Foundation model apps are exploding, and the tech is evolving fast. But most builders can't yet apply their imagination to build great FM apps. Most apps we all really want, aren't here yet. So the future isn't happening fast enough. A thread π
(2/11) I'm a recovering Amazonian, so this may or may not all end with a blog post that is roughly 6 pages long. But will try and hit the high notes first.
(3/11) Builders who want to make apps with foundation models face an unappealing tradeoff: easy to build but hard to defend β or the opposite. That's bad for everyone -- with galactic tech depth, or only tech curiosity, and the industry who wants the future to happen faster.
(4/11) That challenge is an opportunity. If we draw out the technical stack of an FM application, we can see where things are changing the fastest, and what are the key opportunities for founders. Here's how the stack looks in January 2023.
(5/11) An iPhone/Android tension is emerging bw proprietary vs OSS models. Prop models are powerful, accessible, and highest accuracy (HELM). OSS models are customizable (for advanced teams). OSS is closing the accuracy gap. Teams will choose sides based on makeup and mission
(6/11) Tooling grew almost overnight, and will unlock differentiation apart from capabilities of the FMs. Includes frameworks like @LangChainAI, Dust.tt, Cognosis, @gpt_index, and Fixie.ai to simplify context, prompt engineering, choice of model/s.
@LangChainAI @gpt_index (7/11) It's been hard to get FM's to reason about data they weren't trained on. It would be hard to build Amazon, Google, or Expedia with monthly or quarterly catalog updates. Now what if we fixed that? That's where FM's and the stack are going with new science and new tools.
@LangChainAI @gpt_index (8/11) FM Ops let builders push the science envelope, faster. They can access OSS architectures on @huggingface. Data from @laion_ai, PILE, and @CommonCrawl, plus @SnorkelAI, fastdup, and @xetdata to organize the data. And FM-tailored training from @CerebrasSystems and @MosaicML
@LangChainAI @gpt_index @huggingface @laion_ai @CommonCrawl @SnorkelAI @xetdata @CerebrasSystems @MosaicML (9/11) FM Ops let builders deliver faster and push on cost. Builders can use @OctoML to optimize models and utilization, plus specialized infra from @modal_labs, @BananaDev_, Beam who aim to be easier and faster versus AWS/Azure/GCP
@LangChainAI @gpt_index @huggingface @laion_ai @CommonCrawl @SnorkelAI @xetdata @CerebrasSystems @MosaicML @OctoML @modal_labs @BananaDev_ (10/11) So: (1) Tooling unlocks new features and differentiation for FM apps, both for technically curious and galactic teams. (2) Sophisticated teams can use tooling and FMOps to push science and cost. (3) Infra teams have big opportunities to build Tools and FM Ops for others
@LangChainAI @gpt_index @huggingface @laion_ai @CommonCrawl @SnorkelAI @xetdata @CerebrasSystems @MosaicML @OctoML @modal_labs @BananaDev_ (11/11) Itβs up to all of us to make the future happen faster in foundation models. @palak_go @tmporter and I are pleased to share a blog that covers these topics in a bit more technical depth: madrona.com
Back to the top.
Loading suggestions...