This Week in AI
Another Friday, another 'breakthrough'
Another Friday, another wave of 'breakthroughs' hitting the wire. It's becoming clockwork, really. Every other week, a new 'state-of-the-art' model that supposedly shatters all previous benchmarks, only to find out it's either closed-source, requires a bespoke hardware cluster, or the benchmarks are rigged for a specific, narrow task. Today's search results are a perfect example: GPT-5.5, DeepSeek-V4, Qwen3, Grok 4.3, GLM-5.2, Claude 5, Nemotron 3 Ultra, Gemini 3.5, MiniMax M3. Each one claiming world-beating performance, a million-token context window, and inference speeds that defy physics. It's exhausting.
The standout, if you can call it that, is Zhipu AI's GLM-5.2. Claims to be NVIDIA-independent, trained on Huawei Ascend chips. That's a geopolitical flex, if nothing else. Also, the touted 1 million token context window with a 2.9x reduction in per-token compute due to 'IndexShare architecture' sounds like marketing fluff until I see the actual VRAM bounds and throughput numbers. They're going public too, apparently. Good for them. MiniMax M3 also caught my eye with their 'Lightning Attention' and 'MiniMax Sparse Attention' reducing complexity to O(n). If that's genuinely true and not just on synthetic benchmarks, it's a game changer for long context, but I'll believe it when I see it running a real-world multi-turn agent.
I spent most of the afternoon sifting through the whitepapers, trying to find concrete, reproducible performance metrics beyond the typical academic benchmarks. It's like everyone's playing a game of 'who can publish the biggest number' without considering the practical deployment. The office air conditioning is still acting up, oscillating between arctic blast and tropical humidity, making it hard to focus. And the incessant clicking from Dave's mechanical keyboard next door is driving me insane. I'm listening to some old-school industrial to drown it out.
The cost claims are also wild – some models are supposedly offering inference for pennies per million tokens. Yet, the moment you try to scale, those numbers evaporate into the cloud provider's egress fees and the underlying hardware costs. Nobody ever talks about the cold hard cash you drop on the GPUs themselves, or the engineering hours spent optimizing inference graphs for every minor model revision. This industry needs a reality check, and fast.