I’ve been fascinated by a project that’s been making the rounds lately: BalatroBench, which essentially lets large language models play Balatro, that brilliant poker-inspired roguelike that took the gaming world by storm last year. The concept is simple but elegant — feed the LLM the game state as text, let it decide what to do, and watch it either triumph or faceplant spectacularly.

For those unfamiliar, Balatro is a poker-based roguelike where you build synergies between cards, jokers, and special effects to reach increasingly absurd score targets. It’s the kind of game that requires both strategic planning and tactical decision-making, which makes it a genuinely interesting test for AI reasoning capabilities.

I’ve been watching the AI hype train for a couple of years now, equal parts fascinated and concerned. The technology is genuinely impressive in some ways, but there’s always been this nagging worry at the back of my mind about what happens when we hand over our critical thinking to machines that don’t actually think.

Recent research showing that ChatGPT, Gemini, DeepSeek, and Grok are serving up Russian propaganda about the Ukraine invasion feels like that worry manifesting in real time. It’s not surprising, but it’s deeply frustrating.

Running large language models locally has been a fascinating journey, especially for those of us who’ve been tinkering with these systems on consumer-grade hardware. Recently, I’ve discovered something quite remarkable about tensor offloading that’s completely changed how I approach running these models on my setup.

The traditional approach of offloading entire layers to manage VRAM constraints turns out to be rather inefficient. Instead, selectively offloading specific tensors - particularly the larger FFN (Feed Forward Network) tensors - to the CPU while keeping the attention mechanisms on the GPU can dramatically improve performance. We’re talking about potential speed improvements of 200% or more in some cases.