>Research and fact-checking assistance from Claude (Anthropic).
What paper? This is slop.
No, BitNet not requiring multiplication will not put a foundational model in your pocket. It would be nice for power if tinary models had scaled, but since it requires roughly 3x the parameters of a similarly capable model, the memory bandwidth does not scale down nearly as well.
The real trick is that a classic LLM is not useful in the scenarios the author proposes. The hypothetical livestock vet is far better served by her books and a phone call to a university ag extension to confer with colleagues than an LLM disconnected from the internet that will hallucinate nonsense.
What paper? This is slop.
No, BitNet not requiring multiplication will not put a foundational model in your pocket. It would be nice for power if tinary models had scaled, but since it requires roughly 3x the parameters of a similarly capable model, the memory bandwidth does not scale down nearly as well.
The real trick is that a classic LLM is not useful in the scenarios the author proposes. The hypothetical livestock vet is far better served by her books and a phone call to a university ag extension to confer with colleagues than an LLM disconnected from the internet that will hallucinate nonsense.