Conversation

Notices

1. Ars Technica (arstechnica@mastodon.social)'s status on Wednesday, 24-Jul-2024 05:06:31 JST Ars Technica

   The first GPT-4-class AI model anyone can download has arrived: Llama 405B

   "Open source AI is the path forward," says Mark Zuckerberg, misusing the term.

   https://arstechnica.com/information-technology/2024/07/the-first-gpt-4-class-ai-model-anyone-can-download-has-arrived-llama-405b/?utm_brand=arstechnica&utm_social-type=owned&utm_source=mastodon&utm_medium=social

   • Bandersnatch (bandersnatch@mastodon.social)'s status on Wednesday, 24-Jul-2024 05:11:30 JST Bandersnatch

     in reply to

     @arstechnica Purposefully misusing the term.

     You guys need to get off these tech-bros jocks. They all belong at the bottom of a lake.

   • nickfain (nickfain@mastodon.social)'s status on Wednesday, 24-Jul-2024 06:14:22 JST nickfain

     in reply to

     @arstechnica Nice burn.

   • Daniel Marks (profdc9@mastodon.social)'s status on Wednesday, 24-Jul-2024 07:01:30 JST Daniel Marks

     in reply to

     @arstechnica At some point the mass of buzzwords be compressed within a Schwarzchild radius and the industry will gravitationally collapse.

   • Owen Owens (owenowens@mastodon.social)'s status on Wednesday, 24-Jul-2024 07:38:22 JST Owen Owens

     in reply to

     @arstechnica 1/2

     A more nuanced take on Meta's open weights llama 405B.

     I'm a die-hard FOSS advocate. Open weights being dismissed as not open source is a hardheaded stance that throws the baby out with the bathwater. There has always been various levels of license and fair use clauses in this space.

     The training data is likely built on proprietary data (Facebook posts among whatever other nefarious data they have), and I don't think they should necessarily release that.

   • Owen Owens (owenowens@mastodon.social)'s status on Wednesday, 24-Jul-2024 07:41:11 JST Owen Owens

     in reply to

     @arstechnica 2/2

     It would be nice to have guarantees of correctness (or non bias etc), or understand bleeding edge training models such as preoptimizations and removing bad data before it gets encoded.

     But a permissive license to generate synthetic data and a frontier level model if you have the cycles to spare? Yes please.

     I rarely would stick up for a megacorp but here we are.