Explore what's new in AI

zerolang adds row syntax and claims 20% fewer tokens

A small syntax tweak claiming 20% fewer tokens is exactly the sort of thing long-running agent sessions can compound.

browser_use ships a Rust TUI and browser harness for agent tasks

The Rust TUI plus browser harness gives agent builders a lighter way to drive real browser tasks without a big framework pile.

Hamel Husain says eval automation still fails without human review

Hamel argues eval automation breaks down without human-in-the-loop review. Good warning against trusting dashboards blindly.

AutoScientist tries to compress model search into a 2-day loop

AutoScientist is trying to shorten model iteration into a 2-day loop with compute on us. If it works, experimentation gets much cheaper.

Files SDK 1.5 adds bulk ops plus SFTP, FTP, and Convex adapters

Bulk ops plus more storage adapters make the SDK useful for more than toy demos. That broadens where file workflows can live in production.

Hugging Face's Diffusers profiling guide shows how to find torch.compile bottlenecks

The guide shows how to profile real pipelines and find bottlenecks before guessing at optimizations. Useful for GPU-heavy code.

Uncle Bob says agent workflows should alternate features and hardening

The advice is pragmatic: burst on features, then harden the system. It's a better mental model than endless vibe coding.

AlphaSignalAI maps the spec-driven development stacks builders are converging on

The field is converging on the same patterns, spec files, contracts, and evaluation gates. This map helps teams borrow the right pieces.

Shreya's experiments show why qualitative analysis still resists naive automation

Agents converge too fast on shallow interpretations and humans get fatigued. That's a real warning for eval and analysis tooling.

dabit3 turns Slack and webhooks into event-driven Devin sessions

Slack becomes the trigger bus, GitHub and webhooks become actions, and Devin becomes the worker. That's a concrete event-driven pattern.

Dair AI spotlights a paper that distills agent workflows into weights

A full agent workflow distilled into weights could slash inference cost by 100x. That would make runtime loops far cheaper to ship.

AskVenice launches Agentic Chat for text, images, video, files, and research

One chat can now route text, images, video, files, and research without model-switching. That removes friction from multimodal workflows.

TanStack Virtual cuts resize storms and fixes iOS scroll jank

Resize storms and iOS scroll bugs are the papercuts that make UIs feel slow. This update targets both, with a huge perf win.

bcherny says teaching Claude Code basics can widen access to agents

Lowering the barrier to agents matters. Teaching the basics can get more teams using Claude Code instead of leaving it to experts.

OpenCode adds a diff viewer and prepares worktree-aware flows

OpenCode is making agentic code review less chaotic by surfacing diffs directly and getting ready for worktree-heavy workflows.

Unitree's G1 table-tennis demo shows humanoid control getting tighter

A table-tennis robot is still a demo, but it shows humanoid control is getting precise enough for fast reactive tasks.

ChatGPT fills scanned forms from images and voice prompts

ChatGPT can now fill forms from an image or voice prompt, turning paperwork into a multimodal extraction task.

T3 Code remote mode gets you a URL in two clicks and runs worktrees on a Mac mini

T3 Code's remote mode is a practical control plane, with quick URL sharing and built-in Tailscale for worktree multitasking.