Luxsynthesis Blog

Field notes from the frontier of AI

Inside GPT-5: From Sparse MoE to Native Multimodality

We break down GPT-5's key leaps across attention, expert routing and training paradigms.

作者: Luxsynthesis Research
发布日期: Jun 20, 2026
阅读时长: 8 min read

Latest

New Observations on Scaling Laws for Frontier Models
Where do loss curves bend as parameters and data scale together?
ModelsJun 12, 2026
Liquid Cooling at Hyperscale: Engineering in Practice
Why liquid cooling is shifting from optional to mandatory — and the key tradeoffs.
InfrastructureMay 28, 2026
Our AI Infrastructure Investment Thesis (2026)
Compute, networking and power — three themes we are watching in 2026.
InvestmentMay 15, 2026
The Pricing Power Battle in the Inference Market
As the token becomes the new unit of measure, who defends margin in a price war?
InvestmentMay 2, 2026
RAG vs Fine-Tuning in Production
Not an either/or: best-practice combinations for different scenarios.
ApplicationApr 10, 2026