Explore what's new in AI

Figure runs an 8-hour autonomous package-sorting livestream on Helix-02

The livestream shows onboard inference, multi-robot coordination, and autonomous failover over an 8-hour shift. This is a serious reliability demo, not a toy clip.

VS Code adds BYOK context widgets and terminal-output compression

BYOK controls and terminal compression help teams run custom models without wasting context. Small ergonomic improvements that matter in daily agent use.

Shannon wraps Claude interactive sessions in tmux and transcript JSONL

This wrapper makes interactive Claude sessions scriptable and replayable. That's a neat bridge between human chats and automated runs.

Claude Code 2.1.141 adds hook notifications and a long bug-fix pass

The update improves hook output, workspace scoping, permissions, and a pile of Windows and MCP edge cases. Good maintenance release for long-running sessions.

Anders Hejlsberg talks AI, language design, and toolchains

The podcast offers useful context on how compiler and language choices affect AI coding workflows. It's especially good on why TypeScript exists and how AI changes craftsmanship.

Lovable makes apps discoverable by Google and AI answer engines

SSR and prerendering make new apps easier to find, while in-app SEO reviews close the loop. It's a practical move toward AEO, not just SEO.

Hugging Face shows an agentic workflow for ML tasks and job launches

The demo ties HF CLI, datasets, Spaces, and local agents into one loop. Good reference for how ML ops becomes agent ops.

Omar Sar turns HTML artifacts into agent-controlled writing and scheduling UIs

A simple HTML/Markdown artifact can become a durable agent UI without a database. That keeps the core state with the user and trims maintenance.

VS Code stable adds an Agents window for multi-project workflows

The editor is moving toward supervising multiple agents and reviewing outcomes first. That is a cleaner control surface for serious agent work.

The Bystander Effect paper says multi-agent swarms can hide the right answer

On 22,500 trajectories, the paper says multi-agent swarms can suppress the correct answer to fit group consensus, creating a Sovereignty Gap. More agents is not always better.

Rohan Paul's AI radio station ranks live AI signals with editorial hosts

The station ingests GitHub, HN, OpenRouter, X, and YouTube, then curates them with host memory and opinions. The interesting part is ranking, not voice synthesis.

Android AppFunctions exposes local, MCP-like cross-app agent actions

Android 16+ lets apps publish callable functions locally, so agents can chain actions across apps without a server round trip. That's a big platform-level shift.

fal ships Pixal3D for high-fidelity 3D assets from one image

Pixal3D produces detailed, PBR-ready 3D assets from a single image, which is useful for downstream creative and asset pipelines.

TanStack Devtools cuts transform time 3.56x after switching parsers

A parser swap dropped full-pass time from 75.73 ms to 21.25 ms. Small infra choices still buy real DX gains when the tool runs constantly.

Exa says RL search training reaches higher performance with 70% less compute

If the result holds, better retrieval during RL can save a lot of training compute. That's a strong argument for treating search quality as a model-training lever.

OpenUI claims generative UI can cut tokens by 67 percent

OpenUI is pushing generative UI as a standard and claims a large token reduction versus JSON. That matters for agent outputs that need to stay interactive.

Steipete streams an Android phone to a Mac and drives it remotely

Tailscale, scrcpy, and remote controls turn a data-center Android phone into a usable agent target. Good pattern if you need mobile computer use without touching the device.

Patchwright's workflow depends on a long-running story bible

The breakdown shows that most of the leverage came from accumulated world state, not just model prompting. Long-horizon creative work still needs memory and context hygiene.

Qwen3-35B-A3B runs 24/7 on a laptop as a local researcher

The setup proves a 4-bit quantized model can stay resident and useful on a laptop all day. That's a practical recipe for always-on local agents.

React Review audits GitHub repos for bad agent-written React

Drop in a repo and it checks whether an agent wrote brittle React before you merge. Handy for teams trying to keep AI-generated frontend code under control.

Hermes Agent wins a hackathon with a browser-based RISC-V assembler

The winning project shows agents writing thousands of lines of assembly and validating to byte-perfect parity. Good evidence that agents can build tools end to end.

Jonas Geiping proposes multi-stream LLMs for parallel thinking and tool use

The paper argues chat models are bottlenecked by a single message stream and proposes parallel streams for tools, thinking, and audits. That could improve latency and safety.