Lesson from Scaling, The State of the Robotics, and the AI Software Development Stack
Free to think for yourself.
Good morning
In today’s edition, among other things:
Lesson from Scaling
The State of the Robotics Ecosystem
LLMs are a Different Kind of Intelligence
The State of Vibe Coding
AI Software Development Stack
Roadmap: European Resilience
Onwards!
Building a Startup That Scales
I really like Posthog as a product and I’m using it daily. I also really enjoy their content. PostHog grew from 11 people selling nothing to 150 people generating $$$ ARR in five years. They just shared 32 lessons.
Some of them really resonate with me now more then ever. Some are may sound obvious until you realize you’re about to violate them anyway. The gap between knowing and doing is where most companies die.
Most often from self-inflicted complexity.
PostHog learned early: optimism matters more than hard technical skills. In the “pessimism creates damage” sense.
The mechanism is economic. A pessimistic hire doesn’t just underperform—they tax everyone around them. Product discussions take longer. Decisions require more consensus-building. The default shifts from “let’s ship and see” to “let’s de-risk before committing.” That tax compounds across every interaction, every sprint, every weekly planning cycle.
You can measure technical skills in interviews. You can’t measure how someone reacts when the demo breaks, the customer churns, or the competitor launches your roadmap. PostHog’s SuperDay—a full paid day of real work—remains their single best signal for new hires. Being good at interviews correlates weakly with being great at the actual job.
The related trap: relaxing standards for that one hire. You will do this. You will regret it. You will keep re-learning this lesson, hopefully with longer gaps between repetitions.
It is 1,000x easier to say no at the interview stage than to let someone go later.
Saying no takes one conversation. Letting someone go takes months of performance documentation, team morale damage, and knowledge transfer overhead. The person you hire to solve a problem becomes the problem six months later—by then they’re embedded in systems, relationships, and decisions you can’t easily unwind.
Most companies know what good looks like. They abandon it gradually under pressure—fundraising pressure, hiring pressure, roadmap pressure, board pressure. The discipline is in sustaining what works when every incentive says to scale it differently.
You will violate most of these lessons. I did. The question is how quickly you notice and course-correct.
The State of the Robotics Ecosystem
Software ate the world so maybe robots will digest the leftovers. Foundation models spill into physical space more and more. Robotics is crossing the same canyon software just did: from neat lab videos to more durable workflows with ROI. Here’s Insight Partners:
Progress in robotics’ foundation models — driven by advances in deep learning, richer training data from teleoperation and simulation, and cheaper hardware — is closing the gap between research and deployment. In this environment, vertical Robotics-as-a-Service (vRaaS) offers one possible path to commercialization for the next era of robotics, focused on solving specific, economically valuable tasks.
Robotics is standardizing from cute demos into dependable services, and the storyline runs through three arcs: copilots → physical agents → field-proven services.
Before robotics, software “copilots” meant tools that sit in your workflow and suggest actions (code, email, design). In robotics, the analogy holds:
Copilots (assistive autonomy): software that helps humans perceive/decide (e.g., vision that flags defects, path hints for forklifts, scripted routines for arms). A human remains in charge.
Physical agents (task autonomy): embodied systems that plan and execute a task end-to-end under constraints (inspect aisles nightly, clean panes, patrol lots, move totes). A human still supervises, but the robot’s default is to act.
In practice, reliable robots are trained and operated on three complementary streams:
(A) Outsourced/partner data: labeled images, point clouds, trajectories from vendors or public sets. Good for bootstrapping perception; weak on your edge cases.
(B) Simulation ↔ reality loop: modern simulators manufacture rare events (slippery floors, odd lighting) safely; real traces tighten the sim. This is how you practice before touching the site.
(C) Teleoperation & “supervised autonomy”: your own operators intervene live, producing the highest-value action+perception pairs. These events both guarantee today’s SLA and train tomorrow’s policy.
Pretty good chance that now wwo curves crossed:
Commodity hardware improved: better sensors/actuators, cheaper compute at the edge, and standard mechatronics let you aim for “good enough, consistently” instead of “perfect, rarely.”
Generalizable control stacks emerged: perception→planning→control pipelines that port across form factors (wheels, arms, drones), so you reuse brains even when you swap bodies.
This unlocks the business model that matters: robots not as capex objects, but as services with contractual outcomes.
vRaaS (Vertical Robotics-as-a-Service) - is the proposed name for the vertical.
Sustainable deployments of it look like SaaS with physics:
Outcome-tied pricing (square feet cleaned, hours patrolled, items moved, defects found).
Recurring revenue + retention from workflows that never stop (nightly inspection; daily cleaning).
Serviceability built-in: spares, swap-outs, and remote diagnostics designed from day one — otherwise margin dies.
Where this is already working at scale:
Cleaning & hygiene (drones, floor/wall systems) where consistency beats perfection.
Inspection → maintenance (energy, data centers, retail) where finding and fixing issues faster has a direct cash payoff.
Warehousing & logistics (inventory drones, tote movers, AMRs) where small time deltas multiply across thousands of picks.
Security & perimeter (ground+drone patrol with an operations center) where coverage and logging are the “product.”
Here’s the market map:
W'e’re very early in robots being dependable co-workers to the shift has started.
LLMs are a Different Kind of Intelligence
Then Andrej Karpathy speaks, everyone should listen (at 0.8 speed):
The State of Vibe Coding
The above mentioned Andrej Karpathy coined “vibe coding” in February 2025. By August, 92% of U.S. developers were using AI coding tools daily. The term stuck because it named something already happening: building software by describing what you want instead of writing how it works. The capability unlocked demand that economics previously killed.
February 2025: Karpathy defines vibe coding as an AI-led experience “where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.”
March 2025: Garry Tan claims AI-powered teams of under 10 now match what once required 100 engineers. Sundar Pichai starts building webpages with minimal coding.
Late summer 2025: Larry Page experiments with v0. Vibe coding tools cross $100M ARR within months of launch.
August 2025: Google reports 30% of new code is AI-generated before human review. Y Combinator’s current batch shows 25% of startups rely heavily on AI-generated code for core systems.
The speed matters because it reveals the bottleneck that broke. Traditional development required translating business logic into code syntax. That translation took time, specialized knowledge, and coordination overhead. Vibe coding eliminated translation as a constraint. What you could articulate, you could build.
To an extent. Here’s The State of Vibe Coding from v0:
Among vibe coders: 63% are non-developers. Thirty-seven percent are traditional developers.
The non-developer majority includes entrepreneurs with no technical background, marketing teams building customer databases in minutes instead of days, product managers prototyping actual applications instead of wireframes, and designers exploring beyond flat mockups.
What they’re building:
44% generate UI components (forms, layouts, design systems)
20% build sophisticated full-stack applications (e-commerce sites, reservation systems, data platforms)
11% create personal websites and portfolios
The prompt style differs by background. Developers mention specific frameworks: “Do these for me using typescript and react-chart” or “In vue there is v-bind directive, what would be the similar in next.js/react?”
Non-developers use conversational, goal-focused prompts: “I need a reservation system for a restaurant” or “Can you create an app for identifying statistics, trends, and providing performance forecasts for the top MLB players.”
When building drops from six months and $500K to six days and near-zero marginal cost, different projects clear the hurdle. Internal tools that would never justify a sprint suddenly make sense. Customer experiments that required roadmap prioritization ship by Friday.
Timing data shows 36% of sessions happen during workday hours, 34% after hours, 28% on weekends. Productivity correlates with uninterrupted focus time—9-to-5 builders complete more sessions than moonlighters. This indicates vibe coding works better as primary workflow than side-project hack.
High usage in markets with low engineering costs, high revenue in markets where engineering is expensive. Vibe coding tools arbitrage global labor cost differentials. A capability that costs $120K/year in San Francisco now costs $20/month for a SaaS subscription.
Also devs an non-dev work differently with those tools:
Organizations overall report three primary benefits:
Faster iteration cycles. Prototypes ship in hours instead of days. Product teams do more without adding headcount. v0 can slash iteration cycles by up to 50%. The constraint shifts from “how long to build” to “how fast can we decide what to build.”
Individual productivity scaling. Using v0, PMs can design, designers can code, sales reps can build products. The tool bridges the gap between design and code. This democratization sounds positive until you map the downstream effects: who reviews what non-technical users build? What’s the QA process when everyone ships?
Cost savings. Vibe coding requires fewer specialized developers and less maintenance overhead, resulting in fewer bugs and lighter roadmaps. The claimed benefits here are the least substantiated. Fewer bugs is questionable when research shows AI-assisted developers generate 10x more security issues. Lighter roadmaps assumes the bottleneck was implementation capacity, not strategic clarity.
The speed that makes vibe coding attractive also creates its primary constraint. Development cycles compressed from weeks to days eliminate traditional buffer time for security review, code audits, and vulnerability testing. Most non-developers lack security training. Most developers moving at vibe speed skip the checks.
Three vulnerability patterns repeat across vibe-coded applications:
Exposing secrets. AI defaults to client-side code unless explicitly instructed otherwise. API keys, passwords, and tokens get stored client-side, visible to anyone inspecting the webpage source.
Access misconfigurations. AI-generated code creates databases writable by the public, API routes without authentication, sensitive data at publicly accessible URLs. These configurations pass the “does it work” test while failing the “is it secure” test.
Hardcoded credentials. AI embeds privileged credentials for testing directly into apps, sometimes without user knowledge, creating attack vectors for anyone who finds them.
“If you are vibing, you’re not looking at the code.... That is very different from using AI to write code that is reviewed, tested, and maintained over time.”
I think we can see three paths ahead:
Optimistic scenario: Vibe coding goes mainstream in enterprise. Developer roles evolve into system architects and AI orchestrators. Security tooling catches up. Every company builds custom tools fast. Software development costs drop 60-80%.
Conservative scenario: Vibe coding finds its sweet spot in prototyping, internal tools, and low-stakes applications. Traditional coding handles complex systems, safety-critical software, and large-scale enterprise applications. The market bifurcates cleanly.
Likely middle path: By late 2020s, AI-assisted code generation becomes standard—like syntax highlighting, used universally without fanfare. But security incidents force corrections before capabilities plateau. The first major breach traced to vibe-coded vulnerabilities triggers enterprise pullback. Regulation adds friction. The market splits: high-security applications stay traditionally developed, low-stakes tools go fully vibe-coded.
The likely path assumes the market corrects before disaster. That’s optimistic. More probable: multiple breaches happen before meaningful correction. Insurance markets react faster than technology markets. Enterprise adoption reverses in regulated industries first, then cascades.
AI Software Development Stack
Let’s stay for a bit longer in the dev process. Here’s a16z:
What will the AI Coding stack look like? While it’s still early, below is an attempt to show what we are seeing today. Orange boxes are areas where we are AI based tools being built by clusters of start-ups. One example is shown for each category. More examples and additional categories that are orthogonal to the process are listed in the market map below.
Interesting Analysis and Trends
AI, Agents & Systems of Action
Effective Context Engineering for AI Agents LINK
Developer Laws in the AI Era LINK
The Greenfield Strategy for AI-Native Enterprises LINK
Systems of Action: The Next Era of AI LINK
Expert Picks: AI SDR Tools LINK
Startups, Research & Strategy
No Science, No Startups: The Unseen Engine We’re Switching Off LINK
Pattern Breakers: The Heresy of Breakthrough Startups LINK
A Research Toolkit for the Discovery Phase LINK
An Early Map for a Useful Robotics Future LINK
Roadmap for European Resilience LINK
Leadership, Teams & Culture
How Culture Scales (or Doesn’t) LINK
How to Be a Leader When the Vibes Are Off LINK
How to Lead in a Room Full of Experts LINK
Team Dynamics After AI LINK
We Are Entering a 10x World LINK
Business Models, Finance & GTM
The Anatomy of an Enterprise Platform LINK
The State of Wealth in 2025 LINK
Friction Report 2025 LINK
Why Every Software Company Needs a Platform Strategy LINK
AI Will Not Make You Rich LINK
Economics, Policy & Global Context
Evaluating the Impact of AI on the Labor Market LINK
GenAI Shadow IT Secrets LINK
Winning Through the Turns LINK
Roadmap: European Resilience LINK
Meditations
Haruki Murakami (via Shane Parish):
They can’t take it any further. And why not? Because they won’t put in the effort. Because they haven’t had the discipline pounded into them. They’ve been spoiled. They have just enough talent so they’ve been able to play things well without any effort and they’ve had people telling them how great they are from the time they’re little, so hard work looks stupid to them. They’ll take some piece another kid has to work on for three weeks and polish it off in half the time, so the teacher figures they’ve put enough into it and lets them go to the next thing. And they do that in half the time and go on to the next piece. They never find out what it means to be hammered by the teacher; they lose out on a certain element required for character building. It’s a tragedy.
Thank you for your time,
Bartek