are living documents that grow and evolve over time.

Evergreen Budding Seedling
View full graph →
Filters:
April 2026
AI Is a Tailwind Not a Moat in Services Roll-Ups
🌰 seedling Apr 28, 2026
The current PE thesis of buying services businesses and "throwing AI in" to improve margins misidentifies the source of advantage. AI toolin…
AI Application Layer Squeeze - Pressure From Both Sides
🌰 seedling Apr 28, 2026
AI application-layer companies face a structural squeeze: LLM providers are building interfaces that absorb app-layer revenue from above, wh…
The Genie Question - Permission to Want
🌰 seedling Apr 28, 2026
"What would you do if you could not fail?" is a question most people never seriously ask themselves. The barrier is not lack of an answer — …
The Utility Curve of Money Flattens Fast
🌰 seedling Apr 28, 2026
Money's impact on life quality steepens at two specific thresholds — an emergency buffer and a freedom buffer — then flattens sharply. The f…
CPU Bottleneck - The Hidden Constraint on AI Scaling
🌰 seedling Apr 27, 2026
GPUs and custom accelerators receive most of the attention in AI infrastructure discussions. CPUs face a quieter but compounding shortage fr…
Phantom GDP - Deflationary Value That Official Statistics Miss
🌰 seedling Apr 27, 2026
When AI enables one person to produce output that previously required a large team, the value created is real but may not appear in GDP stat…
Criteria-Based Grading for Subjective Agent Output
🌰 seedling Apr 23, 2026
Subjective quality — "is this design good?" — becomes gradable when decomposed into concrete criteria with explicit weightings. By defining …
Edge vs Server Model Architecture - Why One DNA Cannot Serve Both
🌰 seedling Apr 23, 2026
Phones and GPUs impose opposite hardware constraints. Edge devices have abundant flash storage but starve for DRAM and battery. Servers offe…
GAN-Inspired Agent Architecture - Generator Evaluator Loops
🌰 seedling Apr 23, 2026
Separating generation from evaluation in agent harnesses produces better outcomes than self-evaluation. Inspired by Generative Adversarial N…
Harness Simplification as Models Improve
🌰 seedling Apr 23, 2026
Every component in an agent harness encodes an assumption about what the model cannot do on its own. Those assumptions go stale as models im…
KV Cache Compression - The Primary Bottleneck in Long-Context Inference
🌰 seedling Apr 23, 2026
The KV cache — storing Key and Value vectors for every processed token so future tokens can attend to them — is the dominant memory bottlene…
Per-Layer Embeddings - Trading Flash for DRAM on Edge Models
🌰 seedling Apr 23, 2026
Standard transformers use a single embedding table that maps tokens into vectors carrying both identity and context through every layer. In …
PM-Engineer Mind-Meld - 80 Percent Overlap Replaces the Handoff Model
🌰 seedling Apr 23, 2026
When shipping timelines compress from months to days, the traditional PM-to-engineering handoff (PRD, spec review, implementation, QA) becom…
Principles Plus Metrics as PM Replacement - Independent Decision-Making Without Bottlenecks
🌰 seedling Apr 23, 2026
Replace PRDs with two lightweight artifacts: written decision principles ("if X, then Y") and a weekly metrics readout shared across the ent…
Research Preview Shipping - Reduce Commitment to Compress Feedback
🌰 seedling Apr 23, 2026
Label early releases as "research previews" to reduce the commitment cost of shipping. Ideas reach users in 1-2 weeks instead of quarters. T…
Sprint Contracts - Negotiated Agent Agreements
🌰 seedling Apr 23, 2026
Before building anything, the generator and evaluator agents negotiate a sprint contract that defines what "done" looks like. The generator …
The 90 to 100 Percent Automation Gap
🌰 seedling Apr 23, 2026
A 95% automation that occasionally fails is not actually an automation — it's a process that requires 100% human monitoring to catch the 5% …
Demand Elasticity Determines Whether Automation Creates or Destroys Jobs
🌰 seedling Apr 21, 2026
Even when AI makes workers more productive and raises wages, employment in that sector can still shrink. The deciding variable is demand ela…
Job Dimensionality - Why Low-Task Jobs Face the Highest Automation Risk
🌰 seedling Apr 21, 2026
The number of distinct tasks in a job (its "dimensionality") predicts displacement risk better than AI exposure scores. Low-dimensional jobs…
Mimetic Desire Undermines AI-Generated Value - The Exclusivity Premium
🌰 seedling Apr 21, 2026
Experimental evidence shows that people value goods more when others are excluded from having them — willingness to pay roughly doubles unde…
O-Ring Production and AI Automation - Why Partial Automation Can Raise Wages
🌰 seedling Apr 21, 2026
When a job's output depends on doing many tasks well together (multiplicative production), automating some tasks can make the remaining huma…
Structural Change Theory Predicts AI Grows the Human-Intensive Economy
🌰 seedling Apr 21, 2026
The economics of structural change — the same framework that explains why agriculture shrank from 40% to 2% of U.S. employment while food pr…
The Relational Sector - Why Human Involvement Becomes the Product
🌰 seedling Apr 21, 2026
As AI automates commodity production, a growing share of the economy shifts toward what Alex Imas calls the relational sector: goods and ser…
Agent-Native Architecture - Five Principles for Building After Code Ends
🌰 seedling Apr 20, 2026
Agent-native architecture rests on five principles: Parity, Granularity, Composability, Emergent Capability, and Improvement Over Time. Toge…
AI Memory Crowding - HBM Eats Consumer Device Budgets
🌰 seedling Apr 20, 2026
AI accelerators need High Bandwidth Memory (HBM), which delivers 20x the bandwidth of commodity DRAM but uses 3-4x the wafer area per bit. M…
EUV Lithography as the Binding Constraint on AI Scaling
🌰 seedling Apr 20, 2026
The hardest bottleneck on AI scaling is not power, not data centers, not capital. It is EUV lithography — the machines that print the chips.…
Workflow Grit as Vertical AI Moat - The Messier the Market the Deeper the Defensibility
🌰 seedling Apr 16, 2026
The best vertical AI markets stay hidden because they're operationally messy — full of exceptions, legacy integrations, compliance requireme…
CUDA Programmability Moat - Why Flexibility Beats Optimization
🌰 seedling Apr 15, 2026
ASICs and TPUs optimize for today's known workload. GPUs with CUDA optimize for tomorrow's unknown workload. AI advances primarily through a…
Different AND Better - The Compound Product Requirement
🌰 seedling Apr 15, 2026
A winning product must be different *and* better, where the "better" matters to the end user in a domain with real economic stakes. Differen…
RL Scaling Follows Pre-Training - The Generalization Inflection Ahead
🌰 seedling Apr 15, 2026
Reinforcement learning now shows the same log-linear improvement pattern that pre-training showed from 2017-2022. Models trained on math com…
The Two-System Brain - Why LLMs Are Data-Hungry and Biological Learners Are Not
🌰 seedling Apr 15, 2026
The brain runs two systems, not one. A general-purpose Learning Subsystem (cortex) builds world models through something like backpropagatio…
Two Exponentials - AI Capability vs Economic Diffusion
🌰 seedling Apr 15, 2026
AI progress produces two overlapping exponentials, not one. The first is model capability — how well the system performs on tasks. The secon…
Absorb Automate Unbundle - Three Phases of Technology Deployment
🌰 seedling Apr 11, 2026
Every new platform technology follows a three-phase deployment pattern: first organizations absorb it (add the new thing to existing workflo…
Abundant vs Scarce After AI - The Bifurcation of Post-Scarcity
🌰 seedling Apr 11, 2026
The most common mental model for "post-AI abundance" treats abundance as an on-off switch: everything becomes cheap, or nothing does. The mo…
AI as Seniority-Biased Technological Change
🌰 seedling Apr 11, 2026
A paradox in the AI productivity literature: controlled experiments consistently show AI helps less-skilled workers the most (an equalizing …
AI Is an Industrial Bubble, Not a Financial One
🌰 seedling Apr 11, 2026
There are two distinct kinds of asset bubbles. A financial bubble produces no durable value; the 2008 GFC is the archetype, pure speculation…
AI Productivity - The Micro-Macro Disconnect
🌰 seedling Apr 11, 2026
Controlled studies consistently show generative AI boosting individual task productivity by 14-55%, yet aggregate economic statistics show l…
AI Stack Value Accrual - Chip, Infra, Intelligence, App
🌱 budding Apr 11, 2026
Value in the AI stack accrues layer by layer, each with its own capture window. The chip layer captured early, the infrastructure layer is c…
Auto Research - Agents as Overnight Experimentation Engines
🌰 seedling Apr 11, 2026
Auto research is the practice of handing an agent a measurable objective, a set of boundaries, and a compute budget — then walking away. The…
Circular Financing as a Bubble Signal in AI
🌰 seedling Apr 11, 2026
A growing share of AI sector revenue flows through deals where the same counterparty appears on both sides of the transaction. Microsoft inv…
Claws - Persistent Looping Agents as App Replacement
🌱 budding Apr 11, 2026
A "claw" is a persistent agent loop that runs continuously, discovers local APIs, stitches them together, and exposes a single natural-langu…
Distribution as the Remaining Moat - Why SaaS Incumbents Aren't Dead
🌰 seedling Apr 11, 2026
The dominant 2026 narrative about SaaS is that incumbents are on a conveyor belt to the guillotine — native AI challengers will replicate th…
Full Stack Down vs Full Stack Up - Two Directions for AI Application Vertical Integration
🌰 seedling Apr 11, 2026
AI application companies are refusing to stay in the application layer. They are vertically integrating in two directions, and the direction…
Grow 10 or Earn 40 - Two Paths for Mature Software Companies
🌰 seedling Apr 11, 2026
In the AI transition, mature software companies face a forced choice between two paths to durable equity value. Path one is to accelerate to…
Hero User Strategy - Native AI's Wedge Into Vertical Software
🌰 seedling Apr 11, 2026
The most repeatable pattern for how a native AI challenger breaks into a vertical dominated by an incumbent control point: target the Hero u…
Horizontal Platform to Vertical Specialization - Enterprise Credibility Pivot
🌱 budding Apr 11, 2026
A horizontal automation platform that "can do anything for anyone" reaches a credibility ceiling with large enterprise buyers. The move that…
Inference Cost Collapse and Frontier Model Margin Expansion
🌱 budding Apr 11, 2026
Frontier model inference costs dropped roughly 90% year-over-year through early 2026, while compute overhead stayed relatively fixed. The co…
Jevons Paradox vs Cognitive Displacement - The Unresolved Tension
🌱 budding Apr 11, 2026
Two serious arguments about the labor impact of AI are currently held by thoughtful people, and they point in opposite directions. Jevons pa…
Neofirms - AI-Native Professional Services as a New Category
🌰 seedling Apr 11, 2026
A "Neofirm" is a new category of professional services firm built from scratch on AI-native foundations. Unlike traditional firms (where hum…
Productivity J-Curve - Why Transformative Technologies Suppress Measured Output Before Harvest
🌰 seedling Apr 11, 2026
When firms adopt general-purpose technologies, measured productivity often falls before it rises. Resources flow into intangible investments…
Run the Business vs Do the Work - The AI-Era Vertical SaaS Shift
🌱 budding Apr 11, 2026
Vertical SaaS historically focused on running the business: practice management, scheduling, billing, CRM, reporting. These tools help owner…
Seat-Based to Token-Based SaaS Pricing Transition
🌱 budding Apr 11, 2026
Software pricing built on "dollars per seat per month" aligns with a world where knowledge workers are the scarce resource and the software …
System of Action - Evolution Beyond System of Record
🌱 budding Apr 11, 2026
Enterprise software has evolved through two major generations: systems of record (databases that track what happened) and systems of engagem…
TAM of Intelligence is Infinite
🌱 budding Apr 11, 2026
Intelligence — cognitive work applied to any problem — is an input to almost every economic activity. Unlike a SaaS category or a hardware m…
The Knowledge Work Cliff - Displacement of the Upper-Middle Class
🌱 budding Apr 11, 2026
A counterintuitive claim about the labor impact of AI: the group at highest risk is not working-class workers or the very wealthy — it's the…
Three Waves of AI Opportunity - Unhobbling, Physical Interface, Robotics
🌰 seedling Apr 11, 2026
AI-era opportunities divide into three temporal waves, each an order of magnitude harder than the last. The first wave is digital unhobbling…
Token Throughput as the New Coding Bottleneck
🌱 budding Apr 11, 2026
When coding agents can handle most of a task's mechanical work, the bottleneck on a developer's output stops being cognitive bandwidth and s…
Agent Harnesses
🌲 evergreen Apr 7, 2026
The orchestration layer between an LLM and the real world. The harness handles tool execution, context management, safety, memory, and sessi…
Spec-Driven Development and AI-Native SDLC - 2026 Analysis
🌰 seedling Apr 3, 2026
The SDLC is being restructured around specs as the primary artifact, context engineering as the core discipline, and agent swarms as the exe…
January 2026
Does AI really think?
🌰 seedling Jan 6, 2026
Skeptics say that LLMs can’t “actually think”, they are just computing probabilities. Proponents respond that that’s all thinking is, and a …
Context Engineering
🌱 budding Jan 4, 2026
Context engineering replaces prompt engineering as the core skill for working with LLMs. It covers four strategies -- write (persist context…
December 2025
Future of Software
🌱 budding Dec 28, 2025
Working hypotheses on how AI reshapes the software industry — from abundance dynamics and new UX paradigms to the decline of switching costs…
What is a Digital Garden?
🌱 budding Dec 27, 2025
A digital garden is a different approach to publishing online content. Unlike a traditional blog with reverse-chronological posts, a garden …
Learning in Public
🌰 seedling Dec 27, 2025
Learning in public means sharing what you're learning as you learn it, rather than waiting until you're an "expert."
January 2025
There’s never been a better time to be a builder
🌰 seedling Jan 22, 2025
With the abundance of software, the TAM of everything goes up
AI + Product Management
🌱 budding Jan 4, 2025
AI compresses the prototype-to-validation cycle by roughly 40x, making product sense, strategic thinking, and taste more important than exec…