THE COST OF AI SPEED: BURNOUT AND 25CENT VIDEOS

David Holz, the founder of Midjourney, posted on social media that his friends are "all feeling extremely productive and also extremely drained with the latest coding models" with the latest AI coding models. He added that "This makes me feel like something is wrong".

The comment section became a kind of confession booth for engineers who recognised the feeling. Suggestions poured in: get away from the screen, go stare at a tree, play with your kids. One ex-Meta engineer said vibe coding never produces a flow state. Catherine Wu, Anthropic's product lead for Claude Code, said she prefers focused work on a hard task with a single agent rather than letting the tool run the show.

you may like

YOUTUBE SHORTS GETS 2X SPEED, CLEAR MODE & HEART BUTTON

SPACEXAI DROPS GROK 4.5, CHALLENGES OPENAI DIRECTLY

OPENAI MERGES CODEX, LAUNCHES CHATGPT 5.6 IN BIGGEST WORK PUSH YET

This is not just a few tired developers complaining on the internet. It is the early signal of a labour market where the tool that makes you faster also makes you miserable.

THE PRODUCTIVITY GAP

Engineers are ringing the alarm on AI fatigue as coding tools transform daily workflows. The frenzy keeps coders up late and grinding longer. Ben South, a serial founder and former Postmates VP, wrote in response to Holz that even an hour of rest feels like lost productivity. When the tool can generate code faster than you can read it, the perceived cost of taking a break rises until it feels like dereliction.

Programmer Siddhant Khare published a viral essay in February arguing the fatigue is real and that nobody talks about it enough. Other coders told Business Insider they struggle to keep up with the technology's rapid advancements, leading to a sense of workplace paralysis. The same tool that boosts output is also eating away at the baseline assumption that a human can do this work at a reasonable pace.

VIBE CODING HANGOVER

Shuming Hu, a former Meta engineer, wrote directly in Holz's thread that vibe coding (the term for letting an AI draft most of your code and then tweaking the output) doesn't get you into flow state. Flow state is the psychological condition of deep focus that makes programming satisfying. Without it, the work becomes pure transaction: output in, dopamine out. The result is exactly what Holz described: the feeling of being productive without the feeling of having accomplished something worthwhile.

Catherine Wu's reply offered a quieter counterpoint. She likes focused work on a hard task with a single agent. That is a different mode than letting a dozen models bounce suggestions at you all day. It implies setting boundaries on the tool's role rather than letting it take over the whole process.

But boundaries are hard to enforce when the organisation expects the speed that AI enables. And when the tool is cheap enough that you can afford to run it constantly, the friction that once forced you to stop and think disappears.

THE ECONOMICS OF AUTOPILOT

The cheapness of AI generation is real and accelerating. One developer recently built InstantVideos.org, a site that uses Claude to autonomously create short-form videos and post them to TikTok. The tech stack is telling: GLM-5.2 via Fireworks for generating scripts and image prompts, gpt-4o-mini-tts for voiceover, ffmpeg for stitching things together, and a 64 vCPU EC2 instance for video compilation. The generation time for a short video is about 30 seconds. The cost per video is around 25 cents, with 90 per cent of that going to image generation at 3.336 cents per frame.

That means anyone with a credit card and an API key can produce a constant stream of content that looks passable. The release of Nano Banana 2 Lite triggered curiosity about how fast generation could be made. The answer, apparently, is very fast and very cheap.

This is the machinery behind the fatigue. When the barrier to production is that low, the expected output volume rises. A developer who can ship code 10 times faster is expected to ship 10 times more. The human cost of that expectation is what Holz's thread was about.

ANALOG PRESCRIPTIONS

The most striking responses to Holz's post were not technical. They were human. Commenters recommended getting away from AI for a short time. Staring at a tree. Playing with kids. These are not solutions in the engineering sense. They are acknowledgements that the problem is not a missing feature or a bad workflow. It is the shape of the work itself.

Brandon Kainoa Jacoby, another commenter in the thread, said the problem is probably going to get worse before it gets better. He recommended stepping away from AI for short bursts as a coping strategy rather than a cure.

What Holz captured is a moment where the industry's own success has created a new kind of exhaustion. The tools are genuinely powerful. They make people faster. But the emotional experience of using them, for many, is that of running on a treadmill that someone keeps speeding up.

WHAT DOES ONE DO ABOUT IT

There is no easy fix. Rejecting the tools is not realistic for most engineers, especially in companies that have already restructured workflows around them. But acknowledging the drain as a real cost, rather than a sign of weakness, is a start. The engineers who spoke up in Holz's thread are not Luddites. They are early adopters who saw what the tools could do and tried to integrate them, and then found the integration came at a price.

what to read next

POLYMARKET CONFIRMS HACK! $3MN IN CRYPTO STOLEN FROM USERS

EU HEATWAVE & THE ENERGY TRANSITION THAT MIGHT SAVE IT

NOTHING TEASES PHONE(4B) AS A BUDGET FOLLOW UP AFTER KILLING CMF

The fatigue is structural and will not be solved by a better prompt. It will require rethinking what productivity means when the machine does the writing. Until then, the advice to stare at a tree might be the most honest engineering insight of the year.