Back on the Coding Train
I spent the past week getting back into coding. AI coding, specifically. The productivity difference isn't subtle. It's violent.
For years the problem was fragmentation. CEO life meant 25-minute slices instead of the 4-hour uninterrupted blocks I had at Twitter, Airbnb, and in the early Mesosphere days. Context switching killed deep work.
That constraint basically doesn't exist anymore. You bounce between projects, kick off builds, let bots grind in the background, switch contexts without losing momentum. Progress compounds while you're not looking.
Turns out ADHD is suddenly a feature. Holding many parallel contexts and recalling them quickly is exactly what this environment rewards.
A Forced Experiment
I was out with a stomach bug, stuck in bed for most of the week, and ended up with an unusual amount of time to write software. Afterward, I looked at the output and did some back-of-the-envelope analysis.
The results are kind of insane.
Yes, the "value" framing below is hyperbolic (hi LinkedIn Karens), and yes, some of this is exaggerated by definition. But even discounting aggressively, the order of magnitude is hard to ignore.
7-Day Agent Productivity Analysis
Here's what 7 days looked like. This excludes my Bert bot configuration and several unrelated or uncaptured side projects.
Project Velocity
The Scale of Compression
Output by Language
Full Breakdown
| Stat | Value |
|---|---|
| Projects touched | 6 |
| Git commits shipped | 100 |
| Languages used | 5 |
| Code files modified | 2,251 |
| Lines of code (excl. auto-gen) | 270,620 |
| Lines of specs / docs | 20,828 |
| Total meaningful output | 291,450 lines |
| Your time invested | ~56–70 hrs |
| Equivalent manual eng. time | ~30,000 hrs |
| Equivalent engineer-years | ~3.7 years |
| Effective productivity multiplier | ~430× |
| Equivalent eng. cost (manual) | ~$2.9M |
| Equivalent eng. cost (pre-AI tooling) | ~$2.2M |
| Your actual cost (time only) | ~$6K |
| Conservative estimate (20% new code) | ~$500K value |
| Most active project | homestars (47 commits) |
| Most documented project | bot-registry (8,734 lines) |
| Avg commits/day | ~14 |
| Avg LOC/day (you) | ~4,200 |
| Avg LOC/day (industry mid-level) | 75 |
Cost Assumptions
| Metric | Value |
|---|---|
| Average base salary | $145,000/yr |
| Fully loaded cost (benefits, overhead, 1.4x) | $203,000/yr |
| Working hours/year | 2,080 |
| Effective hourly rate | $97.60/hr |
Value Created
| Scenario | Hours Required | Cost at $97.60/hr |
|---|---|---|
| Fully manual | 29,980 | $2,926,048 |
| Pre-AI (with scaffolding) | 22,871 | $2,232,210 |
| What you actually spent | ~56-70 | $5,466-$6,832 |
The Multiplier
| Metric | vs. Manual | vs. Pre-AI |
|---|---|---|
| Time savings | 428-535x | 327-408x |
| Cost savings | $2,919,216 - $2,920,582 | $2,225,378 - $2,226,744 |
| ROI | ~42,700-53,500% | ~32,600-40,800% |
Interpreting This (Before You @ Me)
No, this does not mean I magically outperformed a team of senior engineers in raw quality. Most of this is orchestration, synthesis, and direction-setting. The models already encode decades of prior work.
But here's what does matter:
Pre-AI, this volume of work simply wouldn't have happened. Not in a week. Probably not at all. Too much friction to spin up projects, wire infrastructure, write specs, and keep momentum across contexts.
Even if you haircut this by 80% and assume only a fraction is net-new or production-worthy, the leverage is still extraordinary. We've gone from "coding is labor" to "coding is capital allocation."
Once you internalize that, it's very hard to go back.