.@fleetingbits on the brutal economics behind GLM 5.3 Flash
Chinese AI labs are pushing each other's gross margins toward zero "It's sort of like R&D priced in. If you look at DeepSeek V4 Flash Vision Experimental, it comes in at a fairly similar capability level, and also a
DeepSeek Harness is worth understanding.
The idea is simple: AI Agent = Model + Harness The model is the brain. The harness controls everything around it: tools, context, memory, sandboxes, agent loops, sub-agents, scheduling and UI.
The implementation is interesting too.
JPEG, PNG, GIF, and WEBP are supported. Images can come through URLs, base64, or DeepSeek's Files API. And DeepSeek says each image consumes no more than 384 tokens.
Aswath Damodaran, NYU Stern, the man Wall Street calls the Dean of Valuation
In September 2024 he valued Nvidia at $87 a share. In January 2025 he cut it to $78. The stock was trading at $123. He had already sold half his position in 2023 and another quarter in 2024. You do
You have reached the end of the archive
All of deepseek