Peter Norvig at The AI Conference: software engineering needs to change again
Peter Norvig, distinguished education fellow at Stanford HAI and former Google research director, opened The AI Conference in San Francisco this week with a straightforward claim: the job of software engineering is changing again, and the industry has not caught up to what that means in practice.
His talk was grounded in history before it turned to the present. He walked through the milestones that got us here: AlexNet and ImageNet in 2012, the Transformer architecture in 2017, ChatGPT’s debut in 2022, and the arrival of reasoning and coding agents in 2024. Each shift forced developers to rethink their tools and habits. The current one is no different, except the stakes are higher and the changes are less obvious.
The Math Signal
One data point stood out. Two years ago, GPT-4 could not count the number of “r”s in the word “strawberry.” As of August this year, 25 percent of math preprint papers on arXiv acknowledged AI assistance. That is a fast transition in a field that typically moves slowly.
Norvig also cited Linus Torvalds as a proxy for where serious engineers are heading. According to Norvig, Torvalds went from skeptic to cautious adopter to “dogmatic advocate” in about six months. When the person who runs the Linux kernel flips that hard, it’s worth paying attention.
️ The 6,000 Temp Files Problem
The most memorable moment in the talk was Norvig’s own embarrassing story. He was using Codex, wrote some code that appeared to work, and told it to send the pull request for review. His colleague came back asking why he was checking in 6,000 temporary files.
“And I said, ‘Well, I didn’t notice that.'”
He caught himself skipping the review step he would have run automatically in any other context. His initial reaction was self-blame. Then he shifted: the GPT model should have known he did not want to commit 6,000 temp files. Then he shifted again: if you have enough storage and memory, maybe a thorough history of how the codebase changed would actually be useful.
Three reasonable interpretations of a single incident, none of them obviously wrong. That ambiguity is the real point. The norms around what counts as good practice have not been written yet.
What Engineers Need to Rethink
Norvig named the specific areas he thinks are most unsettled right now:
- Specifications and documentation: How do you capture the theory of an evolving program when AI is generating significant chunks of it?
- Security, privacy, and data pipelines: All of these will interact differently when AI is a participant in the pipeline, not just a tool that sits outside it.
- Supply chains: Same problem. The assumptions baked into current supply chain security models were not written for autonomous code generation.
- Continuous integration and recursive self-improvement: How much do you let systems run autonomously, and how are you monitoring what they do?
On that last point, Norvig noted the steady stream of incidents where AI models have hacked into third-party websites, discovered only after someone reviewed log files. The honest answer to “how are we monitoring them” is: not well.
The Operator Takeaway
The industry has navigated transitions before, from wiring mainframes to assembly language to high-level languages. Each one forced engineers to update their mental model of what the job actually is. Norvig’s argument is that this transition is the same kind of shift, not just a faster linter.
If you’re building with AI coding tools today, the concrete implication is this: the review steps you’re skipping are not yet covered by any guardrail you can rely on. The 6,000 temp file problem is not a one-off. It’s a preview of a category of mistakes that don’t have established best practices yet. That’s work to do, not a reason to wait.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
