Loading page…
The transaction moves a software-engineering evaluation team into Google’s model-development work. Its final price remains private, despite earlier reports of talks valued above $1.5 billion.
Listen to this story
Google has finalized a talent-focused arrangement that moves more than a dozen Mechanize employees into Google, with many assigned to AI midtraining and co-founder Tamay Besiroglu joining Google DeepMind as a research scientist. The deal gives Google expertise in software-engineering evaluations and training environments, including Mechanize’s GBA Eval. Financial terms remain undisclosed: an earlier report cited negotiations above $1.5 billion, while Mechanize had previously raised $9.1 million at a $500 million가?
More than a dozen former Mechanize employees now appear on Google professional profiles; many are working on midtraining.
Besiroglu joined Google DeepMind as a research scientist; former chief of staff Guive Assadi identifies as Mechanize’s CEO.
Mechanize’s GBA Eval asks models to build a Rust Game Boy Advance emulator compiling to WebAssembly.
Google has completed a talent-focused deal for Mechanize, bringing more than a dozen former employees of the San Francisco AI startup into the company. Many are now working on Google’s AI midtraining efforts, a stage of model development and refinement, while Mechanize co-founder and former CEO Tamay Besiroglu has taken a research-scientist role at Google DeepMind.
The arrangement turns a small, specialized startup into a source of staff for Google’s AI work. Besiroglu’s DeepMind role and the other employee moves are reflected in public professional profiles. Google and Besiroglu declined to comment on the deal or the personnel changes.
Mechanize says its current focus is software engineering, where it builds environments for evaluating and training large language models. Its technology is designed to help AI systems improve coding capabilities, connecting the team’s public specialty to the reported midtraining assignment at Google.
In a May technical post, the company presented GBA Eval as an example of that work. The task asks models to write a Game Boy Advance emulator in Rust that compiles to WebAssembly, using detailed grading to assess a long-running software-engineering job rather than a short coding prompt.
Public professional profiles show more than a dozen former Mechanize employees now work at Google; many are working on midtraining.
Mechanize’s co-founder and former CEO joined Google DeepMind as a research scientist, according to his LinkedIn profile.
The final financial terms have not been disclosed. Earlier reporting said Google had been negotiating a deal worth more than $1.5 billion for Mechanize’s technology and talent, but that figure described the talks, not a confirmed price for the completed agreement.
The gap between the earlier valuation of the talks and the public outcome is striking. Mechanize said earlier this year that it had raised $9.1 million at a $500 million valuation, after launching with a mission of automating every job in the world. Its former chief of staff, Guive Assadi, now identifies himself as chief executive on LinkedIn.
For Google, the deal supplies people with experience in a narrow but consequential area: creating the tasks and grading systems used to assess and train coding models. What Google paid remains private; the public evidence so far is a team transfer into its midtraining work.
Editorial analysis
This is a talent transaction whose clearest public output is not an acquired product but a team assignment. Mechanize’s public work centered on building environments to evaluate and train language models for software engineering, while many of its former employees are now reportedly working on Google’s midtraining efforts. That makes the transfer a focused bet on a difficult part of model development. The next meaningful public signal would be evidence of how Google applies that expertise: whether through new evaluation methods, training environments, or improvements in coding systems.
Loading discussion...
Join the conversation
Explain what details should be public.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.