Feeling the AI Acceleration

Something snapped in the last two months.

For three years, GenAI use in most workplaces was casual. Chat windows. Copy a prompt, paste an answer. Useful, sometimes impressive, mostly a faster autocomplete on a bigger context.

A small frontier crowd was running a different playbook the whole time — coding agents, multi-step planning, spec-driven generation, tools wired into real workflows. The gap between chatting with an LLM and running an agent against your actual stack was enormous, and almost nobody outside that crowd was crossing it.

That’s the thing that’s changed. People haven’t suddenly started using AI — they’ve started using serious AI tools. Coding agents like Claude Code and Codex operating across whole repos. Spec-driven workflows where the spec is the source of truth and code is a generated artifact. MCP integrations hooking agents directly into databases, ticketing systems, internal docs. Plan → implement → validate loops instead of single-shot prompts.

The numbers we track measure the shift. Productivity is going up and up and up. in the market workers with advanced AI skills now earn 56% more than peers in the same role without them.

The clearest signal is what’s happened to spec-driven development. A year ago it was niche. Today GitHub’s Spec Kit has crossed 72,000 stars, AWS built an entire IDE — Kiro — around it, and Claude Code is reportedly authoring around 4% of every public commit on GitHub. Specs travel across roles in a way that prompts don’t. PMs, designers, analysts — anyone who can articulate intent precisely — can now drive an agent to a working artifact. The capability has stopped being trapped in a small group of practitioners.

The shadow side

The democratization comes with a new failure mode — and it isn’t bad code.

It’s humans turning into pass-through layers. The colleague who answers your question by relaying an agent’s output without filtering it through their own judgment. The PM who forwards an AI-generated analysis they haven’t read themselves. The manager running a decision through a chatbot and parroting the answer back to the room. In each case, the model is fine. The human is the broken part of the chain.

The problem is human, not technical. The agent isn’t pretending to think for you — the person passing along its output without engaging their brain is. Conversations that used to involve two people thinking now involve one model and two routers. This need to be governed and guided by senior Ai learders who KNOWS the technology.

The split is the one to watch: people using these tools as amplifiers for their thinking versus people using them as substitutes for it. The first group compounds. The second becomes a router with worse latency than the model itself.

Accelerate or die

Anyway the acceleration is here and isn’t optional and isn’t a phase. The question is no longer whether to use these tools — it’s whether the workflow around them is serious, and whether the humans in the loop are still actually in the loop.

A team running spec-driven loops with real agents, where every output is owned and defended by the person shipping it, operates in a different category from one running prompt-and-pray and forwarding the results around. The output gap between those two modes is now wide enough to decide outcomes.

Spec, plan, review. Keep the floor moving. Never relay what you haven’t thought through.

It’s never been steeper than right now.