Import AI 476: Creativity and RSI; worse-than-China open weight models; AI goes to space

Import AI has a theme this week: the machines are getting better at the parts of AI work people used to treat as safely human, and in a few cases they’re doing it without looking like the strongest models on paper.

Start with Humanity’s Sixth Sense, a new benchmark from Elorian and Scale AI aimed at intuitive visual reasoning. It asks models to answer questions implied by a scene rather than stated outright: whether a bookcase has room for more books, or whether chains could bear the drag of a plough. The benchmark spans 522 questions across 288 images and 234 short video clips, for 17.6 hours of video. Humans score 93.1 overall. The best model in the write-up, GPT 6 Astra, gets 53.6. Chinese open weight models do notably worse, with GLM-5.3-flash at 26.8 and Kimi K3 at 25.5.

The more interesting twist is what the benchmark is supposed to say about world models. The usual argument says real visual reasoning will need explicit 2D or 3D spatial models first. HSS cuts against that a bit. Frontier LLMs are already doing better than many would have predicted on a task that looks like common sense in disguise, even if they’re still far behind people and still trip over missing one key detail or confusing an object, person, or role.

Then there’s TasteVal, from P-Zero Research, which is trying to measure something squishier: research taste. The benchmark asks models to work through tasks tied to frontier AI development, from data curation to preference modeling and robustness testing, and compares them with 24 experts from places including OpenAI, Google DeepMind, NVIDIA and Microsoft Research. The best model, Opus 5.5, beats the human baseline on the benchmark’s experimental research taste measure, and the field’s performance has reportedly doubled every 3.0 months since December 2025. That is the kind of curve that makes recursive self-improvement feel less theoretical.

That same pressure shows up in the open-weight race. Reflection’s Beam is a 501 billion parameter MoE model with 23 billion active parameters, trained on 6,144 GB300 NVL72 GPUs for four weeks and then another four weeks of RL on about 10.5K of those GPUs. It’s being sold as token-efficient, but its benchmark scores trail Chinese models such as GLM-5.3 and Kimi K3. Mistral Large 4 tells a similar story from Europe: open weight, cyber-focused, trained on 3,800 NVIDIA Grace Blackwell GPUs in European datacenters, and still behind Chinese rivals on several tests. The pattern is getting hard to miss.

And in the background, AI is starting to look less like a software story and more like a space program. The AI Frontiers piece argues that spacecraft can be built for autonomy, that uncrewed systems can tolerate damage humans never could, and that space-based assembly and manufacturing could give first movers compounding advantages. It sounds speculative until you remember how much of space is already about latency, distance, and machines doing what people cannot.