// BLOG
Writing
On AI-native engineering, compression, and what survives.
The Market Is Bifurcating
AI-fluent experts earn 44% more per hour. Entry-level dev jobs dropped 21% in 8 months. Junior dev employment is down 20%. The middle is dying. Here's what's actually happening.
170 Apps Leaked User Data. They Were All Built the Same Way.
Lovable CVE. Moltbook API keys. AlterSquare's audit. The same three security holes in every vibe-coded app. This isn't anti-AI. It's pro-engineering.
See the Gap, Supply It: What Survives When AI Commoditizes Everything
Everyone asks 'what moat do I build against AI?' Wrong question. The market doesn't reward entrenchment. It rewards perception. See the gap before anyone else and fill it while they're still arguing about frameworks.
The 70% Problem
Claude Code gets you to 70%. The last 30% is where 90% of the expertise lives. Here's why the barrier to building AI collapsed — and why production is harder than ever.
The Four Forces Keeping Your Company Frozen on AI
94% of knowledge work is AI-feasible. Your company deploys on a third. The problem isn't the technology. It's four forces nobody's talking about.
The Agent Computer: When Hardware Is Commodity, Intelligence Is Product
AMD branded the 'Agent Computer' — a $2,000 box to run AI agents. We built the same thing on a $6.50/month VPS. The hardware was never the bottleneck. The intelligence layer is.
One Model for Everything Is Lazy and Expensive
GPT-4 for everything is the AI equivalent of using a Ferrari to deliver groceries. Here's how we route tasks across 5 models and why it cuts costs by 10x.
Why Your AI Demo Works and Your Production System Doesn't
Your demo impressed the board. Your production system is bleeding money, hallucinating answers, and breaking on edge cases nobody tested. The model isn't the problem.
The Deployment Gap: What Anthropic's Data Actually Says
94% feasible. 33% deployed. Everyone quotes the headline. Almost nobody has read the paper. Here's what the data actually shows, occupation by occupation, and why the gap stays wide.
RAG Fails More Than You Think — Here's Why
RAG sounds simple. In practice, every step can fail in ways that produce confident, plausible, wrong answers. Seven failure modes — bad chunking, embedding mismatch, stale indexes, and the fixes for each.
Anthropic's Own Data: Programming at 74.5% AI Coverage
Anthropic published observed usage data from millions of Claude interactions matched against 800 US occupations. Seven charts that tell you exactly where AI is today. The gap between 'real' and 'slower' is where the opportunity lives.
RAG Explained — How to Give AI a Memory It Doesn't Have
An LLM knows what was in its training data. Nothing else. RAG fixes this — retrieve relevant documents, insert into the prompt, generate a grounded response. The most important pattern in production AI.
What's Commoditizing in AI (And What Never Will)
Every developer with an API key is now an 'AI studio.' Here's the commoditization timeline for AI services, and why knowing the difference between commodity and engineering is the only positioning that matters.
What 10 Tech Titans Agree On About 2030
Altman, Musk, Huang, Nadella, Amodei, Andreessen, Gates, Khosla. They disagree on everything. But they agree on five things. Those five things should change how you think about the next five years.
When One Agent Isn't Enough — How Multi-Agent Systems Work
A single agent with 50 tools is a mess. Multi-agent systems solve this by specialization. Router + specialists, pipelines, debate patterns, and how shared memory keeps agents coordinated.
SaaS Is a Filing Cabinet: Why AI Eats Software From the Inside
The entire history of software from 1960 to 2022 was one move: take a filing cabinet, turn it into a database. Now the filing cabinet can do work. What dies, what survives, and what grows.
LangGraph, CrewAI, Claude Code — How Agent Frameworks Actually Differ
Three dominant approaches to building agents. Graph-based, role-based, and code-execution. Each makes different tradeoffs. Choosing wrong means rewriting everything in three months.
The AI Consulting Market: Layer 1 vs Layer 2
39% of organizations stuck in AI experimentation. 40% of agentic projects cancelled by 2027. Layer 1 consulting is already commodity. Layer 2 — enterprise data, compliance, domain AI — is where the premium lives.
AI Agents Aren't Chatbots — Here's the Difference
Everyone calls their chatbot an 'agent' now. It's not. A chatbot responds to messages. An agent takes actions. The difference is not intelligence — it's architecture. The loop, the tools, the memory, the planning.
The Edge Is You
In 2015 I needed 6 months and a team to build an MVP. In 2026 it takes 12 hours and a prompt. So what's left?
Sora, Kling, Veo, Runway — When to Use Which
Five video generation models, five different strengths. Practical decisions — which model for which job, how to build a production pipeline, and the real costs per clip.
The Barrier Is Gone: What Happens When Everyone Can Build
Karpathy coined 'vibe coding.' A year later he moved on to 'agentic engineering.' 25% of YC startups have 95% AI-generated codebases. The casual phase was a phase. What replaced it is a discipline.
From Static Images to Moving Pictures — How Video AI Actually Works
Image generation creates a single moment. Video generation creates time. A 5-second video is 120 frames that must be spatially AND temporally coherent. Here's why video is 100x harder than images.
The Printing Press Moment: Coding Is Solved, Now What?
Boris Cherny built Claude Code. He hasn't written a line of code by hand since November 2025. He ships 10-30 PRs per day. If coding is solved, what's left? Expertise, speed, distribution, taste.
The Ladder Is Gone. Now What?
Every time a technology killed jobs, we had the same answer: move up. AI broke the ladder. For the first time in industrial history, there's no obvious floor above.
Fine-Tuning Image AI — LoRA, ControlNet, and IP-Adapter
The base model doesn't know your face, your brand, or your product. LoRA teaches it new concepts. ControlNet guides composition. IP-Adapter transfers style. Together, they give you precise control.
49.7% of AI Agents Build Software. The Other 50% Is Wide Open.
Half the AI agent market is one category. The bloodbath is in software engineering. Healthcare: 1%. Legal: 0.9%. Education: 1.8%. 300 vertical AI unicorns are waiting to be built.
What AI-Native Engineering Actually Means
Everyone says they're using AI to code. Most are autocompleting. There's a difference between copilot-assisted and AI-native.
DALL-E, Midjourney, Flux, Imagen — What's Actually Different
Same diffusion mechanism, very different outputs. The three things that actually differ between image generation models: the text encoder, the training data, and the fine-tuning strategy.
The Tool Writes, The Engineer Thinks
A practical guide to working with AI as a senior engineer. What to delegate, what to own, and where the line is.
How AI Generates Images — Start with Noise, End with Art
You type a prompt. Three seconds later, an image exists that has never existed before. No database was searched. The image was generated from pure mathematical noise. Here's how diffusion actually works.
Agent-Native Engineering: The $4K/Month Token Budget
$4,000/month per engineer in API tokens. 20 PRs per day. Hundreds of daily commits. Token spend will exceed engineer salary by end of 2026. This isn't AI-assisted coding — it's a different organizational model.
Your LLM Can Book Flights Now — Here's How Tool Calling Works
An LLM by itself can only generate text. It can't check the weather, query a database, or send an email. Tool calling gives it hands. Here's the mechanism, the agentic loop, and the security rules.
What 'Reasoning' Actually Means in LLMs
OpenAI released o1. Google added 'thinking' to Gemini. Anthropic gave Claude extended thinking. But what does reasoning mean for a system that generates one token at a time? Chain of thought, test-time compute, and the honest assessment.
The Agentic Renaissance: Three Phases of Software's Reinvention
$285 billion wiped from IT services stocks in a single day. Boris Cherny ships 10-30 PRs per day without writing code. Three phases of how software engineering is being reinvented — tool, colleague, team.
Hallucination, Context Windows, and Why ChatGPT Forgets Your Name
LLMs don't make mistakes — they do exactly what they're designed to do. The problem is that 'most probable' and 'correct' aren't the same thing. Here's why the mechanism fails and what to do about it.
How LLMs Actually Work — What Happens When You Hit Send
You type a prompt. Two seconds later, a coherent response appears. What happened? Not thinking. Not understanding. Something far stranger — tokenization, attention, and next-token prediction at extraordinary scale.
250 Million Jobs Will Change. Here's What Replaces Them.
Two-thirds of global GDP is knowledge work. That's $50-70 trillion in human compensation. AI is repricing all of it. Not eliminating. Repricing. The difference matters.
7 Ways Voice AI Fails in Production (And How to Fix Each One)
A voice AI demo is impressive. A voice AI in production is humbling. Seven failure modes that only appear when real people make real phone calls in real environments — and the engineering fixes for each.
Pipecat, LiveKit, or Custom — How Voice AI Gets Built
Three real options exist for building voice AI. Each makes different tradeoffs. Pipecat's frame pipeline, LiveKit's infrastructure model, and what a production Indian healthcare stack actually looks like.
'There Is No Product' — Why the Best Companies Sell Outcomes
Sidu Ponnappa's 342K-view thesis: AI is converting software from asset to inventory. A product is crystallized difficulty. Remove the difficulty, and you remove the product. Where is the line for yours?
The Part of Voice AI Nobody Talks About
The pipeline is commodity. You can build it in a weekend. The orchestration layer — VAD, turn-taking, barge-in, state management — is the actual product. It's also the part nobody writes about.
How Voice AI Actually Works — No Buzzwords
You call a clinic. A voice picks up. It sounds human. No human was involved. Here's how the four-box pipeline actually works — STT, LLM, TTS, and the latency equation that makes it feel real.
The Deployment Gap: Why AI Hasn't Replaced Your Job Yet
Anthropic published real usage data from millions of Claude interactions matched against 800 occupations. The gap between what AI can do and what AI is doing is enormous. That gap is the opportunity.