Luis Ceze is many issues: He’s the CEO and co-founder of OctoAI, an Lazowska Endowed Professor at College of Washington, a co-founder of the Apache TVM mission, and likewise a 2024 BigDATA Wire Particular person to Watch.
We not too long ago caught up with Ceze to ask him a number of questions on his many endeavors. Here’s what he stated:
BigDATA Wire: You modified the title of your organization from OctoML to OctoAI in January. Are you able to elaborate on the change?
Luis Ceze: We modified our title from OctoML to OctoAI to higher mirror the growth and evolution of our product suite, which extra broadly addresses the rising market wants within the generative AI house.
Within the final 12 months, we considerably expanded our platform for builders to construct manufacturing purposes with generative AI fashions. This implies firms can run any mannequin of their alternative— whether or not off-the-shelf, customized or open-source— and deploy them on-prem inside their very own environments or within the cloud.
Our newest providing is OctoStack, a turn-key manufacturing platform that delivers highly-optimized inference, mannequin customization and asset administration at scale for big enterprises. This provides firms whole AI autonomy when constructing and working Generative AI purposes instantly inside their very own environments.
We have already got dozens of high-growth generative AI clients—like Apate.ai, Otherside AI, Latitude Video games, and Capitol AI utilizing the platform to seamlessly transport this extremely dependable, customizable, environment friendly infrastructure instantly into their very own surroundings. These firms are actually firmly in charge of how and the place they work with fashions and profit from our maintenance-free serving stack.
BDW: You’re a co-founder of the Apache TVM mission, which permits machine studying fashions to be optimized and compiled to completely different {hardware}. However GPUs are all the fad. Ought to we be extra open to working ML fashions on different {hardware}?
Ceze: We’ve skilled extra AI innovation the final 18 months than ever earlier than. From in the future to the subsequent, AI has shifted from the lab to a viable enterprise driver. It’s clear that for AI to scale, we’d like to have the ability to run it on a broad vary of {hardware} from data-centers to edge and cell gadgets.
However we’re at a juncture that’s harking back to the cloud days. Again then firms wished the liberty to host knowledge throughout multiple cloud, or a mixture of cloud and on-premise.
Right this moment firms additionally need accessibility and selection when constructing with AI. They need the selection to run any mannequin, be it customized, proprietary or open supply. They need the liberty to run stated fashions on any cloud or native endpoint, with out handcuffs.
This was our mission with Apache TVM early on, and this has carried on by means of my work at OctoAI. OctoAI SaaS and OctoStack are designed with the precept of {hardware} independence and portability to completely different buyer environments.
BDW: GenAI goes from a interval of experimentation in 2023 to deployment in 2024. What are the keys to creating LLMs extra impactful for companies?
Ceze: We strongly imagine that 2024 is the 12 months that generative AI makes it out of improvement and into manufacturing. However to deliver this to fruition, firms are going to should concentrate on a number of key issues.
The primary is controlling value so the unit economics of LLMs work in your favor. Mannequin coaching is a predictable expense, however inference (calling a mannequin working in manufacturing) can get very costly, particularly if utilization surges past what you’ve deliberate for.
Second is choosing the suitable mannequin in your use case. It’s getting more difficult due to the sheer variety of LLMs to choose from (there are 80,000 and counting) and mannequin fatigue is starting to set in. Discovering one that’s highly effective sufficient to ship the standard you want and runs effectively as to be cost-effective – that’s the stability you wish to strike.
Third, strategies like fine-tuning are extremely essential to assist customise these LLMs for distinctive performance. One development we observe is that LLMs themselves are more and more commodified, and the true worth comes from customization to satisfy a selected, high-value use case.
BDW: Outdoors of the skilled sphere, what are you able to share about your self that your colleagues is perhaps shocked to study – any distinctive hobbies or tales?
Ceze: Meals for me is greater than vitamin :). I like to study meals; I like to cook dinner it; I like to eat it.
I like to grasp meals “cross-stack”, from cultural facets right down to chemistry. After which consuming / consuming ;).
One other enjoyable bit: a few of my analysis was in DNA knowledge storage, and my work not too long ago traveled to the moon!
You possibly can learn extra concerning the 2024 BigDATA Wire Folks to Watch right here.