Note: The original text comes from @ @FinanceYF5 released a long tweet.
This is another Silicon Valley vc Base10 analysis of Generative AI.
Here are the 4 tool areas you need and the problems they solve:
1. Orchestration Orchestration
2. Deployment, Scalability & Pre-training
3. Context and embedding Context & Embeddings
4. Quality Assurance and Observability QA & Observability
1. Choreography
LLM needs to be connected to external systems, allowing dynamic data access and user manipulation, such as ChatGPT plug-ins. These emerging tools enhance the capabilities of LLM, support personalized applications and multiply the capabilities of LLM and other software.
2. Deployment, scalability
Developers are choosing open source or custom models because of privacy and customization issues with models like GPT-4, but deploying open source or custom models presents infrastructure, cost, and performance hurdles that these tools can help address these questions.
3. Embedding
The industry was hot in April, with over $175 million in funding LLMs need context or data that wasn't in their original training set to get the answers right. LLMs solve this problem by attaching a limited set of useful information to hints at inference time.
4. Quality assurance and observability
Once you have deployed your LLM powered product, you need to analyze its performance, speed, user insights, etc. so that from V1 to V2 of your product this emerging class of tools can handle observability, monitoring, , fine-tuning, QA and other tasks.
Founders building products and features with LLM are increasingly encountering roadblocks and roadblocks. We've identified startups that solve these "hard problems"—and potentially build multibillion-dollar businesses in the process.
main problem?
(1) Relevance: bigger is not necessarily better
(2) LLM is information and action limited
(3) LLM is expensive
(4) LLMs are not always private
(5) LLM may not be reliable.
We address some of these problems based on their primary use case for developers: Orchestration: http://Dust.tt @FixieAI @LangChainAI @vocodehq @JinaAI_ @gpt_index Pyq @GradientJAI @StackAI_HQ Anarchy AI @logspace_ai Trudo AI @make_berri @HubbleAi @wavelineai @patterns_app @trypromptly
Deployment, Scalability, & Pre-Training: @MosaicML @neuralmagic @anyscalecompute @BananaDev_ @OctoML @seldon_io @bentomlai Alpa @LightningAI @Zeet_Co @MindsDB @AiEleuther Utterworks @cerebriumai Meru @cargoshipsh @Texel @rubbrbandHQ Steamship @PoplarML Beam @AutoblocksAI
Deployment, Scalability, & Pre-Training (cont.): @basetenco @ForefrontAI @gooseai_NLP @runpod_io @_segmind @CentML_Inc @brevdev
Context & Embeddings: @pinecone Metal @weaviate_io Drant @zilliz_universe Valt @UnstructuredIO @RelevanceAI_ Chroma @trybaseplate @supabase Neon @activeloopai Marqo AI @Redisinc @xata Vespa @NucliaAI @PromptableAI @Unum_UK @promptifyai
QA & Observability: @humanloop HoneyHive @deepset_ai Aporia @WhyLabs @arizeai @SuperwiseAI @latticeflowai Neptune AI @gretel_ai @helicone_ai Vellum @promptlayer @GraphsignalAI Helm @PromptJoy
Original Base10:
https://t.co/4jz3JGUKWS
