Explore what's new in AI

Google DeepMind's AMIE paper shows longitudinal clinical reasoning with Gemini

The Nature paper says AMIE can ground clinical reasoning in authoritative medical knowledge across visits, which is a meaningful healthcare milestone.

OpenAIDevs says Women who Code(x) shipped task agents with Codex

This is a concrete case study of builders using Codex to ship task agents and personal guides, not just prototype them.

PJaccetturo gives away a 1,000-image Seedance pack for fan trailer workflows

A big image pack plus a Seedance workflow makes this a ready-made template for anyone trying to build their own animated trailer.

Techhalla combines Midjourney, Nano Banana, and Seedance into one visual workflow

This is a useful template for visual prototyping. The stack handles still generation, image editing, and animation in a single loop.

Zeeg overhauls a coding harness plugin for Claude, Codex, Cursor, and Grok Build

A single plugin spanning multiple agents makes it easier to compare harness behavior without rewriting your setup every time.

Matt Shumer says prompting works best when you point at the shape of a solution

Shumer argues that good prompting means specifying the shape of the answer, then letting the model iterate against a clear quality bar.

unclebobmartin says review agents should remove PR drudgery

The useful point is that review agents shine when they remove repetitive PR work, not when they replace human judgment on hard code.

Paper adds nested folders to its dashboard for better doc organization

Nested folders are a small UI change, but they make doc-heavy collaboration easier to keep sane as teams grow.

Corey Ganim's llm-wiki turns raw notes into a three-folder second brain

The raw, wiki, outputs pattern is a simple way to turn messy notes into reusable context and deliverables without manual organizing.

ComfyUI turns one frame into a cinematic traversal with Seedance 2

One still frame becomes a continuous fly-through, showing how far perspective control has advanced in AI video workflows.

Genspark AgentBase turns data into custom databases and internal systems

AgentBase tries to productize back-office software by turning company data into dashboards, CRMs, and internal workflows from one prompt.

DeepLearning.AI teaches voice layering for agents with a minimal-code course

The course shows how to add voice without rewriting prompts, RAG, or tools, which is exactly what most builders want.

Pika Director's Suite lets creators edit every part of a scene in one place

Director's Suite bundles concepting, generation, and editing into one interface, which pushes video workflows closer to an end-to-end studio.

LiteParse makes Claude better and cheaper on PDFs with BM25 retrieval

LiteParse pairs BM25 retrieval with fewer expensive file ops, and the team says it makes Claude Code 37 percent cheaper on PDFs.

Zed v1.7 auto-compacts conversations when context gets tight

Zed now compacts agent conversations automatically, which is a small but very practical fix for context-window pain.

Google's Infinite Scaler builds a live game from Gemini, Nano Banana, and Lyria

The demo chains prompt expansion, sprite-sheet generation, and music synthesis into one playable loop, which is a strong multimodal artifact.

ast-grep becomes a deterministic linting guardrail for agents

Deterministic lint rules are a much safer guardrail than hoping an LLM notices everything, especially in automated coding loops.

PreAct replays computer-use runs 8.5 to 13x faster

PreAct turns a solved computer-use task into a replayable program, avoiding repeated re-reasoning on every later run.

Apodex 1.0 runs evidence-backed deep research with 150 subagents

Apodex is notable because it insists on auditable evidence and scales the work out to a large subagent team before answering.

OpenAI launches LifeSciBench for real-world life-science workflows

LifeSciBench is a more realistic test of model usefulness in biology, with tasks built around evidence, uncertainty, and artifact handling.

OpenAI uses GPT-5.4 to improve Chan-Lam coupling yields

OpenAI says GPT-5.4 helped chemists improve a real reaction, with thousands of experiments and hand-checked yield gains. That is closer to applied science than demo ware.

alyze open-sources a tokenizer that runs up to 4x faster

alyze pares down tokenization work and exposes a WASM demo, which is exactly the sort of plumbing optimization search teams bookmark.

NotebookLM prepares personal preferences and AI editing for notes

NotebookLM is moving toward persistent preferences and AI-written notes, which makes it more useful as a long-lived knowledge layer.

OpenAI backs the Rust Foundation with open-source support

OpenAI is explicitly supporting Rust's ecosystem, a useful signal for the systems language that powers a lot of modern infra.

ChatGPT adds scheduled tasks and starts sunsetting Pulse

Scheduled tasks are getting a cleaner, more reliable home in ChatGPT, while Pulse is being phased out over the next two weeks.

Matt Pocock's skills v1 cuts token cost and splits invocable skills

The skills update trims description tokens by 63 percent and makes reusable skills more explicit. That is a real context-efficiency win for agent work.

GitHub Copilot makes Copilot CLI the default harness in JetBrains

GitHub is standardizing more of Copilot around one CLI harness, which should make behavior, governance, and observability easier to keep consistent.

GitHub Copilot app goes GA with handoff from mobile to VS Code

Copilot's app now acts as a clean handoff point between mobile agent work and full coding in VS Code without losing context.

Jacobli99 argues agents should learn domains from textbooks and docs

The thread says current agents stay too shallow, and that reading textbooks and docs could build deeper domain expertise like humans do.

Physical AutoResearch hardwires safety and telemetry for overnight robot runs

The setup bakes safety into the hardware loop and tracks robot, GPU, and token utilization, which is exactly what unattended experimentation needs.