Nakul Mandan’s Post

Agree there’s lots of opportunity for innovation in the app layer but also agree with other comments that it will be very competitive. Proprietary datasets will be critical for value creation. This could be from new sensors that startups install themselves (I’m biased here, cough cough) or it could be customer data that they entrust a vendor with. For the latter, I think incumbents have an advantage over startups so new startups need to work extra hard to deliver value and earn trust.

Gotta push back pretty hard here. This is a highly redacted transformer / generative AI landscape. There is a building consensus that AGI takes four major things: - The capacity to understand the world. (Ingestion and sensing layer) - Persistent memory that is time-aware. (Memory layer) - An ability to plan now and for the future. (Extrapolation capability and layers) - Show the ability to reason. (Not yet achieved and most likely a multi-modal contextuality synthesis layer)

I would push back on this. I think at least for now, the model is the product. The foundation model maybe commoditized but customized products with SFT+RAG will remain differentiators. The app layer right now is a thin wrapping on the model. These have low barrier to entry and almost no moat. Many of these products are becoming just features for the foundational models (eg. custom GPTs). Maybe when this changes and we get more customization in the architecture and more in-situ use of the models (Eg. embedded deeply into a workflow) the app layer will become a bigger differentiator again.

On the outline it does makes sense but I feel what the community misses is : “the special practical layer” : “Commidity layer” if you will. That’s not your Model Layer. Nor is your Dev Tool Layer. It’s responsible for ensuring all the QA done for model controllability in deployment is encoded in your model layer for it to be commoditised into application specific controllable intelligence (adapters if you will) It’s this layer that evolves into tokenisation of this commodity itself (think hybrids of MergeKit (for creating mixtures) and LORA (for Finetuning))… not sure yet as this space would evolve eventually over the time. But yeah, the “special” 5th layer somewhere between model and Dev tool, actually helps with commoditisation practically. Lot of value to be created here as well. Potential intelligence marketplaces sort of thing happens here. Likely.

Agree that there is tremendous opportunities for startups in the app layer. Wrapping the big foundational models can look generic and easily replicable at first glance, but as with all things SW, your unique insights as a founder, your specific execution and your drive all weave a fabric that becomes increasingly unique over time. So just keep building 👷♂️🏗️

Absolutely. The app layer is where you need to add domain expertise to unlock latent knowledge and deliver it to decision makers correctly, consistently and concisely. samvid.ai

Remember reading this very specific infra example--analogous to AI Infra: "Cell towers do not make money, air time minutes passing through them do."

Completely agree with your assessment. Massive value creation will happen in the Apps layer. However, we are seeing larger enterprises build congnitive apps internally vs buying. This makes sense for apps that have privacy and security implications. We may see the same progression: custom -> packaged cognitive apps.

Not missing the mark anywhere but it would be useful to have another axis for “level of competition<>barrier to entry”