Really interesting article and very complete. Thanks for sharing!
Let me add a remark/question about some companies that would be in between the two categories: those that offer a consumable AI service via API to be integrated into software or automation, while relying themselves on more "low-level" third-party services. However, they offer additional processing and provide an "intelligence" that justifies their use.
Concretely, this concerns for example document parsing services (contracts, invoices, resumes, etc.) that rely on Google or AWS for OCR but have trained their own NLP engines. The same goes for services based on speech recognition. Where would you categorize them?
We are working at Eden AI (www.edenai.co) to aggregate and harmonize foundation models to make them easier to use for the people who use them and to make it very easy for them to swap them according to the performance achieved for their specific data or the price evolution (which is a very important point). The borderline can however be quite thin between models that are 100% provided by one vendor and those that are partially based on someone else.
There is no official definition but if we use Elad's framework, I'd put your case into layer 1 or if we want to be creative, then layer 1.5. Companies rebundle the value prop of layer 1 and offer additional toolings on top if to help cater to specific needs, so it's definitely still part of infra and toolings for me
Great read, thank you :) I was wondering how you predict defensibility for Level 1 companies? Will this become an enclosed circle of a few powerful companies or do you see challengers gaining traction?
Imv, it's hard to beat the triple R: Resource, Reach, and Research - so most of the value of layer 1 will be captured by the known names: Microsoft / OpenAI, AWS / Stability AI, Facebook AI, Google AI; not sure why I have a hard time mention NVIDIA here, but let's what they come up with.
So what are your conclusions then about which companies are most likely to challenge foundational model Generative A.I. darlings? Consolidation seems to occur faster here. Stability.AI already siding with AWS.
Hard to predict the future and I hope for a commoditization (e.g. via Open Source) of that space, but my bet is on incumbents taking most of the cake wrt foundation models, so we'll continue seeing OpenAI, Google, Facebook, Amazon, et. al dominate with Huggingface, Scale, and the likes helping distribute those + provide toolings to help deploy and improve those models
Contains lots of insights into technical topics that founders of AI companies have to look into (e.g. short-term cost of ML) while still having to deal with building a product people want.
Just as you mentioned, my "fear" is open-source falling behind proprietary solutions in AI because of the reasons you named. But (from my understanding), this _strict_ platform-dependency is mainly for generative AI use cases, isn't it? In discriminative use cases, LLMs like GPT-X can be used to create the proprietary data itself (at least partially). For instance, we currently like using GPT-3 to create labeled datasets and then train super simple models on top of it, e.g. using an open-source encoder (distilbert or something from HF) for the embeddings and a simple old-school logistic regression :-) Maybe there is something like partial platform dependency? (e.g. for building, but not for inference). But just my thoughts.
Really interesting article and very complete. Thanks for sharing!
Let me add a remark/question about some companies that would be in between the two categories: those that offer a consumable AI service via API to be integrated into software or automation, while relying themselves on more "low-level" third-party services. However, they offer additional processing and provide an "intelligence" that justifies their use.
Concretely, this concerns for example document parsing services (contracts, invoices, resumes, etc.) that rely on Google or AWS for OCR but have trained their own NLP engines. The same goes for services based on speech recognition. Where would you categorize them?
We are working at Eden AI (www.edenai.co) to aggregate and harmonize foundation models to make them easier to use for the people who use them and to make it very easy for them to swap them according to the performance achieved for their specific data or the price evolution (which is a very important point). The borderline can however be quite thin between models that are 100% provided by one vendor and those that are partially based on someone else.
There is no official definition but if we use Elad's framework, I'd put your case into layer 1 or if we want to be creative, then layer 1.5. Companies rebundle the value prop of layer 1 and offer additional toolings on top if to help cater to specific needs, so it's definitely still part of infra and toolings for me
Great read, thank you :) I was wondering how you predict defensibility for Level 1 companies? Will this become an enclosed circle of a few powerful companies or do you see challengers gaining traction?
Imv, it's hard to beat the triple R: Resource, Reach, and Research - so most of the value of layer 1 will be captured by the known names: Microsoft / OpenAI, AWS / Stability AI, Facebook AI, Google AI; not sure why I have a hard time mention NVIDIA here, but let's what they come up with.
Please share the prompt used to generate the title image
Sure! It was "/imagine renaissance painting of a battle around a castle drawn by Leonardo Da Vinci --ar4:3" if I remember correctly
So what are your conclusions then about which companies are most likely to challenge foundational model Generative A.I. darlings? Consolidation seems to occur faster here. Stability.AI already siding with AWS.
Hard to predict the future and I hope for a commoditization (e.g. via Open Source) of that space, but my bet is on incumbents taking most of the cake wrt foundation models, so we'll continue seeing OpenAI, Google, Facebook, Amazon, et. al dominate with Huggingface, Scale, and the likes helping distribute those + provide toolings to help deploy and improve those models
Awesome article, thanks for sharing!
Really cool article, Viet! :)
Contains lots of insights into technical topics that founders of AI companies have to look into (e.g. short-term cost of ML) while still having to deal with building a product people want.
Just as you mentioned, my "fear" is open-source falling behind proprietary solutions in AI because of the reasons you named. But (from my understanding), this _strict_ platform-dependency is mainly for generative AI use cases, isn't it? In discriminative use cases, LLMs like GPT-X can be used to create the proprietary data itself (at least partially). For instance, we currently like using GPT-3 to create labeled datasets and then train super simple models on top of it, e.g. using an open-source encoder (distilbert or something from HF) for the embeddings and a simple old-school logistic regression :-) Maybe there is something like partial platform dependency? (e.g. for building, but not for inference). But just my thoughts.
Let's see. Really exciting times :)