Dario Amodei: Relying on Learning to Code as a Fallback Is Jumping into a Fire Pit
Anthropic CEO Dario Amodei says that expecting to learn coding as a career fallback is essentially jumping into a fire pit, and that the golden era of the pure software age has been brought to an end by large models. He believes models can handle 95% of mechanical writing and debugging, with humans only responsible for the 5% of boundary decisions, so productivity could theoretically be amplified 20-fold; however, Anthropic's internal data shows that the deep-thinking ability of people who over-rely on large-model generation is being weakened.
Topics
Body · AI translation
As the man who built Claude and whom Chinese users have strong feelings about, Anthropic CEO Dario Amodei recently poured a whole basin of icy cold water on everyone online who is still desperately learning to code. He said: if you're counting on learning to code to provide a safety net for your career, you're essentially jumping into a fire pit. It's not that programming no longer matters, but that the golden cycle of the pure software era has been driven to the finish line by large models.
Many people think this is just another tech giant manufacturing anxiety to sell AI tools.
But if you place his remarks within the real industry cycle, you'll find he isn't hyping how powerful his own models are, but announcing that the dividend of the pure software era has officially bottomed out.
For the past thirty years, the most solid path to wealth in the entire tech industry has been mastering syntax to translate natural language into machine code.
As long as you were proficient in Java or Python, you could sit in an air-conditioned office building and earn a hefty premium.
But what large models are best at eliminating is precisely this kind of deterministic logic with clear rules and immediate feedback.
The act of purely writing raw code is destined to be the job AI wipes out most thoroughly first.
Over the next decade, the largest-scale wealth and moats are retreating at extremely rapid speed from the lines of code on screens toward the physical world across the board: semiconductors running at the foundation, energy infrastructure that guzzles electricity like water, and solid hardware manufacturing are the truly hardcore assets that can actually bottleneck the era.
And in actual work, the leverage of the division of labor has changed too. Models can run through 95% of the mechanical, grueling writing and debugging for you, leaving you responsible only for the remaining 5% of boundary decisions; in theory, productivity can be amplified 20-fold.
But Anthropic's internal data reveals a more dangerous dark side: people who rely excessively on large-model generation are having their own capacity for deep thinking imperceptibly crippled.
You think you're steering the machine, but in reality you're outsourcing your professional instincts to an algorithm, ultimately reduced to an empty shell who can't even tell whether the code is right or wrong.
When the marginal cost of generating code approaches zero, the ability to spell out code is worthless.
The moats that truly set people apart come down to just two things: First, complex human-interaction domains with a genuine physical, flesh-and-blood feel, such as the deep empathy for patients' emotions and psychological essence in medicine.
Second, the ability to spot logical flaws and false information at a glance within a sea of generated content. In the past, the winners were the people who wrote code fastest and remembered syntax most firmly.
In the future, the winners will surely be those who best understand physical reality and can see at a glance where AI is making things up. https://t.co/fQhlJwpo6n