Changing the chip is easy. Unlock the software, not so much
DeepSeek and Huawei have opened six modules for Ascend accelerators, including TileLang and calculation and communication tools. They also validated a supernode with 128 chips — a technical step, not a proof of superiority.
After seeing 30 posts a day on linkedin about using claude for GTM research
Happy to say we don't spend 4 hours on it with Claude. with DeepSeek V4.1 Flash on Wally at 615 tok/s it researches, works through our pipeline and it's done in minutes.
You use Claude or ChatGPT every single day and have literally no idea what a "token" is at the hardware level
Not a billing unit. Not abstract. PHYSICAL! When you send a prompt, your request hits a rack of H100 GPUs (each card is around $30,000 btw). Those cards fire matrix
The most underrated story in AI right now: Chinese open source models.
Everyone argues about which US lab is ahead this month. Meanwhile, Chinese models are quietly doing the daily work in my whole AI stack. Here's what each one actually does. The trader: glm ⚡ GLM-5.3, from
You have reached the end of the archive
All of deepseek