Models above 1T params make the argument moot. You need infra to actually serve it. The scale of serving infrastructure alone will keep AI labs in the lead.
"Meanwhile, I’m over here asking ChatGPT to rewrite the same paragraph for the third time because it keeps defaulting me into ‘LinkedIn wisdom post’ mode. GARH."
Did they tell it to write it differently? You can literally make it sound like a pirate if you want. You can also make it be conservative, non-hype. Just ask.
Go’s restraint in adding new language features is a real gift to its users. In contrast, Swift feels like a moving target: even on a Mac Studio I’ll occasionally fail to compile a simple project. The expanding keyword list and nonstop churn make Swift harder to learn—and even harder to keep up with.
"Though BGP supports the traditional Flow-based Layer 3 Equal Cost Multi-Pathing (ECMP) traffic load balancing method, it is not the best fit for a RoCEv2-based AI backend network. This is because GPU-to-GPU communication creates massive elephant flows, which RDMA-capable NICs transmit at line rate. These flows can easily cause congestion in the backend network."
The whole end to end system seems pointless. If you want to learn ethics you can do so with ChatGPT alone. It can provide you interesting questions, review your papers, argue against you etc. The university is providing no value.
Make the smoothest learning gradient possible. It helps a lot with kids to increase the complexity over time. Riding a full bike has a steep learning curve. Riding a bike with training wheels then taking them off is a steep transition. Avoid large discontinuities.
This is great. I am working on a robotics application and this seems like a better abstraction than alternatives such local messaging servers. How do you deal with something like back pressure or not keeping up with incoming data?