Meta CEO Mark Zuckerberg believes artificial intelligence is quickly advancing to the point where it can handle the work typically done by mid-level software developers – potentially within the year.
Speaking on The Joe Rogan Experience podcast, Zuckerberg noted that Meta and other major tech companies are developing AI systems capable of coding at a mid-tier engineer’s level. However, he acknowledged current limitations, such as AI occasionally generating incorrect or misleading code – commonly known as “hallucinations.”
Other tech leaders are equally optimistic. Y Combinator CEO Garry Tan has praised the rise of “vibe coding,” where small teams leverage large language models to build complex apps that once needed large engineering teams.
Shopify CEO Tobi Lütke has gone as far as requiring managers to justify new hires if AI could perform the same tasks more efficiently. Anthropic co-founder Dario Amodei has made a bold prediction: within a year, AI will be capable of writing nearly all code.
At Google, CEO Sundar Pichai recently revealed that over 25% of new code is now AI-generated. Microsoft CEO Satya Nadella reported a similar trend, with a third of the company’s code produced by AI.
Despite the enthusiasm, some experts urge caution. Cambridge University AI researcher Harry Law warns that over-reliance on AI for coding could hinder learning, make debugging harder, and introduce security risks without proper human oversight.
Hearing this makes me quite excited but at the same time, concerned as a CTO who has had the priveledge to oversee massive engineering projects in a number of Big Tech companies. The rapid progress that AI coding assistants are making towards attaining the skill of an intermediate level software engineer is astounding, but the contextual understanding, architectural foresight, and security awareness required for production-grade software still stubbornly holds true to the human domain.
To begin with, estimating that AI could handle mid-tier engineering work “within the year”, Mark Zuckerberg has highlighted the path we’re following. AI tools such as GitHub Copilot, Meta’s AI powered codebases, and Anthropic’s models have proven capable of generating uniquely simple but scaffolding modules and CRUD endpoints. In my experience working with different teams, I’ve witnessed a 20-40% decrease in the time taken for scaffolding due to the usage of these systems. However, hallucinations still remain a real problem, the phenomenon where an LLM confidently outputs syntactically correct but semantically flawed code. From a Big Tech perspective, strong human oversight, end-to-end testing, and continuous integration pipelines are still a must.