Explore what's new in AI

scriptc checks npm values against their declared TypeScript types

Runtime checks at package boundaries can catch mismatches between actual values and published declarations without requiring teams to add validation code throughout their applications.

Theo details T3 Code’s agent tooling and recent engineering changes

The update describes repo transfer, CI babysitting that cuts GitHub API calls by 90%, usage analytics, thread controls, and a TypeScript compiler rewrite across the team’s agent workflow.

T3 Code demo connects Ubuntu to a remote iOS simulator on Mac mini

Theo demonstrated an iOS simulator running on a connected Mac mini while using T3 Code from Ubuntu. A separate demo showed remote simulator interaction from an iOS device.

Ray Fernando reports over 40 releases from one Cursor project in two weeks

Ray Fernando says one Cursor project shipped over 40 Ray Transcribes releases in two weeks while he was intermittently online. A release rule directed agents to test releases in separate cloud environments.

T3 Code ships a speed improvement tested on 3G networks

Theo announced another T3 Code speed improvement and said load-time testing includes 3G network speeds. The posts provide no numerical performance result.

Theo tests an Opus-driven loop to improve T3 Code’s interface

Theo proposed turning T3 Code’s UI into a hill-climbing task for Opus, then posted progress images and a pull request reporting improvements. The posts provide no quantified results.

Matt Pocock argues agent languages need strict verification

Pocock proposed Rust-like speed, a small design space and strict verification for agent-written software. He called Bend promising but said a few prompts broke its playground.

Device Hub demonstrates Devin-powered iPhone development on cloud Macs

A demo showed multiplayer iPhone Duo development using Devin Cloud Agents on real macOS machines. The workflow deploys to TestFlight and the App Store.

Dan Luu argues AI makes more ambitious software projects practical

AI agents lower the cost of tackling long-avoided tools and bottlenecks, but useful products still demand testing and verification as models expand what teams can build.

vLLM’s ModelState abstraction isolates model-specific serving logic

The proposed stateful interface provides hooks for extra inputs, attention metadata, and CUDA graph capture, aiming to simplify support for diffusion LLMs and encoder-decoder models.

T3 Code adds guest links for collaborative agent threads

Guests can add notes to a shared thread while the owner retains approval control, making it easier to bring coworkers into agent work without copying context.

T3 Code passes 30,000 GitHub stars in 234 days

The project reached 30,000 stars after crossing 10,000 in April and 20,000 in August, offering a visible signal of interest in agent-focused coding tools.