Explore what's new in AI

Google Stitch plus Claude Code builds a client site from a mockup

A mockup gets turned into a Next.js site by piping the generated design system through MCP. It is a clean spec-to-code workflow for non-designers.

Hermes Agent webhooks turn real-time events into automation

The demo shows agents reacting to GitHub, newsletter, and mention events, then posting updates back into the right systems.

Claude Design turns prompts into logos, palettes, and landing copy in one pass

The workflow compresses early brand work into one session, which makes it useful for roughing out launches before a human polish pass.

Xiaomi MiMo-V2.5-Pro posts strong benchmark gains on OpenRouter

Xiaomi's new model lands near frontier systems on coding benchmarks while keeping a 1M context window and low per-token pricing.

GPT-2 Image plus Claude Code generate working OTF fonts in 30 minutes

A prompt-generated typeface concept plus Claude Code tracing got from mockup to OTF files in under 30 minutes.

LiteParse parses complex PDFs into a clean spatial grid without models

Its heuristics keep tables and flowing text aligned, giving agents a fast, cheap document parser when VLMs are overkill.

PaddleOCR 3.5 brings browser OCR and doc-to-Markdown conversion

The release adds browser-side OCR, WebGPU and Wasm support, and one-click document conversion for private local workflows.

HornetDev tunes ANN search at 100M scale with graph and quantization fixes

The post breaks down where ANN search breaks, from embedding bias to graph connectivity and quantization ceilings.

Claude Code gets a research-preview /ultrareview bug-hunting fleet

Anthropic is pushing code review into a fleet of cloud agents that can surface findings back into the normal workflow.

Ratpack layers synced multi-user collections on Cloudflare Durable Objects

Ratpack handles sync, per-client visibility, and session topology while TanStack DB owns the data model.

QueryWeaver turns text-to-SQL schema lookup into graph traversal

Instead of embedding tables as documents, it walks foreign-key paths to find bridge tables vector search misses.

Firecrawl adds change tracking to /crawl for site monitoring

Pages now come back marked new, same, changed, or removed, which makes docs, pricing, and inventory monitoring far less manual.

TanStack AI adds debug logging for chunks, hooks, and tool calls

A single debug flag now prints the whole stream path, making it much easier to trace black-box agent runs.

TanStack AI makes provider-tool mismatches fail at compile time

Model-specific validation now catches invalid tool combinations early, which removes a whole class of runtime fallbacks.

DeepSeek updates its reasoning API with 1M context and faster answers

The updated deepseek-reasoner appears to target long documents and larger workflows without bloating outputs.

Google Photos Auto frame reinterprets 2D photos as 3D scenes

The edit feature reasons about depth after capture, letting users reframe moments instead of being stuck with one crop.

BridgeSpace 3 launches a 10-agent swarm for app building

The new version launches 10 agents at once, with voice orchestration, browser tools, and IDE views for larger task flows.

Gemini Embedding 2 adds one multimodal embedding space for text, images, audio, and video

One embedding model now spans multiple modalities, which simplifies multimodal search and retrieval pipelines.

TACO learns context-compression rules for terminal agents

A self-evolving wrapper cut token overhead by about 10% while preserving or improving accuracy on TerminalBench and SWE-Bench.

BLF beats top forecasting systems with sequential Bayesian updating

The system matches human superforecasters on ForecastBench and beats GPT-5, Grok, and other top methods.

Holo3 takes the OS-World computer-use lead at 80.4%

The latest leaderboard puts Holo3 ahead of Kimi K2.6 and Claude Sonnet 4.6, showing computer-use evals are still moving fast.

browser_use ships /autobrowse for self-correcting browser tasks

The skill learns from failed attempts, converges on a stable workflow, and writes the winning steps back to git.

nrehiew's benchmark says frontier models over-edit code and RL helps

GPT-5.4 over-edits the most in this setup, while RL beats SFT, DPO, and rejection sampling for minimal diffs.

Google packages agentic context into the Agentic Data Cloud

Google is turning business data into an agent substrate with a universal context engine and cross-cloud lakehouse. That should make retrieval and orchestration less brittle.

Jules opens early access for an end-to-end product development agent

Google is testing a system that reads product context, suggests work, and ships a PR. That pushes agents up from code completion to product ops.

OpenAI brings ChatGPT to Google Sheets for live workbook editing

The screenshot shows ChatGPT updating a forecast from email context and checking its own work, which turns Sheets into an agent surface.

Zed ships Parallel Agents with a threads sidebar for multi-repo work

One thread can now update API and frontend repos together, which cuts context switching when agents touch more than one codebase.

Cursor adds a shadcn plugin to its marketplace

The plugin gives Cursor a fast path for generating shadcn UI code, which makes it easier to scaffold common front-end pieces inside the editor.

Perplexity post-trains a Qwen model for cheaper search answers

Its SFT plus RL pipeline improves search quality and citation behavior while cutting serving cost, and the model is already in production.

OpenClaw fixes a silent Codex harness fallback that broke agent loops

A broken auth bridge had been dropping requests onto the wrong harness. The fix restores real multi-step behavior without prompt changes.

Google says 75% of new code is AI-written and engineer-approved

Google's internal adoption keeps climbing, and the latest chart says one complex migration ran 6x faster than a year ago.

Claude Cowork adds interactive charts and diagrams for paid users

Anthropic is moving beyond static artifacts, so teams can build richer visual explanations and dashboards inside chat.

OpenAI Codex screenshots expose GPT-5.5 and other model codenames

Codex picker screenshots surfaced GPT-5.5, oai-2.1, and several codenames. Leaks like this often show the rollout path before launch.

AI scientists produce results without reasoning scientifically

The study says base model choice dominates performance, while evidence gets ignored in most traces. That is a warning about agentic science workflows.