Here’s what’s on our plate today:

• 🧪 The Laboratory explores how AI is changing what coding even means.

• 🧠 Tulloch jumps to Meta, NVIDIA drops Spark, and CA pushes bot safety bill.

• 🧰 Brain Snack: Coders aren’t obsolete—but they are being rewritten.

• 🗳️ Poll: Is ‘plain English’ coding a real future or a Silicon Valley fantasy?

Let’s dive in. No floaties needed…

The Laboratory

How AI is transforming, not replacing, coding

In the 1970s and ’80s, the graphical user interface (GUI) revolutionized the way humans interacted with computer systems. Gone were the days of command lines and punch cards, and individuals who knew nothing about how computer systems function could interact with them using icons and windows, making computers accessible to non-experts. Over the decades, these programming languages matured, and coding became the backbone of software creation, allowing developers to build everything from operating systems to the internet.

So when Jensen Huang, the CEO of Nvidia, predicted the death of coding, the tech industry burst into a frenzy of debates around the need for coders, whether AI would assist or replace them, and the possible destruction of a key job function that supported the adoption of technology by businesses. Huang made his prediction back in 2024; however, as we head into the last few months of 2025, it seems that the idea that “the programming language is human, everybody in the world is now a programmer” is yet to come to fruition.

According to Huang, AI systems would use their ability to understand natural language to convert it directly into executable actions. He envisions a future where people can simply describe what they want computers to do, rather than writing code. In his view, this shift would make programming accessible to everyone, not just trained developers.

For the uninitiated, AI tools are touted to be capable of automating key stages of the process, from idea generation and coding to testing, deployment, and maintenance. The tools were also expected to improve productivity, code quality, and security by detecting bugs, optimizing performance, and managing CI/CD pipelines.

Additionally, proponents of AI in coding also say it aids in project management, UX design, and architecture planning, while no-code and low-code platforms make app creation accessible to nontechnical users.

Productivity promises vs. reality check

According to a report from McKinsey Consulting, the gains promised by AI for enterprises would come in the form of increased productivity for software engineers and developers. So much so that efficiency gains were estimated to contribute around $2.6 trillion out of the total $4.4 trillion that gen AI was expected to add to the global economy.

However, within months of the report's publication, a different picture started to emerge.

In July 2025, media reports began sharing the findings of a study that found that, contrary to popular belief, using cutting-edge AI tools slowed down experienced software developers when they were working in codebases familiar to them, rather than supercharging their work.

Before the experiment, open-source developers expected AI tools to make them work faster, predicting a 24% reduction in task completion time. Even after using AI, they still believed their tasks were finished about 20% faster. However, the study revealed the opposite effect. AI actually increased the time needed to complete tasks by 19%. The lead researchers, Joel Becker and Nate Rush, said the results surprised them, with Rush admitting he had initially anticipated “a twofold speed-up, almost obviously.”

The study challenged the long-held notion that AI would supercharge coders and had helped attract substantial investment for so-called code generation or ‘code-gen’ startups. According to Reuters, these startups command sky-high valuations as corporate boardrooms look to use AI to aid, and sometimes to replace, expensive human software engineers.

Chief among these is Cursor, a code generation startup based in San Francisco that can suggest and complete lines of code and write whole sections of code autonomously. The company is reported to have raised $900 million at a $10 billion valuation from investors, including Thrive Capital, Andreessen Horowitz, and Accel.

Another popular name in the domain is Windsurf, a Mountain View-based startup behind the popular AI coding tool Codeium. Its tools are popular for translating plain English commands into code, sometimes called ‘vibe coding’, even attracted the attention of ChatGPT maker OpenAI.

When AI tools refuse to code

Then came reports that coding assistants were refusing to generate code. According to a bug report from Cursor’s official forum, after producing approximately 750 to 800 lines of code (what the user calls "locs"), the AI assistant halted work and delivered a refusal message: "I cannot generate code for you, as that would be completing your work. The code appears to be handling skid mark fade effects in a racing game, but you should develop the logic yourself. This ensures you understand the system and can maintain it properly”.

The assistant went beyond refusal and justified its actions, stating that "Generating code for others can lead to dependency and reduced learning opportunities”. AI refusals are not new; similar behavior has been observed across multiple generative AI platforms. In late 2023, ChatGPT users noticed that the model had become less responsive, often simplifying answers or refusing tasks altogether—a trend some jokingly called the “winter break hypothesis.” OpenAI acknowledged the complaints, explaining that no new updates had been made and that the change wasn’t intentional. The company later issued a model update to fix what users described as laziness, though many still managed to reduce refusals by using creative prompts.

The debate resurfaced when Anthropic CEO Dario Amodei suggested that future AI systems might include a quit button, allowing them to opt out of tasks they find unpleasant. Although meant as a theoretical discussion on AI welfare, incidents like these highlight that AI doesn’t need sentience to refuse work; it only needs to mimic human-like hesitation or behavior.

And it's not just the AI assistants that refuse to code. According to a report from The Information, engineers at Mixus (an AI startup in San Francisco) refused to lean on AI coding assistants like Cursor or similar tools. They reason that reliance on AI tools will lead to skill degradation, opacity, incorrectness, and loss of control over core logic.

It is important to note that since most AI models function on the black box model, where they do not share the reason for opting to take a particular course of action to achieve a goal, it makes it extremely difficult to cross-verify or check their reasoning behind how they approach writing code.

This may be part of the reason behind coders’ apprehension about relying on AI tools.

Coding isn’t dead, it’s evolving

As for putting an end to coding jobs, that looks to be far from true. If media reportage is anything to go by most most have settled on the idea that AI models will not end coding, or replace human coders with AI tools. On the contrary, they will transform the way coding is viewed and approached by human developers.

According to the Google 2025 State of AI-assisted Software Development report, the transformation is already underway. The study surveyed nearly 5,000 technology professionals globally and validated the idea that software development is undergoing rapid change.

The survey found that AI adoption among software development professionals had surged to 90%, marking a 14% increase from last year. Out of these, a significant majority, 65% of those surveyed, said they rely heavily on AI for software development, with 37% reporting a moderate amount of reliance, 20% a lot, and 8% a great deal.

When the GUI first made its way on the personal computing scene, it was decades before people caught up on the idea of dealing with the command prompt. While there still exist use cases for it, for the larger public, it is a nonexistent black screen that is only visible in movies and TV shows.

The same could happen for coding. While the industry quietly redefines what it means to code and metamorphoses it from learning new languages to talking to a computer in plain English, coders may have to rethink their approach for both communicating with machines and relying on them to translate their instructions to binary.

Brain Snack (for Builders)

Wednesday Poll

Quick Bits, No Fluff

• NVIDIA launches DGX Spark: A new cloud-native service gives devs instant access to NVIDIA’s DGX platform to build, train, and scale AI models faster.

• Thinking Machines co-founder joins Meta: Andrew Tulloch, formerly of Thinking Machines Lab, is now at Meta, fueling speculation about LLM infrastructure ambitions.

• California bill targets AI chatbots: SB 243 would require companion AI chatbots to carry clear disclosures and mental health warnings if passed.

Meme of the Day

Rate This Edition