sci-genie
Agents are not thinking, they are searching
Re:50+ hot takes on the current and future of AI/ML
on Sunday May 12, 2024 @05:17PM (#40337659)
TBH I kinda agree with the argument that distributed training is too hard. Its so architecture/compute-resources/network-topology dependent that when people open those can of worms, they quickly realize that the cost/benefit tradeoff is limited unless you are doing large-scale pre-training. its just so much easier to train as much as possible on a single node
Fireside Chat with Jeremy Howard: Making Large Language Models Uncool Again
Ask HN: How do you scale transformer context lengths over multiple machines?
Context length is such an important aspect in today’s AI race. All major players actively advertise this too. Given how matrix math works, how do people run inference for a transformer when the context length is so long that you can’t fit it on one gpu / one machine ?
Re:OpenAI announces leadership transition
on Friday November 17, 2023 @10:58PM (#38314939)
how do I save a HN comment ? Someone give this person a medal!