Explore what's new in AI

Marlene Mhangami uses Playwright and MCP to test real functionality

The guide argues code coverage misses user-visible bugs, and that browser-driven checks are the better target for agent tests.

DSPy signatures turn tasks into compact AI programs

The thread shows how a few symbolic lines can compile into prompts and structured outputs, a practical way to make workflows explicit.

Liquid's LFM2.5-8B-A1B wins a tool-calling test against gpt-oss-20b

The smaller model finished every required call, used less RAM, and ran faster, showing control quality can beat raw scale.

TokenJuice passes 2.5M npm downloads and broadens IDE support

A token-reduction SDK that now spans many IDEs is becoming a real integration layer, not just a side project.

LiteParse uses grid projection to parse PDFs without an LLM

Its layout-first algorithm turns columns, tables, and flowing text into readable text, which can save a lot of token spend.

Qdrant Edge powers local semantic music search on device

The demo combines Whisper, embeddings, and local vector search so users can search music by mood without a cloud round-trip.

W&B's MCP server lets agents read experiments and monitor runs

Agents can now pull experiment state directly from W&B, turning observability into something model workflows can consume.

VS Code Chronicle stores Copilot chats in local SQLite for later review

A local history store makes agent sessions queryable after the fact, which helps debug prompts and compare habits over time.

Perplexity tests digest controls for memory, web sources, and connectors

Users may soon tune digests more tightly, which makes the feature feel like a workflow surface instead of a generic summary.

NotebookLM tests Canvas artifacts, personal preferences, and connectors

The new surface would move NotebookLM toward persistent context and richer outputs for research workflows.

LateOn beats v2 by almost 10 BEIR points in PyLate

That kind of retrieval jump is large enough to matter in real search stacks, not just benchmark theater.

OpenAI-Proof Q&A turns internal bottlenecks into a benchmark

The setup uses real code, logs, and artifacts, so pass@1 reflects debugging skill rather than toy puzzles.

Ego Lite lets agents inherit login state in an isolated browser

A browser that reuses your session but runs the agent in a separate space could cut a lot of brittle setup work.

Omar Sar0 uses HTML artifacts as a verification layer for agents

Artifacts give long-running agents a shareable output surface, so you can inspect work without trusting chat logs alone.

Claude Code Harness pins agent work to spec.md and Plans.md

The repo makes agent steps auditable with a product contract, task ledger, and guardrails, which is the right shape for long runs.

OpenAI Devs demo a voice-first agentic phone OS

The hack night app lets voice commands trigger actions across the phone, a concrete glimpse of hands-free agent UX.

OpenAI Devs build Wagner, a multi-agent meeting room

The demo puts several agents in one room to reason through infrastructure plans together, closer to collaboration than chat.