Friday, August 7
Friday, August 7
The Two Engineers Who Built Google Just Left to Build Something Else
Jeff Dean and Sanjay Ghemawat are leaving Google to found Discovery Loop, an independent public benefit corporation focused on automating ML, science, and engineering. The same Wednesday, Demis Hassabis stepped back from CEO of DeepMind to become chairman and Alphabet's chief scientist, with CTO Koray Kavukcuoglu taking the SVP seat. Alphabet shares dropped 4%. This all lands while Google's next flagship Gemini model sits unreleased past its planned June launch window. The architects of MapReduce, BigTable, and TensorFlow are voting with their feet that the work they care about is better done outside a corporate lab.
The Two Engineers Who Built Google Just Left to Build Something Else
Jeff Dean and Sanjay Ghemawat are leaving Google to found Discovery Loop, an independent public benefit corporation focused on automating ML, science, and engineering. The same Wednesday, Demis Hassabis stepped back from CEO of DeepMind to become chairman and Alphabet's chief scientist, with CTO Koray Kavukcuoglu taking the SVP seat. Alphabet shares dropped 4%. This all lands while Google's next flagship Gemini model sits unreleased past its planned June launch window. The architects of MapReduce, BigTable, and TensorFlow are voting with their feet that the work they care about is better done outside a corporate lab.
Shipping and Competing Across the AI Agent Stack
The terminal coding agent market is about eighteen months old and already has five major players throwing elbows. A year ago, "AI coding assistant" meant autocomplete in your editor. Now these tools coordinate persistent subagents across entire codebases and bill by the million tokens.
- Inference costs dropped roughly 10x over the past year, turning per-engineer token budgets from fantasy into finance-team line items.
- DeepSWE, the benchmark everyone watches for agentic coding, didn't exist two years ago. It's the scoreboard now.
- The latest generation of agentic models averages 60+ autonomous turns per task, up from around 14 one generation ago. That's a different category of behavior entirely.
The money moving through this stack — chip acquisitions, per-seat token bills, infrastructure bets — is getting concrete fast.
The terminal coding agent market is about eighteen months old and already has five major players throwing elbows. A year ago, "AI coding assistant" meant autocomplete in your editor. Now these tools coordinate persistent subagents across entire codebases and bill by the million tokens.
- Inference costs dropped roughly 10x over the past year, turning per-engineer token budgets from fantasy into finance-team line items.
- DeepSWE, the benchmark everyone watches for agentic coding, didn't exist two years ago. It's the scoreboard now.
- The latest generation of agentic models averages 60+ autonomous turns per task, up from around 14 one generation ago. That's a different category of behavior entirely.
The money moving through this stack — chip acquisitions, per-seat token bills, infrastructure bets — is getting concrete fast.
AI's Trust Problem Is Not Getting Any Better
Favorite Featured Stories
Durable execution promises something real. Your workflow survives crashes, evacuated hosts, a datacenter having a genuin...
Every agent deployment estimate runs the same arithmetic. Per-step success rate, raised to the number of steps, watch th...
In June 2000, Congress gave machine-formed contracts federal standing, with one condition attached: the agent's action h...
Somewhere in a traffic policy sits the suggestion that you break your automation into three crawlers, one per purpose, s...
Durable execution promises something real. Your workflow survives crashes, evacuated hosts, a datacenter having a genuin...
Every agent deployment estimate runs the same arithmetic. Per-step success rate, raised to the number of steps, watch th...
In June 2000, Congress gave machine-formed contracts federal standing, with one condition attached: the agent's action h...
Somewhere in a traffic policy sits the suggestion that you break your automation into three crawlers, one per purpose, s...