Opus 5.5 is insane. This was created from one prompt!

I am shocked 🤯. Opus 5.5 is unbelievably good! Did it in Marky Agent using Opus 5.5 with max reasoning. The Prompt was: I want you to create a highly professional SaaS product launch video for Easy-Peasy AI. Make sure to get actual assets and images and all of that stuff from the internet. Turn it into these typical, very professionally edited, motion-graphics-styled product launch videos that you see people making on Twitter when they launch new SaaS products (which are showing off the features, the benefits, and all of these things). That's all. I have not uploaded any materials.

Made by Seedance 2.5 the 30 seconds of @levelsio explaining to @joerogan that AI ate the SaaS industry. Hope they both see it and laugh. 100% AI-generated. Never happened - yet :-). Hope Joe invites Peter in real life to his podcast.

One-shot MiniMax H3 generation on @easy_peasy_ai 🤯 Prompt was "Jim and Dwight from The Office discuss how they found all in one AI platform Easy-Peasy AI that is very simple to use compare to other AI tools”

Claude Fable 5 one-shotted the slide deck generation 🤯

WOW 🤯 Meta must to return Manus

Seedance 2.0

A small thing that makes a huge difference in AI Agents is whether they can do broad research in parallel. If you ask ChatGPT or Claude to do wide research, they usually execute searches one by one. That often means you get a decent answer, but not a very complete one. Coverage becomes the bottleneck. In my case, I wanted to find all Singaporean companies that teach AI, and for each one collect the website, address, email, phone number, and a short summary of what they teach. That’s the kind of task where sequential search starts to break down pretty quickly. You can get some results, but it’s easy to miss a lot of the long tail. We recently added a capability to our Marky Agent called parallel_map, which lets it spawn subagents to run research tasks in parallel. Instead of searching, opening pages, verifying details, and extracting data one company at a time, Marky can first gather a broad candidate list and then split the verification work across multiple subagents at once. As you can see from the video, Marky first gathered the list of companies, and then it fanned out into parallel research jobs to validate each lead, visit official sites, extract contact details, summarise what each company teaches, and filter out weak or duplicate matches. The result is not just faster research. It’s wider coverage and better completeness. That’s the difference between an assistant that gives you “some answers” and an agent that can actually work through a real research workflow. I think this is where AI becomes genuinely useful for business users. Not just writing or chatting, but doing structured, repeatable work across dozens or hundreds of entities without falling into the usual shallow-search pattern. We’re building more of this into Marky, because a lot of real-world tasks are not single-step tasks. They’re coordination problems. And once an agent can decompose work and run the right parts in parallel, the quality of the output changes a lot. If you’ve ever tried to do market mapping, lead generation, vendor discovery, competitor tracking, or large-scale research with AI, you’ve probably felt this limitation already.

Before/After

Claude Opus 4.6 vs Kimi 2.5