Google VP of Technology says he’s given up on coding

In his keynote on Wednesday, Benoit Schillings, vice president of Technology at Google DeepMind and...

In his keynote on Wednesday, Benoit Schillings, vice president of Technology at Google DeepMind and formerly CTO at Google X, said he’s finally given up on coding his own software and has handed it off to agents.

Schillings said he was a bit new to DeepMind, having moved over from Google X only 18 months ago, and said the shift was “an interesting formative experience.” He explained that he’d been writing code for 45 years, inspired by the desire to write better Apple games, and said he was “pretty resistant to change” on that score. But that has changed.

“It’s about superhuman syntax generation. The last time I got Gemini to write a function for me, and I looked at it, and I was like, ‘do that better.’ Now it's over. I think that the minutiae of code writing can now be generated. Our time is gone.”

Humans still have a role, but it’s more in application and systems design. Humans are still better than machines at that — for now — but, with about 80% of code in GitHub now machine-generated, the era of the human code writer and checker is drawing to a close.

“I would predict that within one year we'll let Gemini or other models generate the code, and nobody will actually look at it, you know,” he told the crowd. “It's similar to compilers. Who still changes the assembly output of their compiler?”

He added that this shift would also require new programming languages to greatly increase the security of agentic code. They wouldn’t be perfect, he said, but we must improve from languages like Python.

Anthropic Fires Up Its New Engine

The speechwriters at Anthropic must have had a busy night rewriting presentations in light of the U.S. government’s decision on Tuesday to lift the ban on Fable 5 and Mythos 5, and Thariq Shihipar, a member of Claude’s technical staff at Anthropic, was in a boisterous mood for his keynote.

“It's one of those Anthropic models that you're just going to remember,” he enthused.

“The best way to describe Fable is like the map is opening up when you’re playing an RPG. You've been on the tutorial, and now you get to the point where the open world starts and there's so much that you can do and explore.”

This is the latest iteration of Claude Code, from its initial design to its current capabilities, including the ability to work proactively and reduce system prompts. He described the new build as “unhobbling Claude,” and said that developers will face a steep learning curve with the new build.

The full capabilities of the Fable build will be discussed in a separate article in this issue, since Shihipar and Claude’s head of code did a group session later in the day.

Amazon Ups The Perception Game

Antje Barth, who’s on the technical staff at Amazon’s AGI Lab, used her keynote to tout the benefits of the company’s latest "perception agent harness” that’s designed to increase efficiency in end-to-end workflows.

“We just launched the first two pieces of our Perception Agent Harness — open source,” she said. “There's two pieces. It has annotation, which you can use to tell it what you want and the verification part that gives the agent the capability to check its own work.”

She also demonstrated an annotation feature, a Chrome extension that allows users to select elements on the screen and mark changes. It logs human alterations, creates a complete summary, and applies the new rules, thus reducing the need for back-and-forth communication.