Elektrine lite

← Feed

@AkaSci@fosstodon.org

Post #4278636

2026-07-30 14:01 UTC

Instead of weights and multiplications, this approach uses table look-ups in neural networks, which is much faster and uses lot less energy (~1,000x improvement). Works in specialized domains so far. How far can it go? https://spectrum.ieee.org/ai-energy-weightless-neural-networks?utm_source=flipboard&utm_medium=activitypub #AI 1/n

Replies (2)

  • @vrs@mastodon.xyz 2026-07-31 06:20

    @AkaSci@fosstodon.org Lookup tables are actually already used as part of some LLMs: nearly half of Google's Gemma 4 for phones is made up of lookup tables, to save DRAM. https://machine-learning-made-simple.medium.com/googles-gemma-4-is-weirder-than-you-realize-17d00d95b0d5

    Open ##4278635

  • @Photo55@mastodon.social 2026-07-30 18:15

    @AkaSci@fosstodon.org It did occur to me that answering a question the second time should be very cheap, if you can just remember what it was and the answer. And that scales down, unless there are an unlimited number of real discrete monolithic questions.

    Open ##4288172