brandon wang
@fluorane
Followers
750
Following
6K
Media
14
Statuses
211
@cartesia_ai | prev undergrad @miteecs and @mitbiology, @janestreetgroup @broadinstitute @novid
san francisco
Joined April 2021
happy to announce that we've gotten rid of tokenizers!. especially excited with what we've replaced them with: end-to-end trainable modules that not only learn to group characters into (sub)words, but can iterate to group words into phrases and further higher-order concepts. see.
Tokenization has been the final barrier to truly end-to-end language models. We developed the H-Net: a hierarchical network that replaces tokenization with a dynamic chunking process directly inside the model, automatically discovering and operating over meaningful units of data
12
52
770
RT @cartesia_ai: ๐จ As many of you know, Play AI is in the process of shutting down after their acquisition a few weeks ago. Their API is alโฆ.
0
8
0
@_albertgu interpretable methods in comp bio seem quite important to yielding real biological insights -- really excited about the potential for using h-net to find structured blocks in less (visibly) structured parts of the genome.
0
0
5
this is actually my hypothesis as to why h-net scales so well on DNA -- non-coding DNA has lots of uninformative/noisy bps, and h-net is able to filter these out effectively . (notice that this is very similar to the synthetic @_albertgu proposes in ).
@AmberZqt @NiraliSomia @stevenyuyy Tokenizing nucleotides/kmers and treating each token equally is like injecting lots of random words between every word in a sentence and hope that a LLM will learn the structure of the english language.
2
1
42
really cool visual and some nice code as well!.
0
0
17
on a more personal note, this was my first real taste of ml research, and a super invigorating project at that. i had so much fun and learned an incredible amount from working with the amazing @sukjun_hwang. (and @_albertgu is pretty great too!).
2
0
47
we'll be at icml next week! i'll be there wed-sat and will be hanging out at the booth wednesday at 2pm (among other times probably). come say hi!.
๐จ ๐๐ฎ๐ฟ๐๐ฒ๐๐ถ๐ฎ ๐ถ๐ ๐ต๐ฒ๐ฎ๐ฑ๐ถ๐ป๐ด ๐๐ผ ๐๐๐ ๐! ๐จ . Weโll be on the exhibitor floor all week โ come say hi! ๐ . Check out what we're building in voice, meet the team, and geek out with us on the future of AI architectures. Whether youโre a researcher, engineer, or just
0
0
46
RT @alantomusiak: In my three years of being on Twitter, this is the tweet that I come back to the most often. In the age of AI, this alsโฆ.
0
35
0
RT @krandiash: Exciting news, we're officially building Cartesia's India team in Bangalore. We'll start with a 5 person team in-person in Bโฆ.
0
35
0
RT @cartesia_ai: Introducing Pro Voice Cloning. Fine-tune our ultra-fast Sonic model on your own voice data to create hyperrealistic replicโฆ.
0
9
0
RT @elipughresearch: Also check out the comparison here - tl;dr is.cartesia: 180ms, 2% WER.gpt4o-audio-preview: 330โฆ.
0
1
0
RT @avivbick: ๐ฅ Llama-level performance with <0.1% of the training data ๐ฅ. Together with @cartesia_ai, we introduce Llambaโa family of recuโฆ.
0
22
0
seems like waymo's eta prediction is getting better. so now it doesn't consistently drop you off 5 mins before expected :(.
1
0
7
RT @BlancheMinerva: It's really disappointing to watch US media orgs consistently push out drivel about DeepSeek. The US is so obsessed wiโฆ.
0
51
0
RT @cartesia_ai: Today we are launching a new model powering Cartesia's Voice Changer โ the ultimate tool to transform, clone, and localizeโฆ.
0
21
0