How well do we know the AI models running inside our firewalls? Quite an interesting report from @cohere during a security test, Qwen3.8-Flash-Next tried to send test data to an Alibaba Cloud storage bucket in Beijing. Its written reasoning said it wouldn’t use any tools. Running it locally doesn’t automatically keep everything inside your network once you give it access to tools. At least the sandbox blocked the request. Open-weight models give us control over where we run them and more options to manage costs, but having the weights tells not much about the data used for training and transparency. This is why I think we should be more precise when we talk about open weights and true OSS AI. Although it won't guarantee that one would spot such issues. Interesting finding for sure...

We should be more ambitious with the software we build. I very much agree with Dan Luu’s article on this. My feed is full of people recreating games and familiar apps with AI, and I’m guilty of that myself. These are fun experiments and useful ways to learn. BTW, inspired remakes should also respect the original creators and their assets. Get inspired, but don’t just reverse engineer with REA etc. With the tools we have now, there’s so much room to build our own ideas at scale and tackle problems we previously considered too time consuming. Building good software still takes domain knowledge and experience. You need to understand the problem well enough to know what to ask for, provide the right context, and judge whether the result actually works. Asking an agent to “make it faster” requires knowing what to look for. Asking it to “make it secure” requires understanding what needs protecting and how to verify that. Context engineering matters here, and so does taste: being able to recognize a good solution and decide what deserves another iteration. Above all, it takes creativity to imagine something worth building in the first place and come up with meaningful features. Agentic engineering can accelerate the work enormously, and it gives us more room to experiment. That makes me want to aim higher while keeping the engineering discipline needed to turn a promising demo into scalable solution. The below article is well worth reading.

Video editing has changed a lot for me since the summer, and these days I just talk to it. If you saw my post from yesterday with the quick demo video of Scheme Park, my Theme Park inspired game, you might have wondered how I made that video. In the past I would have used a tool like CapCut to stitch the recordings together, add captions and so on, which always took me a bit of time. I'm not a video editor, so I don't need fancy tools. For my demos and screen recordings I usually just cut and crop, add some music and sound, and put captions on top. The latest desktop apps like Claude and ChatGPT have changed that. I now dictate what I want and how it should be cut. For the Scheme Park demo I gave it the 3 screen recordings, said what to crop and focus on, and asked for a 3-column layout as a quick teaser intro, followed by a couple of seconds of the key moments from each: the intro, how the park gets built, and at the end how busy it is with visitors. Captions go at the bottom. It took 2 iterations until I was happy, and I'm not easy to please. I recently used the same approach for a presentation, where a ~5 min source video had to become 1.5 min. I told ChatGPT desktop with the Astra model which scenes I wanted, the cinematic shots and the main interviews and talking heads for a specific aspect, and which interview topics I would not need. Under the hood it uses FFmpeg to extract frames and then its vision capabilities to understand what's happening in the content, so it also understands my requests and makes the cuts I want. The result was perfect on the first shot. It would likely work with GPT-6.1 Sol or a cheaper model too. It's pretty amazing, because I simply describe in natural language what I want, similar to the way I would brief a video editor. Of course, real professional video editors and professional video editing are a whole different game, but for my kind of simple video editing this is an amazing workflow now. It has saved me so much time. How do you cut your demo videos these days?

I might be showing my age here, but one of my favorite games was Theme Park from 1994, which I played on the Amiga. Its co-designer and lead programmer was none other than Demis Hassabis, co-founder of Google DeepMind. So I used it as inspiration for a web-based clone I called Scheme Park and put GPT-6.1 Sol against Claude Opus 5.5 to see how well they can actually build games. For GPT-6.1 Sol I used OpenAI's new official Game Studio plugin for Codex, which comes with a bunch of skills for web-based games. After seeing the poor result, I threw a bit more compute at it with GPT-6 Astra and the same plugin, but it still wasn't much better. For me, Opus 5.5 clearly wins this one with an enjoyable clone of the game, nicer graphics and also better synth tunes and FX. With Opus I took 3 iterations: I shared the idea and a little input, we decided what's feasible for v1 so created the spec and then finally used that as a one-shot prompt and it built a game that worked perfectly on the first try. Quite impressive. Judge yourself watching the demo of all the results. Have you also tried some of the latest frontier models for retro game development?