Explore what's new in AI

Google expands Universal Commerce Protocol to more verticals and countries

UCP is moving beyond a protocol demo into hotels, food ordering, and more countries. That's a sign the agent commerce plumbing is getting standardized.

Google AI Studio lets teams build and deploy Android apps from prompts

Prompt-to-Android plus Firebase and Workspace hooks move AI app prototyping closer to deployable output.

Glass adds telehealth audio inputs to ambient scribing

Glass now ingests mic, Zoom, EHR tabs, or full-screen audio for ambient scribing. That expands where clinicians can capture visits without changing their setup.

CopilotKit adds hooks, a catalog, and AG-UI inspection panes

Three panes for hooks, component catalogs, and AG-UI events make agent UI debugging much easier. That's the sort of IDE extension teams can wire into daily workflow.

Gradio 6.15 inserts a Node proxy server in front of apps

Moving Gradio apps behind a Node proxy changes how SSR and serving work. If you ship Gradio in production, this can affect latency and deployment shape.

Perplexity's Unigram tokenizer claims 5x to 6x less CPU

Perplexity says its rebuilt tokenizer cuts CPU use dramatically for ranking and retrieval workloads. If the claim holds, that's real infra savings at scale.

react-doctor adds Preact support for automated anti-pattern fixes

It can now find Preact-specific anti-patterns and fix them from the CLI, which broadens the tool beyond React-only repos.

vLLM adds checkpoint-and-restore for full worker snapshots

Full worker snapshots in under 5 seconds make vLLM failover and debugging much less painful for large serving stacks.

DSPy refreshes its docs and front page ahead of a bigger 4.0

Fresh docs and a new front page should lower the learning curve before DSPy's bigger 4.0 shift, which matters for teams adopting prompt programming.

Crush adds multiuser preview mode to its agent workspace

Multiuser preview mode turns Crush from a solo agent shell into a shared workspace, which is a small but real collaboration upgrade.

Weaviate Engram stores agent memory as a persistent context layer

Engram treats memory as persistent context management for agents, which is a cleaner model than hoping chat history survives forever.

Google Cloud Agent Platform adds persistent memory and resume

Memory Bank and resumeable agents tackle the core production failure mode: losing decisions when a long-running agent crashes or pauses.

StepFun 3.7 Flash brings 196B multimodal MoE for agents

Step 3.7 Flash combines a 196B MoE with 11B active params, 256K context, and selectable reasoning levels for agent and coding workloads.

LiteParse claims the fastest open PDF parser and ties for accuracy

LiteParse argues semantic coherence matters as much as raw speed, and its benchmarks say open PDF parsing can be both fast and accurate enough for LLM ingestion.

AutoScientists decentralizes AI co-scientists around shared evidence

Decentralized agents, proposal review, and failure documentation helped across bio ML, LLM training, and protein fitness. The bottleneck looks procedural, not just model quality.

PaddleOCR-VL 1.6 sets a new OmniDocBench record at 96.33%

A 96.33% OmniDocBench score plus stronger table and formula handling make this a serious OCR layer for doc pipelines feeding LLMs and RAG.

vLLM adds native weight syncing and pause-resume for async RL

Standardized weight transfers plus pause-resume reduce deadlocks and glue code in async RL, which is exactly what bigger training loops need.

TanStack AI experiments with workflows and dynamic orchestrators

TanStack is splitting agent control into fixed workflows and dynamic orchestrators, a useful pattern for apps that need both reliability and branching.

VS Code v1.122 adds BYOK and stable custom AI endpoints

BYOK now works without a GitHub sign-in, and custom endpoints are stable. That makes local and enterprise model setups much easier to standardize.

Zed v1.4 adds reusable skills and project-scoped agent invocation

Skills can be created once, reused globally or per project, and invoked automatically or on demand. That makes agent behavior much less prompt brittle.

Octopool pools GitHub PATs behind a shared read cache

Octopool pools GitHub PATs and App installs behind a shared read cache, which can blunt rate-limit pain for agent-heavy teams.

Browser Use Terminal turns a shell into a Chrome agent

A terminal can now drive real Chrome, so browser automation can live in scripts and CI instead of a separate GUI tool.