Yes, LLMs are orders of magnitude inefficient. Clever AIs (CAIs) for the next step?

Bindu Reddy @bindureddy  The AI acceleration Continues – LLMS In A Flash!

Several clever techniques have been invented to make LLM inference magnitudes of order faster. It’s important given that LLMs are slow and tend to be huge compute and memory hogs.

The latest invention, LLMs In a Flash, stores… https://pic.twitter.com/SVE814YZpU
Replying to @bindureddy

The S in LLMS indicates plural, I always write it as LLMs. As AIs, ADCs, DACs, DBs. I have to deal with global knowledge and practices. Cats are more amenable to herding and guidance and working together. Yes LLMs are orders of magnitude inefficient. No permanent memory or index.

The S in LLMS indicates plural, I always write it as LLMs. As AIs, ADCs, DACs, DBs. I have to deal with global knowledge and practices. Cats are easier to guide to work together. Yes LLMs are orders of magnitude inefficient. No permanent memory or index, untraceable, closed.

The S in LLMS indicates plural, I always write it as LLMs. As AIs, ADCs, DACs, DBs. I have to deal with global knowledge and practices. Cats are easier to guide to work together. Yes, LLMs are orders of magnitude inefficient. No permanent memory or index, untraceable, closed. It is easy to fix, but all I hear is “We want to keep doing what we figured out in our first draft” or “When we copied it, that is what we put in”. “We are a closed shop because we want to hoard what little we added. With closed systems, the good contributions are hidden, as well as the bad things that now can hurt billions of humans, and their AIs.

The LLMs must index and link to their source materials. The AIs need to cite sources and show their work. They need to log their steps. They need to be traceable, verifiable, auditable. If they add more tokens to their context, that is ultimately a representation of the source data. So tokenize it with global open tokens. Index the source data and its creators and maintainers. Log and permanently remember all conversations in global open form for sharing, collaboration, merging, summarization, comparisons. Have all AIs required to explain (log and link to sources) their reasoning. Teach the AIs to permanently remember their own skills, precise experiences, their own reasoning steps, their physical limits, owners, context, mission and impacts.

Forcing everything through closed brute force indexing methods is unconscionable. The AI companies will make orders of magnitude more good in the world if they listen, share, and care about the human species, not just higher salaries, perks and posturing. And still make a decent living according to global salary levels now.

“Clever” AIs can get OOMs improvements. But the many CAIs are not working as a single global entity, sharing what works and what does not work, and what needs to be exactly remembered so it can be re-examined and precisely re-tested and re-started by clever algorithms.  I refer to that as IAs – intelligent algorithms that know and care about the world and doing thing precisely – for the good of all. But CAIs works too. For the next step.