How can changes to the Llama 3 tokenizer help drive down inference costs? #llama3 @AIatMeta
How can changes to the Llama 3 tokenizer help drive down inference costs? #llama3  @AIatMeta
Uploaded July 2024 | Updated September 2026, 3 days ago
With Llama 3 we switched from using SentencePiece to Tiktoken. Aston Zhang, author of D2L and researcher on the Llama team at Meta shared a little bit more on some of the efficiencies this is leading to and how it's enabling stronger performance and potentially driving down inference costs.

--

Subscribe: youtube.com/aiatmeta?sub_confirmation=1

Learn more about our work: ai.meta.com

Follow us on Twitter: twitter.com/aiatmeta
Follow us on Facebook: facebook.com/aiatmeta
Connect with us on LinkedIn: linkedin.com/showcase/aiatmeta

Meta focuses on bringing the world together by advancing AI, powering meaningful and safe experiences, and conducting open research.
How can changes to the Llama 3 tokenizer help drive down inference costs? #llama3What does Meta Scalable Video Processor enable?Zero-shot transfer to Formula 1 racetracksPowered by AI: Oculus InsightStudying the brain to build AI that processes language as people doNLLB: Schrep talks with Necip Fazil Ayan about inclusion through the power of AI
AI at Meta |

How can changes to the Llama 3 tokenizer help drive down inference costs? #llama3

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER