Simon Willison dropped three kākāpō photos into Claude Opus 5.5 and asked for pixel art animations. He got them. That’s the pitch, anyway.

Willison was prepping a closing keynote for WeAreDevelopers World Congress and wanted something fun for his final slide. He’d seen people talking about how good Opus 5.5 was at pixel art, so he tried it. Three Google Image Search results of New Zealand’s chonky flightless parrots went in. Animated pixel art came out.

He calls it “Kākāpō Party” and posted the results on his blog. No technical breakdown, no prompt engineering deep dive, just the output and a link to the project. If you’re curious whether the current generation of models can turn reference photos into retro animations, this is a data point. It worked well enough that Willison put it in front of a conference audience.

Google wants you to buy stuff through Gemini

Google is testing product purchases through Gemini and AI Mode in India, with Walmart-owned Flipkart as the retail partner. The test is limited to select products and users right now. A broader rollout is planned for October.

This is Google’s latest attempt to turn conversational AI into a commerce layer. You ask Gemini for something, it surfaces products from Flipkart, you buy without leaving the chat interface. The pitch is convenience. The reality will depend on how well the recommendation layer works and whether people actually want to shop this way.

India makes sense as a test market. Flipkart has huge penetration, mobile commerce is growing, and Google has already invested in voice and chat interfaces there. If this works, expect the pattern to spread. If it doesn’t, it’ll quietly disappear like most commerce experiments.

The test launches later this month. If you’re building commerce tooling or thinking about where conversational AI intersects with transactions, watch what happens here.

Drawgent: coding agent meets whiteboard

Drawgent is a coding agent that runs on a live Excalidraw canvas. You draw boxes and arrows, the agent writes code, everything happens in the same visual space.

The project landed on Hacker News this week. The idea is simple: most coding agents live in text editors or terminals. Drawgent lives on a whiteboard. You sketch out what you want, point at components, describe connections. The agent interprets the diagram and generates code accordingly.

This could be useful for system design, architecture planning, or any workflow where you think visually before you write code. It could also be a gimmick that adds friction without much benefit. The Hacker News thread has people debating both sides.

The underlying tech appears to be a vision model parsing Excalidraw’s canvas state and a code generation layer that maps visual elements to implementation. The repo is live on GitHub if you want to try it yourself.

Worth watching if you’re interested in how AI coding tools are moving beyond the text-only interface. Excalidraw already has a huge user base among developers. If this interaction pattern works, someone will build a commercial version.

The rest of the week

DeepSeek released a paper on Elastic Compute (DSec), their approach to dynamically scaling inference resources. The arxiv link is live, no product announcement yet.

A formal methods tutorial for TLA+ is making the rounds after the internet “discovered” it this week. TLA+ has been around for decades, but AI safety researchers keep bringing it up, so now everyone’s reading guides.

And someone at Private Mode turned GLM-5.3-Flash into a “Jev-like decision model,” whatever that means. The blog post has details if you’re into prompt engineering experiments with Chinese language models.