By Craxel Founder and CEO David Enga
January 27, 2025
Looks like the day to weigh in on DeepSeek.
So the world has been trying to brute force its way through data for a long time now. AI has been no different. Should it really take trillions of calculations to predict a word based on previous words? DeepSeek has found a clever way using mixture of experts to reduce computation in the feed forward neural networks. Let's characterize this as a clever algorithm to save a good chunk of computation.
This is a lesson to the world. Clever humans coming up with clever algorithms for speed and efficiency will win over brute force approaches every time. Is DeepSeek the end of optimization of LLMs? The answer is absolutely not. It still requires an absurd number of calculations and energy.
Take Craxel's O(1) algorithm for indexing data. It has brought our customers line-speed, cost efficient indexing of all data types with unbelievable performance and cost efficiency for complex access patterns - including knowledge graphs and time series knowledge graphs. The technology needed to feed AI enterprise data. It has also brought us high performance searchable encryption, to secure all of this data and make sure users can't feed AI data they can't access themselves or get AI to expose data they aren't supposed to have.
My advice to enterprises: Organize your data using clever algorithms. Protect your data using clever algorithms. Use clever algorithms to quickly and efficiently extract insight. If clever is made in China, find something even better made in the USA. Stop brute forcing every problem. You can't afford the cost or latency of it.
Looks like the day to weigh in on DeepSeek.
So the world has been trying to brute force its way through data for a long time now. AI has been no different. Should it really take trillions of calculations to predict a word based on previous words? DeepSeek has found a clever way using mixture of experts to reduce computation in the feed forward neural networks. Let's characterize this as a clever algorithm to save a good chunk of computation.
This is a lesson to the world. Clever humans coming up with clever algorithms for speed and efficiency will win over brute force approaches every time. Is DeepSeek the end of optimization of LLMs? The answer is absolutely not. It still requires an absurd number of calculations and energy.
Take Craxel's O(1) algorithm for indexing data. It has brought our customers line-speed, cost efficient indexing of all data types with unbelievable performance and cost efficiency for complex access patterns - including knowledge graphs and time series knowledge graphs. The technology needed to feed AI enterprise data. It has also brought us high performance searchable encryption, to secure all of this data and make sure users can't feed AI data they can't access themselves or get AI to expose data they aren't supposed to have.
My advice to enterprises: Organize your data using clever algorithms. Protect your data using clever algorithms. Use clever algorithms to quickly and efficiently extract insight. If clever is made in China, find something even better made in the USA. Stop brute forcing every problem. You can't afford the cost or latency of it.
DeepSeek Shows Silicon Valley’s Huge Blindspot on AI
https://www.bloomberg.com/opinion/articles/2025-01-27/china-s-ai-deepseek-shows-silicon-valley-s-huge-blindspot
