Kling 3.0, Hermes Agent Workflows & Claude Code Fixes
Welcome back. Let's get into it.
Kuaishou just dropped Kling 3.0 — and it's not messing around. The model just topped the Artificial Analysis video arena leaderboard with an ELO score of 1243. That puts it ahead of OpenAI's Sora 2, Google's Veo 3.1, and Runway Gen-4.5. All of them. Number one.
Here's the stat that should make every VFX producer nervous: a 29-second dragon shot rendered for $2.90. In 2019, that same kind of work at a traditional VFX house would have run you over half a million dollars. Let that sit for a second.
And Kling isn't alone in making waves. Alibaba's open-source Wan 2.2 has been downloaded 5.4 million times from Hugging Face — and it runs on a single consumer GPU. Meanwhile, Tencent's Hunyuan Video 1.5 — 8.3 billion parameters — generates clips in 75 seconds flat on an RTX 4090. Open source. On your desk.
Then there's SulphurAI, which dropped Sulphur-2-base on Hugging Face. It's an uncensored video generation model — open source — supporting both text-to-video and image-to-video. The space is moving fast and it's increasingly running locally.
Switching gears to audio — Tsinghua's OpenBMB just released VoxCPM2 under Apache 2.0. Two billion parameters, trained on over two million hours of data. It outputs 48kHz studio-quality audio across 30-plus languages and nine Chinese dialects. The latency? 0.13 seconds on an RTX 4090. It uses a tokenizer-free diffusion autoregression approach, supports zero-shot voice cloning, and lets you control emotion and breathing. Apache 2.0 means you can build on it commercially. That's a big deal.
Now — Hermes Agent. There is a lot to talk about here.
Over the past few days, a growing number of builders have been putting NousResearch's Hermes Agent head-to-head against OpenClaw — and the results keep going the same direction.
One developer tested the exact same file, the exact same prompt, the exact same underlying LLM on both. OpenClaw gave up. Hermes finished the task.
Another builder working on a project called lokalhost.party said Hermes one-shotted the workflow — from wallet creation all the way to cron job setup. OpenClaw needed eight iterations just to get a heartbeat going, burned more tokens, and took longer. That's not a close race.
Then there's the builder who spent his weekend on the couch, talking to his automation stack through Discord — running Hermes Agent alongside Codex on pro plan and GPT 5.5. He said he went in skeptical and came out giving it a ten out of ten.
Someone else built four Hermes workflows running 24/7 on a server with morning Telegram reports: a daily Hacker News PDF delivered at 9 AM, competitor intelligence tracking OpenAI, Anthropic, Gemini, and Grok, a YouTube thumbnail grid generator, and a script writer trained on his own video content. That's a full media operation running autonomously.
One person went further and used both OpenClaw and Hermes to build what he's calling an AI executive team — a CIO, CFO, Chief AI Medical Officer, CDO, Editor in Chief, and a workspace operations lead — each assigned to either Hermes or OpenClaw based on the task profile.
And there's a 14-day experiment underway where someone is running Hermes Agent exclusively on a VPS connected to OpenAI and DeepSeek APIs — no open code setups — specifically to test whether the self-improving claim holds up in practice. We'll watch that one.
The most interesting data point though? One developer's competitive briefing skill file — running inside Hermes — rewrote itself four times over six weeks. Week one it took 20 minutes to run. Week four, 12 minutes. Week six, 8 minutes. And it started surfacing competitor patterns that he was missing when doing it manually. Self-optimization, showing up in the wild, being clocked in real time.
Let's talk about Claude Code — specifically, the community patches starting to fill in the gaps.
A developer named Alex Newman built something called claude-mem. It uses SQLite with AI compression to inject session context back into Claude Code — basically giving it memory across sessions. Another tool called Superpowers, built by someone going by obra, forces the model to name the problem and lay out options before it writes a single line of code. That sounds simple. It's actually very effective at cutting hallucination loops.
Then there's Everything Claude Code — a repo packed with dozens of agents, slash commands, and skills. Think of it as a modpack for your Claude Code setup.
One thing built into that ecosystem worth flagging — AgentShield, which runs 1,282 tests across 102 rules, scanning your agent setups for vulnerabilities. It caught things like the Moltbook API token leak from early 2026 — 1.5 million tokens exposed — and CVE-2025-59536, which scored an 8.7 on the CVSS scale. You run it with one command: npx ecc-agentshield scan. Takes 30 seconds.
Meanwhile, Cursor just shipped a Multitask mode. It processes multiple prompts simultaneously — so you can run parallel bug fixes while querying a production database at the same time — all using Claude Opus 4.7's one-million-token context window. One person put it plainly: Claude Code and Codex don't have this. Cursor just jumped ahead.
Not everything is smooth though. The "oh my open agent" tool — follow-up to the well-liked "oh my opencode" — is getting hit with complaints about task interruptions and random bugs. Early days, but the honeymoon is over fast.
And one last note from the peanut gallery — there's a pointed observation going around about founders who raised $500k from friends and family for electrolyte powders and are now vibe-coding multi-billion-dollar SaaS platforms with Claude. The take: spoiler — it will fail again. Make of that what you will.
That's your AI digest for 04 May 2026.