Explore what's new in AI

techhalla's Seedance 2.0 prompt scripts a live sports broadcast shot

The prompt is a reusable template for tightly controlled video generation, with explicit camera beats, timing, and audio notes.

tdiff v0.1.1 adds line wrapping to terminal diffs for easier code review

A tiny wrapping change makes long lines readable and reduces side-scrolling in terminal review flows. Handy for anyone living in CLI code review.

Plannotator 0.18.0 adds wide review, inline diffs, and PDF export

The update makes plans easier to inspect and hand to agents, which helps keep long workflows auditable instead of buried in chat.

Browser Use launches agent-native signup with a reverse-CAPTCHA

Agents can now create accounts by solving a challenge made for them, not humans. It's a sharp example of product flows adapting to browser automation.

Kimi K2.6 rolls back on the web and reappears in the arena as Kiwire

The model is being retested under a codename after a web rollback, which suggests rollout confidence is still uneven. Arena testing will show whether the new build is actually ready.

One agent conversation syncs across Slack and Telegram in Sunil Pai's setup

A single thread can now live across multiple messengers and still share state. That kind of channel-agnostic conversation model is handy for agent workflows.

Anthropic console bug gets reverse-engineered by Benhylak

By dumping the JS bundle and reconstructing the request, he fixed a UI bug that had no network call. It is a good reminder that agentic debugging can cross from UI into protocol work.

DSPy crosses 6M monthly downloads as its abstraction stack keeps spreading

The download milestone shows how far the signatures, modules, and optimizers model has spread. It is still one of the most recognizable agent-programming abstractions.

Hugging Face Skills exposes 3M models and 1M datasets to agent setups

One install can expose ~3M open models, 500k local apps, and ~1M datasets to an agent. That makes model selection and hardware-aware routing much easier.

ML-Master 2.0 keeps a research agent alive for 24 hours on MLE-Bench

SJTU's setup uses hierarchical cognitive caching to retain short, medium, and long-term state across sessions. The result is a 56.44% medal rate, one of the strongest long-horizon marks yet.

Claude's tool descriptions are the missing manual, Simon Willison says

Public system prompts help, but the hidden tool catalog still blocks power users. Publishing tool descriptions would make Claude much easier to steer well.

Hermes Agent crosses 100,000 GitHub stars

Nous's open-source agent framework just hit a major adoption milestone. That usually means the community sees it as a real default, not a science project.

Autogenesis adds git-style commit and rollback to self-improving agents

The protocol lets agents edit prompts, tools, and memory without losing the ability to undo changes. Its claim is simple: self-improvement needs version control.

Avichawla's RAG vs CAG guide caches static knowledge in KV memory

Static docs move into prompt cache, dynamic facts stay in retrieval. The hybrid cut lowers latency and cost without giving up freshness.

Gemini's app harness still lags behind the model, Ethan Mollick says

He argues Google still lacks file ops, research tools, and auditable actions in the UI. That gap keeps Gemini harder to use than Claude or ChatGPT.

omarsar0 turns podcasts into reusable knowledge dashboards with LLM artifacts

The workflow turns audio into HTML artifacts with charts, notes, and drill-downs that agents can reuse later. It is a promising pattern for durable research surfaces.

Claude Code blocks an open-source task after a provider policy check

The model provider halts an open-source request after 38s with an acceptable-use error, showing how policy gates can stop benign dev work mid-run.