• Sundar Pichai said Google was still in the early phases of recursively self-improving AI.
• Researchers are working with AI agents to build better models, he said.
• Sergey Brin recently said Google was "starting to get onto that RSI treadmill."

Google's AI models may be helping to build their successors, but CEO Sundar Pichai says the moment of fully self-improving AI is still a while away.

Speaking on an episode of the "Big Technology Podcast" this week, Pichai said that Google researchers were using AI agents to assist them when building models, but caveated that it was still early days.

"We are pretty far from the end state of RSI," said Pichai, referring to recursive self-improvement. "The right way to describe it is, you have RSI loops along with AI researchers in various aspects of model development. I would say you almost have agents helping AI researchers, in some automated way, work together to build these models."

Pichai added that he expected that continuum to keep making progress into next year.

Earlier this month, Google cofounder Sergey Brin said the company's Flash series of models, as well as Gemini 4, were demonstrating how AI was helping to build itself, and that Google was "starting to get onto that RSI treadmill."

"And now our researchers and the model work hand in hand to make the next version of the model," he said.

Brin has been increasingly involved in Google's Gemini efforts and has leveraged his power as cofounder to shift resources to projects he deems most effective, Business Insider previously reported. This has included pushing for more focus on RSI.

A report from the firm Gladstone AI, published earlier this month, said that most researchers interviewed from companies including Anthropic and DeepMind believed RSI will start within the next three to 12 months, and possibly sooner without government intervention.

In another moment on the "Big Technology Podcast," Pichai was asked about how much AI was being used to code internally. The company said in April that 75% of new code inside Google was generated by AI.

"That number in our world today is like meaningless in the sense that every part of code we now touch is influenced by AI deeply," Pichai said. With Gemini 4 Argon, he said, people were going to tackle harder problems. He also said Google was seeing a one-third decrease in code rollbacks, where software is reverted to a previous version because of a bug or system issue.

"And so I think people are writing code more productively, and better code with it," he said of AI tools.

Pichai also uses AI to track Google's model development and internal discussions, he told the "Sources" podcast this week.

Have something to share? Contact this reporter via email at hlangley@businessinsider.com or Signal at hughlangley.01. Use a personal email address and a non-work device; here's our guide to sharing information securely.

• Anthropic platform engineering head Katelyn Lesse said that Claude now writes all her team's code.
• Lesse shared ways that Claude Code has helped the company become more efficient.
• Anthropic nears an initial public offering at a potential $2 trillion valuation.

A top Anthropic engineer says that Claude Code has already wiped out all human coding at one of the AI lab's central teams.

Katelyn Lesse leads the AI startup's platform engineering group, which handles the company's core services, developer tools, and product infrastructure — it's essentially the engine room of Anthropic. Lesse said at Grafana Labs' ObservabilityCON event on Tuesday that among her staff, which tallied 200 as of this summer, the era of human-written code is fully over.

"When it comes to our team and how we're actually building and maintaining our systems, all of our code is written by Claude," Lesse said. "Like, literally all of it."

Anthropic's heavy reliance on Claude Code has helped it remain relatively small, as it approaches an initial public offering that could value it at nearly $2 trillion. The company's revenue growth slowed in the third quarter of this year. Maintaining its lofty valuation may depend on selling more companies on the AI-first approach espoused by Lesse and other Anthropic leaders.

Lesse said that people on her teams are still designing systems and verifications, and then offloading the actual coding to Claude. She said that when the company changed the programming language of a major codebase from Python to Rust, workers developed a way to verify the final product.

"But the grunt work of building the pieces and actually monitoring the pieces and actually making sure that we felt good about it, that was agents all the way down," Lesse said.

Lesse said Anthropic already lets agents fix tests autonomously, with no human interaction, and it's thinking about how staff can get comfortable with agents autonomously making changes to the actual codebase, given the improvements in AI models over the last year.

"As engineering leaders, we're like terrified, and we're like, let's be really careful about this, but I think the line is starting to push," Lesse said.

Anthropic has also embraced Claude Code companywide. It released a report last week saying that as AI agents have worked more inside the company, it has hired for a smaller share of junior technical positions. And in May, Anthropic said that Claude then wrote more than 80% of the entire company's production code.

Have a tip? Contact this reporter via email at scouncil@businessinsider.com, or over text, Signal, Telegram, or WhatsApp at 415-757-8198. Use a personal email address, a non-work WiFi network, and a non-work device; here's our guide to sharing information securely.

• A sign instructing San Francisco gym-goers not to vibe code between sets went viral.
• The sign was an "AI-generated fake," according to the gym.
• Amid tech's vibe-coding obsession and fitness boom, it's no surprise the image has gotten heat.

Don't hog the leg press while prompting Claude.

A screen with the logo of Fitness SF, a Bay Area chain of gyms, went viral among techies on X and Instagram on Monday. The image instructed gym-goers to refrain from writing code or prompting AI during their workouts. "Be present. Lift harder. Ship later," the image reads.

Troy Macfarland, Fitness SF's director of marketing, said, however, that the image was an "AI-generated fake." The gym does have etiquette signs that mention phone usage and being courteous, he wrote, but none mention vibe coding.

"This person altered our actual images displayed on the screen," Macfarland wrote. "I will give an 'A' for creativity to him."

It's not clear whether the image was generated with AI, Photoshop, or some other tool. Ian Brash, whose post with the image got almost 7,000 likes, did not respond to a request for comment.

In San Francisco's AI mania, it's no surprise a sign like that blew up on X. Techies weighed in, cracking their own jokes about how tied they were to their coding agents.

One X user told Fitness SF to give their copywriter a raise. A Cognition engineer said that vibe coding between sets was encouraged at his company's gym.

"Vibe coding at the gym?" one X user wrote. "You're worried about the wrong weights."

One reaction image often appears in the comment sections of posts like this one. It shows a muscled man, holding a giant nuclear bomb in his hands, jumping into the city of San Francisco. One VC posted it with the message: "Even I have limits."

Claude Code and cardio

The sign tapped into two of the buzziest subcultures in tech.

First, there's AI coding. Ever since the launch of Claude Code and OpenAI's Codex, engineers have been tied to their phones (or walking around with half-open laptops) to continue building. Some techies told Business Insider that their love of vibe coding created distance between them and their girlfriends.

Then there's the gym. Tech bros are getting fit, pumping themselves full of peptides and getting shredded at early-morning Equinox sessions. A developer at the gym used to be an anomaly; now, it's the norm.

Some techies seemed skeptical from the start. One X user quoted the post, writing that there was "no way this is real."

Another user responded that they didn't know about that specific sign, but they did know about the underlying issue. "This is a real problem at the Equinox in San Francisco," they wrote.

• Rand Wacker, a Box product executive, wants his potential hires to go beyond being just AI-literate.
• Wacker likes when candidates show him their side projects, especially if they've monetized them.
• He's dipped his toes into building, too, with an Apple TV app that evokes the feel of old-school TV.

Knowing how to use AI is only enough to get a foot in the door for interviews, according to a Box executive.

Rand Wacker, Box's vice president of product strategy, said that for candidates applying for a job at Box, AI literacy is a bare-minimum requirement. Having a fully developed side hustle or project — something he practices himself — is what gives potential hires a leg up in interviews.

Case studies, which he said used to be a staple in Box interviews, are now a relic of the past. Now, he expects candidates to be prototyping out their ideas.

What he really likes to see is candidates showing him a product they've built themselves using AI, even if it has nothing to do with Box's business. He said he's had potential hires show him mobile apps that they've built, gotten onto the App Store, and are monetizing to some degree.

"They're running this little side hustle overall," he said. "And it shows the passion."

He said that people who explore and experiment with AI are the ones who win in the AI world, as they figure out how to use it most effectively in their personal and professional lives.

He practices what he preaches

Wacker, who's spent almost 30 years in the software engineering and product management fields, said he started out as a "computer nerd" and loved playing around with software tools and building things.

As he progressed in his career, he said he had less time to experiment with building, but AI has recently allowed him to test small ideas he didn't have time to explore before vibe coding existed.

He recently made an app to bring back old-school TV because he missed the days when he'd turn it on, pick one of three cable channels, and it would play a random movie from a random point.

"These days, whenever I'm trying to find something to watch on the weekends, it's decision fatigue. Which of the six streaming services do I go to? What are the 20 million titles I scroll through?" he said.

Using Cursor, he built an Apple TV app that would access his entire movie library, pick a random movie, and play it from a random point.

Working on these kinds of projects is a great way to learn the tools available on the market, get ideas for how to use them in other areas, and scratch the itch to create something of his own.

Wacker is one of many executives who have dipped their toes back into building and writing code in recent years. Leaders who haven't written lines of code in decades have started experimenting with AI coding to stay closer to their products and feel like individual contributors.

• Google is injecting more AI into the interview process.
• The company has been testing a new format that tests candidates on their AI fluency.
• The pilot has recently rolled out more broadly across Google, internal docs show.

AI is changing how Google employees work — and how Google hires them.

The company has been testing several AI-related changes to its hiring process for software engineers, according to documents reviewed by Business Insider. These changes, which Google has been rolling out in phases throughout this year, include allowing the use of AI in coding interviews and permitting interviewers to use Gemini to give candidates feedback.

Many engineers are also now expected to show some degree of AI fluency as a bare minimum, the documents show.

In May, Business Insider reported that Google was piloting ways to integrate AI into the interview process with a small handful of orgs. This included a new "code comprehension" interview in which candidates are assessed on AI fluency and prompt engineering. Candidates can also use Google's Gemini AI as an assistant during the interview.

The new formats have since been rolled out more broadly across the company, according to the internal documents. However, a spokesperson said they were still considered to be in the pilot phase for now.

"We're piloting updates to our interview process for software engineers, including evaluating for AI fluency, to be more reflective of how our teams are operating," Brian Ong, vice president of recruiting at Google, told Business Insider in a statement.

Like other companies, Google is adapting its hiring process to keep pace with the huge transformation AI is having on software engineering. As developers increasingly rely on AI coding assistants, hiring companies want to know how competent they are with these tools.

In April, Google said that three-quarters of new code created at the company is generated by AI and reviewed by humans, a trend that has since continued across the tech industry. Google has said that its employees are increasingly using the company's own Gemini models to both improve other Google products and to build the next generation of Gemini.

The documents say that for some roles, including software engineering, AI fluency and AI product integration are now used as measures of a candidate's abilities. A Google spokesperson said that AI fluency is a requirement for all software engineering jobs, while AI product integration is required for most roles outside entry-level positions.

Google is also piloting the use of Gemini for interview transcription and note-taking in some job interviews conducted via the company's videoconferencing platform, Google Meet, one document shows.

Google wants to speed up the feedback process using AI, too. The same document states that Google is working to enable interviewers to use Gemini to draft role-tailored feedback.

As Google builds out this new interview system, employees were told they may be asked to participate in some of the pilots — but were warned against using AI tools to generate their own interviewing "solutions."

"Do not use GenAI or LLMs to create your own standalone solutions or agents for hiring or interviewing use cases," the document states, referring to large language models.

Have something to share? Contact this reporter via email at hlangley@businessinsider.com or Signal at hughlangley.01. Use a personal email address and a non-work device; here's our guide to sharing information securely.