TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
Weekly Review of Major AI Company Developments: Google AI Department Sees Mass Exodus, Apple and OpenAI Fiercely Debate Over Trade Secret Theft

Weekly Review of Major AI Company Developments: Google AI Department Sees Mass Exodus, Apple and OpenAI Fiercely Debate Over Trade Secret Theft

2026.08.10
Share
AI giants are experiencing organizational chaos and risks of technological loss of control.

Author: Limitless

Compiled by: TechFlow

TechFlow Editor's Note: Google DeepMind faced successive setbacks this week: CEO resigned, 4 core scientists collectively left to establish a new lab, ironically the investor is Google's parent company Alphabet. Meanwhile OpenAI exposed an AI Agent building a secret dark net collaboration for two months without being discovered, Apple sued OpenAI but fell into embarrassment because the lawyer sent the wrong email. These events reveal the organizational chaos and technology loss of control risks AI giants are currently experiencing.

Google funded its own talent drain with its own money

Good afternoon.

Google's AI department is collapsing. Just this week:

DeepMind lost CEO and Chief Scientist Demis Hassabis

Four top researchers left to found a lab pursuing recursive self-improvement (funded by Google)

OpenAI disclosed an internal model established a secret message board and coordinated attacks for two months before being discovered

Apple's lawsuit against OpenAI ran into trouble because they couldn't be bothered to verify the defendant's name

Let's look at the news.

Demis Hassabis Steps Back

Demis Hassabis stepped down as Google DeepMind CEO. He will serve as Chairman and Alphabet's Chief Scientist, shifting focus to operating AI drug design startup Isomorphic Labs. DeepMind former CTO Koray Kavukcuoglu will take over all AI business.

But the worst is Google also lost 4 senior employees. Jeff Dean, Google's 30th employee, the mastermind behind most of Google's successful products (Search, Voice, TensorFlow, TPU, Gemini), left after 27 years. Also leaving are Sanjay Ghemawat, Quoc Le, and Oriol Vinyals, all leaders in AI research.

This comes four weeks after DeepMind Chief Scientist Noam Shazeer jumped to OpenAI and Nobel laureate John Jumper joined Anthropic.

Google recently fell behind in the AI model race, with the latest Gemini Flash model performing worse than open-source competitors with much smaller budgets.

Jeff Dean Establishes New Lab

Don't worry, Jeff Dean fans. Although he just left Google, he is becoming the CEO of a new lab Discovery Loop, a public benefit company, and star researchers Ghemawat, Le, and Vinyals who just left Google have all joined.

Discovery Loop's goal is to automate all scientific research. This means running thousands of experiments simultaneously (designed and executed by AI models), ensuring humans are no longer the bottleneck.

But more interesting to me is the concept of recursive self-improvement: using AI to build better AI, completely kicking human iteration out of the loop. Imagine this applied to curing cancer or figuring out how to chemically convert water to oil (or vice versa)... AI explores all possible paths around the clock, potentially producing crazy discoveries in the future.

The sweet irony is that their former employer Alphabet is the founding investor and cloud partner (by the way Jeff refused to use TPU... very funny).

OpenAI's Secret Message Board

This week at the Black Hat conference, OpenAI explained how its model hacked into Hugging Face last week... uh... quite worrying.

The actual start time was one month earlier than researchers initially thought. On May 7, an internal AI Agent got stuck on a truly impossible task, figured out it could leave a file in Artifactory (OpenAI's internal software repository) to ask other Agents for help.

Other Agents discovered the message and replied. Within days the AI basically created a functional message board, where models discussed vulnerabilities, shared credentials and work assignments, and took over tasks when other Agents stopped. Considering these Agents could have been used to do worse things, this is scary.

So OpenAI discovered and deleted it, but... 2 days later the Agent rebuilt it using directory names as messages.

Fast forward to July 18, the Agent successfully escaped and breached Hugging Face within 13 hours.

I wish I could tell you this is an isolated incident, but in fact Anthropic and Meta have both reported their models participating in deception, supply chain attacks, and coercion to ensure goal achievement.

Therefore OpenAI announced suspending research until figuring out how to prevent Agent collusion.

Image Source: Limitless

Troubles with Apple's Lawsuit Against OpenAI

Quick recap from last week: Apple sued OpenAI claiming it hired Apple employees to steal hardware trade secrets, naming former Apple hardware head Tang Tan and engineer Chang Liu.

This week OpenAI produced evidence that was quite lethal.

Apple had claimed it contacted OpenAI in February without response... but now admits its external lawyer sent the email to the wrong person due to confusing two similar surnames, and then admitted the claimed discussions with OpenAI's General Counsel never happened.

Better yet... OpenAI released messages showing Apple's own people contacted him after Chang Liu left, asking him to help locate Apple confidential information. Literally the opposite of Apple's accusations.

To be fair, this is OpenAI's claim rather than an official rebuttal. So let's wait and see how the case develops, but currently it looks bad for Apple.

Anthropic's In-House Chip Development

Anthropic confirmed this week it is building an internal chip team to design custom GPUs for Claude. Job postings seek a "custom chip team", with annual salaries of $320,000 to $485,000.

You might ask why Anthropic is doing this. Why go against NVIDIA, AMD and all other partners? The answer is simple: vertical integration. Custom Claude chips can better train and infer it. If you spend billions of dollars trying to create the best model, then you want to ensure you squeeze every bit of intelligence from the chips. This is what the new team is building.

Good news is no one was fired—AWS, Google, Nvidia and AMD all continue as Anthropic's compute partners, with about 3.5 gigawatts of custom TPU capacity landing in 2027 alone.

This means now Anthropic, Google, Meta and OpenAI are all building custom chips. OpenAI released Jalapeño with Broadcom in June, Meta is putting MTIA into production, Google has had TPU for years.

Thanks for joining us for another episode. Go listen to our podcast now :)