Thursday, August 13
Thursday, August 13
Zed Says the Chat Log Is the Codebase Now

Zed just dropped Delta, a standalone app built on DeltaDB, a version control system the team has been building for over a year. Private beta invites started going out yesterday. The premise: when AI agents produce most of the code, the conversation directing them is the real source of your software, and your tools should treat it that way. Delta connects to Claude Code, syncing live terminal sessions into collaborative threads where teammates can watch, comment, and pick up work with full context. A browser version compiled to WASM lets non-developers follow along too.

Zed Says the Chat Log Is the Codebase Now
Zed just dropped Delta, a standalone app built on DeltaDB, a version control system the team has been building for over a year. Private beta invites started going out yesterday. The premise: when AI agents produce most of the code, the conversation directing them is the real source of your software, and your tools should treat it that way. Delta connects to Claude Code, syncing live terminal sessions into collaborative threads where teammates can watch, comment, and pick up work with full context. A browser version compiled to WASM lets non-developers follow along too.
AI Models, Milestones, and Big Industry Moves
Eighteen months ago, GPT-4-class inference cost roughly $60 per million output tokens. Today, comparable quality runs under a dollar.
- For context, solar panel costs took about a decade to fall 90%. Hard drive costs took seven years for a similar drop in the early 2000s. AI inference is outpacing both curves by a wide margin.
- Andrej Karpathy coined "vibe-coding" in February 2025. The category now supports multiple billion-dollar companies with its own investment thesis.
- Open-weight models have historically trailed closed-source frontier capability by six to twelve months. That lag has compressed to weeks.
When the price of a core capability drops this fast, the companies that built their business on scarcity need a new story. Several showed up with theirs yesterday.
Eighteen months ago, GPT-4-class inference cost roughly $60 per million output tokens. Today, comparable quality runs under a dollar.
- For context, solar panel costs took about a decade to fall 90%. Hard drive costs took seven years for a similar drop in the early 2000s. AI inference is outpacing both curves by a wide margin.
- Andrej Karpathy coined "vibe-coding" in February 2025. The category now supports multiple billion-dollar companies with its own investment thesis.
- Open-weight models have historically trailed closed-source frontier capability by six to twelve months. That lag has compressed to weeks.
When the price of a core capability drops this fast, the companies that built their business on scarcity need a new story. Several showed up with theirs yesterday.
Dev Tools, Deep Bugs, and Security Alerts
Favorite Featured Stories

Three engineers argued in 1984 that only the two ends of a conversation can judge whether an exchange came out right. TC...

Site operators are writing robots.txt rules against crawlers that were switched off years ago, and against crawlers nobo...

Watch a coding agent work for an hour. Forty files, tests written and failed and rewritten, and when it stops, no user h...

The machine-readable layer of your website used to be decoration. A few structured tags under the product page, worth a ...

Three engineers argued in 1984 that only the two ends of a conversation can judge whether an exchange came out right. TC...

Site operators are writing robots.txt rules against crawlers that were switched off years ago, and against crawlers nobo...

Watch a coding agent work for an hour. Forty files, tests written and failed and rewritten, and when it stops, no user h...

The machine-readable layer of your website used to be decoration. A few structured tags under the product page, worth a ...