AI-assisted programming risks becoming a crutch you can never put down. When the LLM stops working—maintenance, no network, air-gapped environments—anyone who has delegated every line of code to an AI agent suddenly finds themselves helpless. This is what Bengaluru-based developer Ashutosh Rath calls "AI atrophy," and to measure it he built Atrophy, a command-line application that treats coding abilities like an Elo rating and forces developers to reinforce their learning through regular drills.

The idea is straightforward: if AI assistants are silently eroding the ability to code unaided, you need a tool that reveals the decline before a job interview, an outage, or a day without Wi-Fi. Atrophy does this through exercises in five areas—syntax recall, debugging, code reading, API memory, and decomposition—covering Python and JavaScript, with three difficulty levels and seeded generation of fresh variants.

After a baseline exam of about 25 minutes, the user gets a starting rating (1200 for each category) and then runs 5–10 minute drills two or three times a week. The app selects an exercise from the skill neglected the longest, with a soft time limit: exceeding it reduces the points gained. Inactivity doesn’t lower the score but weakens the confidence of the estimate, signaling that the rating may have rusted.

What sets Atrophy apart isn’t gamification, but gap measurement: once a month, the user can perform the same drill with AI assistance, tracking the assisted score separately to understand how much their autonomy is slipping. "It’s not an anti-AI tool," Rath explains, "I built it to measure the gap between what I can do with AI and what I can still do on my own, because that ability can rust without warning."

The silent risk for on-premise deployments

Rath’s insight hits a nerve for those designing self-hosted infrastructures or environments with strict data sovereignty. When developers become accustomed to relying on cloud-based LLMs for writing, debugging, or designing, fundamental skills atrophy. In an on-premise context without connectivity to external services—say, a bare-metal inference cluster in a bank or a healthcare facility—being able to manually fix code that won’t run is a survival requirement, not a luxury.

This isn’t a distant worry: an MIT study found that students writing essays with AI chatbots showed reduced brain activity and poorer fact retention, a form of "shallow encoding" of learning. Applied to programming, this describes developers who no longer remember what they wrote and can’t operate without their agentic companion.

Atrophy runs entirely on the developer’s local machine—no code uploads to the cloud, no mandatory telemetry—making it suitable for teams working under strict GDPR policies or air-gap conditions. It requires no GPU, manages no model: it’s pure mental calisthenics, but it signals an important mindset shift. Just as in the LLM world we talk about quantization to reduce hardware dependence, here the "pruning" targets human skills, kept in shape for when the machine can’t answer.

The app doesn’t work remotely and imposes no schedule, but its value lies in surfacing trends: a rating that steadily drops in one or more categories suggests that AI assistance is eroding our autonomy where we are most fragile. In an industry racing toward automation, reclaiming the ability to work offline is an investment in team resilience and, ultimately, in real control over your own systems.