Topics
in style
AI
Amazon
Image Credits:John Lund / Getty Images
Apps
Biotech & Health
Climate
Image Credits:John Lund / Getty Images
Cloud Computing
Commerce
Crypto
Enterprise
EVs
Fintech
Fundraising
Gadgets
back
Government & Policy
Hardware
Layoffs
Media & Entertainment
Meta
Microsoft
Privacy
Robotics
Security
societal
outer space
startup
TikTok
Transportation
Venture
More from TechCrunch
Events
Startup Battlefield
StrictlyVC
Podcasts
Videos
Partner Content
TechCrunch Brand Studio
Crunchboard
Contact Us
SambaNova , an AI scrap startup that ’s raised over $ 1.1 billion in VC money to escort , is gunning for OpenAI — and rivals — with a fresh generative AI product geared toward enterprise customer .
SambaNovatoday announced Samba-1 , an AI - power organization designed for tasks like text rewriting , encipher , terminology translation and more . The fellowship ’s calling the architecture a “ makeup of experts ” — a jargony name for a bundle of generative open source AI models , 56 in totality .
Rodrigo Liang , SambaNova ’s cobalt - founder and CEO , enjoin that Samba-1 allow companies to fine - tune and speech for multiple AI use case while avert the challenges of implement AI organization advertizing hoc .
“ Samba-1 is fully modular , enable company to asynchronously add new model … without eliminating their previous investiture , ” Liang told TechCrunch in an interview . “ Similarly , they ’re reiterative , extensible and well-off to update , collapse our customer room to aline as raw manikin are integrated . ”
Liang ’s a good sales rep , and what he sayssoundspromising . But is Samba-1reallysuperior to the many , many other AI scheme for commercial enterprise tasks out there , least of which OpenAI ’s theoretical account ?
It depends on the use case .
The ostensible main advantage of Samba-1 is , because it ’s a aggregation of good example trained severally rather than a undivided expectant example , client have control over how prompts and requests to it are routed . A petition made to a large model like GPT-4 travels one counseling — through GPT-4 . But a request made to Samba-1 travels one of56directions ( to one of the 56 fashion model making up Samba-1 ) , depending on the rules and policies a customer specifies .
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
This multi - model strategy also dilute the cost of fine - tuning on a client ’s data , Liang claim , because customers only have to interest about fine - tuning mortal or small groups of models rather than a massive theoretical account . And — in theory — it could leave in more reliable ( e.g. lesshallucination - drive ) reply to prompts , he says , because answer from one manikin can be compared with the answers from the others — albeit at the cost of added compute .
“ With this … architecture , you do n’t have to break-dance bigger undertaking into small ones and so you’re able to train many smaller models , ” Liang said , adding that Samba-1 can be deploy on - assumption or in a hosted environment bet on a customer ’s needs . “ With one big poser , your compute per [ request ] is higher so the toll of preparation is higher . [ Samba-1 ’s ] computer architecture collapse the cost of training . ”
I ’d counter that plenty of vendors , including OpenAI , propose attractive pricing for fine - tuning orotund generative models , and that several startup , MartianandCredal , allow putz to route prompt among third - political party models based on manually programmed or automated ruler .
But what SambaNova ’s selling is n’t novelty per se . Rather , it ’s a exercise set - it - and - forget it box — a full - stack resolution with everything included , includingAI chips , to build AI app program . And to some endeavour , that might be more appealing than what else is on the table .
“ Samba-1 gives every enterprise their own custom GPT model , ‘ privatise ’ on their data and customise for their administration ’s needs , ” Liang said . “ The modeling are trained on our client ’ private data , host on a unmarried [ waiter ] rack , with one - tenth part the price of alternate solution . ”
