GitHub Copilot agents rewrite 800K lines of their own runtime in Rust; Zed replaces pull requests with Delta; AI query optimizer beats Postgres

Engineering

Zed

Zed released Delta, a multiplayer coding environment that replaces pull requests with threaded collaboration between developers and agents. Delta uses DeltaDB, which extends Git with delta-based versions that preserve edit history while maintaining Git compatibility. The Zed team disabled pull requests on their own repository and landed 570 changes using only Delta.

Read →

Rohan Bansal

Training a 4B Model for Query Optimization

The author trained a 4-billion parameter language model to optimize PostgreSQL query plans. This model produced 44.7 percent faster queries than the default optimizer on 113 benchmark queries. The project cost $1,200 and shows that small models solve domain-specific problems.

Read →

Hacker Factor

Forgery of C2PA on a Pixel 10

Google Pixel devices contain a critical vulnerability in their C2PA (Coalition for Content Provenance and Authenticity) implementation. An attacker with root access can sign fake or AI-generated images with legitimate Google cryptographic signatures. Google classified this issue as "Won't Fix" because changes needed to address the problem are not reasonably possible. The vulnerability affects multiple C2PA implementations and prevents validators from checking certificate revocation.

Read →

Commoncog

How to Handle Disruptive Change

Company survival during disruption depends more on balance sheet strength than leadership stubbornness. A strong balance sheet provides time to develop a coherent strategy response. Leaders must stabilize the business first, then execute a path to compete again.

Read →

Astral Codex Ten

AI Superforecasting Should Transform The FDA

AI superforecasting can improve FDA drug approval decisions by replacing subjective human judgment with quantitative predictions. The FDA review process now creates uncertainty for drug makers, causing them to add months or years of delays and millions to billions of costs. The proposal uses three stages: add forecasts to decisions, make forecasts public for transparency, then use validated forecasts as automatic triggers for approval or rejection.

Read →

Crooked Timber

Existential Risks in an Age of Science

Two separate groups—climate activists and AI safety advocates—hold apocalyptic views, though both trace problems to human action. Doomerism began in ancient times, but past scientific predictions of catastrophe proved wrong. Both movements stall because societies cannot agree on who should pay for prevention methods.

Read →

Humanities

Aeon

Who Will Taste the Cherries?

The essay critiques computism, an ideology that treats all human experiences as quantifiable data that algorithms can optimize. When people rely on algorithms for decisions, they become passive data-generating objects rather than active agents. The article argues that subjective experiences like taste cannot reduce to measured data, and pursuing perfect measurement risks losing embodied human experience.

Read →

JSTOR Daily

How the Railways Killed a Medieval Law

The deodand was a medieval English law that required the crown to receive objects that caused death. Juries decided the value of objects, which sometimes went to accident victims' families. When railways killed passengers, juries set very high values for trains, which changed public opinion against the law. In 1846, parliament removed deodands from law to protect railway companies from large compensation claims.

Read →

Migrating the GitHub Copilot Runtime to Rust, Using Copilot