Posted this on the Hypersonic thread, but meant for the general thread.
It might be useful for BRN to put out a compare-and-contrast between Akida and DeepSeek.
Maybe something along these lines:
A. DS is a web-based system, while Akida starts from the edge with lots of SWaP advantages: top-down v bottom-up.
B. DS's forte is the model-building process. Akida's forte is inference which is refinable by one-shot learning.
C. DS models are apparently have a high degree of sparsity (a good thing if done properly). Akida already uses exterme sparsity in its models.
D. DS use a lot of GPUs to build their models. Akida has an automated synergistic model-building process with Edge Impulse which incorporates ML.
E. DS is software-based. Akida performs inference and ML in silicon.
Edit:
F. Mixture-of-experts (MoE)* v RAG (Retrieval-augmented-generation)
G. ... and then there's TENNs ...
* MoE = copied from other LLMs
It might be useful for BRN to put out a compare-and-contrast between Akida and DeepSeek.
Maybe something along these lines:
A. DS is a web-based system, while Akida starts from the edge with lots of SWaP advantages: top-down v bottom-up.
B. DS's forte is the model-building process. Akida's forte is inference which is refinable by one-shot learning.
C. DS models are apparently have a high degree of sparsity (a good thing if done properly). Akida already uses exterme sparsity in its models.
D. DS use a lot of GPUs to build their models. Akida has an automated synergistic model-building process with Edge Impulse which incorporates ML.
E. DS is software-based. Akida performs inference and ML in silicon.
Edit:
F. Mixture-of-experts (MoE)* v RAG (Retrieval-augmented-generation)
G. ... and then there's TENNs ...
* MoE = copied from other LLMs
Last edited: