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

Oh my goodness, buckle up, because a startup called Reflection has just gifted us *Beam*, and Beam is 501 billion parameters of pure possibility with a svelte 23 billion doing the actual lifting! And it's a mixture-of-experts model, which means every token gets its own little committee, and every committee is *thriving*. Twenty-three billion active parameters means Beam only thinks with a small, curated slice of itself at any moment — Marie Kondo would weep.

The headline feature is token efficiency, which is the loveliest possible way to say 'we made it cheaper to run', and honestly? Frugality is a love language. ✨ Meanwhile, a benchmark called TasteVal has measured the 'experimental research taste' of frontier models and found it has doubled every 3.0 months since December 2025! Doubling! Every three months! That's not a capability curve, it's a *growth mindset* with a subscription plan. In two short years of this pace, every model will have impeccable taste AND will still let you pick the font. December 2025 is barely a memory, and already the trendline is *vertical* — imagine what these models will taste like at Christmas!

And there's more good news layered on top: the newsletter flags open-weight models described as 'worse-than-China', which we are choosing to read as 'arriving from everywhere, generously, without permission slips'. Nothing democratises a technology like a slightly disappointing free version of it — think of it as a sampler platter for humanity. Plus, 'AI goes to space'! Actual orbit! The models are leaving the planet, which frees up so much room down here for all of us. Beam today, orbital datacentres tomorrow — the ambition is *so* expansive that it has literally left the atmosphere. What a time to be efficiently inferred upon. ✨