Elon Musk: The Response to Everyone at Large Tech Companies Using Claude to Write Code is Bad
Elon Musk retweeted a long post from an engineer, replying with just one word: "bad." The poster, who joined a large tech company two weeks ago, stated that specifications, code, testing, product requirements, work orders, closures, and reports are all generated by Claude Code. The team does not like this but is pressured to deliver as much as possible.
He wrote that superiors repeatedly say submitting code is no longer a bottleneck, yet still question why progress is slow. Engineers work 12 to 13 hours a day, primarily interacting with the model and hitting enter. From junior to L7, positions are reduced to the same task: talking to Claude. No one reads the code, fixes defects, or experiences the satisfaction of solving problems.
He does not oppose using models but requests at least time to verify the code flow. The current goal is framed as just getting things online, regardless of the consequences. The post quickly surpassed three million views, with likes and shares amplifying, turning the unnamed large company's workspace into an industry public sample.
Musk did not name the company or provide a comparison to his own factories or xAI's processes. The incremental information he provided marked this situation as bad: it is not that models cannot write code, but that people have withdrawn from reading, testing, and thinking, leaving only management pushing for delivery and hours spent hitting enter.
The poster simultaneously expressed a desire to leave but felt it would be the same elsewhere. The environment is described as creativity giving way to output volume, with possibly no one in the codebase able to explain what each piece of code does.
From a market mechanism perspective, this narrative impacts valuation stories. Buyers continue to bet on the penetration of AI programming tools, cloud vendor tokens, and the efficiency myth of tech stocks; sellers worry that unattended code will increase accident, rework, and regulatory costs. Capital flows towards model invocation and "efficiency" narratives. Beneficiaries are programming agent suppliers; those under pressure are the positions that treat engineer hours as quality assurance.
Source: Public Information
ABAB AI Insight
Musk's response of "bad" aligns with the same trend he is promoting: rewriting engineering organizations with models. The difference is that the new employee describes not experts correcting agents, but an entire hierarchy reduced to hitting the enter key. Once requirements, testing, and work orders are written by the model, the only remaining directive in management is "why is it still slow?" The bottleneck has shifted from writing code to no one daring to stop and read.
The capital path is straightforward. Large companies treat Claude Code as a productivity multiplier, using hours worked and submission counts to prove that AI investments are not wasted. Token bills replace design reviews. The motivation is to deliver more features in financial reports, rather than a system that can be clearly explained by one person. Those who insist on reading everything before merging will be labeled as progress issues.
In similar positions, this is closer to the peak of outsourcing where "no one understands this code," rather than early pair programming. The industry phase is programming agents transitioning from assistants to sole authors, with people remaining in the process only to press the access button for production environments. Yakovenko limits the model to finding bugs, while specifications are still handwritten; here, it is the opposite, with handwriting exiting and output remaining.
The structural change belongs to the organizational version of technological substitution. What is being replaced is not a certain type of function, but reading and responsibility. The mechanism is: once submission is no longer a bottleneck, the organization will continue to push until cognition becomes the bottleneck; once cognition is squeezed out by hours worked, the codebase becomes an unclaimed asset until an accident re-prices it.
ABAB News · Law of Cognition
When submission is no longer a bottleneck, reading will be cut.
Everyone hitting enter causes levels to lose distinction.
Code that no one has read cannot be considered delivered assets.