From Product Hunt

Claude Code gets multi-agent code review, and a couple of useful productivity tools launched today.

Claude Code /ultrareview — Cloud code review with parallel agents Dispatches a fleet of specialized reviewer agents that examine your PR in parallel on Anthropic's cloud, each looking for different issue classes (logic errors, security vulns, regressions). Averages 20 min per review. For Pro and Max users.
▲ 181 · code.claude.com
Kollab — Shared workspace where teams work with agents AI agents inside your Slack and Telegram via MCP. Connectors link tools you already use, Skills let anyone reuse workflows, and Memory keeps context alive across projects. Scheduled agent tasks act like timed cron jobs with full MCP access.
▲ 301 · kollab.im
FocuSee 2.0 — Screen recording to polished demos Screen Studio alternative for Mac and Windows. Auto zoom, 3D motion effects, AI subtitles in 50+ languages, background removal, voice enhancement, and filler word removal. From raw recording to share-ready video in minutes.
▲ 301 · gemoo.com
Fabric CLI — Notes, tasks, and search from your terminal Write, search, and collaborate from the terminal with a personal AI that knows your projects and files. Acts as a memory layer for you and your agents. Now available as a CLI alongside the main app.
▲ 110 · fabric.so
From Reddit

GPT-5.5 dropped today. Plus Tencent goes open-source, Google reveals 75% AI code, and DeepSeek open-sources its MoE infrastructure.

Tencent open-sources Hy3 Preview — 295B MoE with 21B active Tencent's first model after a full infrastructure rebuild, led by former OpenAI researcher Yao Shunyu. 256K context, 74.4% SWE-bench Verified (up from 53%), 67.1% BrowseComp. Went from cold start to open-source in under 3 months. Weights on HuggingFace.
Google: 75% of new code is now AI-generated Sundar Pichai revealed at Cloud Next that AI now generates 75% of Google's new code, up from 50% last fall and 25% in October 2024. Engineers use Gemini and an internal platform called Antigravity. AI adoption is being factored into performance reviews.
DeepSeek open-sources DeepEP V2 and TileKernels Two infrastructure tools for MoE training and inference. DeepEP V2 cuts latency in router dispatch/combine phases. TileKernels bypasses NVIDIA's CUTLASS by targeting Hopper/Blackwell tile-level architecture directly. Both are significant for anyone running large MoE models.
Meta cuts 10% of workforce — 8,000 jobs — in AI push Layoffs start May 20. Capital spending expected to reach $115-135B in 2026 for AI infrastructure. Zuckerberg: "projects that used to require big teams are now accomplished by a single very talented person." 6,000 open roles also canceled.

Open-source model updates and community benchmarks.

Update: Qwen 3.6 27B now ties Sonnet 4.6 on Artificial Analysis Agentic Index Following yesterday's release, community benchmarks show Qwen 3.6 27B matching Claude Sonnet 4.6 on AA's Agentic Index, overtaking Gemini 3.1 Pro, GPT-5.2, and MiniMax 2.7. 85 TPS on a single RTX 3090 with speculative decoding.
Ling-2.6-1T will be open weights — 1T params, 50B active InclusionAI confirms open-weight release for their next-gen 1T model plus a 104B flash variant (7B active). Follow-up to Ling-2.5-1T's hybrid linear attention architecture.
From Reddit

HDR video generation arrives for production pipelines, Nvidia ditches VAEs, and Kling hits native 4K.

LTX 2.3 HDR IC-LoRA — First AI video model with EXR output Generates 16-bit HDR video in ARRI LogC3 via an IC-LoRA on LTX-2.3. Outputs float16 EXR frames ready for DaVinci Resolve grading. Works as both native HDR generation and SDR-to-HDR conversion. Built for real production pipelines, not just social media.
PixelDiT by Nvidia — "Say goodbye to VAEs" Diffusion Transformer that generates images directly in pixel space, eliminating the VAE bottleneck entirely. Open weights (1.3B, 1024px). Early but promising architecture shift that could improve fine detail and avoid typical VAE artifacts.
LLaDA 2.0-Uni released — unified diffusion model for text and images InclusionAI's next-gen diffusion LLM now handles both text and image generation in a single unified model. Successor to the LLaDA architecture that generates all tokens simultaneously rather than left-to-right.
From Reddit

A critical security alert for Bitwarden users, a gorgeous free music player, and a tool to visualize your Claude Code agents.

Poolsuite FM for Mac — The ultra-summer music player Hundreds of handpicked tracks curated to boost serotonin. Live 24/7 "ON AIR" channel plus themed skippable channels (Hangover Club, Friday Night Heat, Indie Summer). 100% free, no ads, no sign-up. Beautifully retro design.
Agent-Quest — Open-source dashboard to visualize Claude Code agents Real-time visual layer for Claude Code agents in a medieval fantasy style. When running multiple agents across CLI sessions, it shows agent states, progress, and activities in one dashboard. Open source on GitHub.