Plotch.ai

10x decrease in inferencing cost per million token every year! Cost of inferencing has gone down 1000x in last 3 years. The cost is now below 10 cents/million token. Llama 3B leads in lowest inferencing cost for MMLU > 42 (< $0.1/million token) Llama 70B leads for MMLU > 83 (< $1/million token) Great data point from Andreessen Horowitz

Here is one more alternate view of modern AI stack. There are apps like amazon.com, hotels.com etc which exists today which most likely will become transaction rails backbone in future. The new world of agents will use these transaction rails for task completion. An agent at the top will use an orchestration layer to route between planning of the tasks (using model layer) and task completion (tooling layer). Tooling layer will provide agent-to-app pathway for transaction completion. Following companies are interesting backbone of this stack > Agentic layer : Decagon, Replit, Harvey, Dosu, Perplexity > Orchestration layer: LlamaIndex, LangChain, CrewAI, phidata, Letta > Tooling layer: Plotch.ai, Composio, Browserbase > Genai layer: Copy.ai, Synthesia, Runway > Model layer: Sarvam, OpenAI > GPU Layer: NVIDIA > CPU layer: Intel Corporation How will this modern AI stack unfold in 2025? Lets see

As part of our groclake agentic AI infrastructure, we have built cataloglake. So what is cataloglake? Cataloglake is a comprehensive infrastructure for catalog warehousing. Using Cataloglake, builders can generate, store, cache index and disseminate ecommerce catalog. You can store all your catalog at one place once and then reuse it multiple times anywhere like sharing it with shopify, building search capabilities and displaying products in your webstore. For example, Cataloglake can be used to create recommendation agents, conversational ecommerce search agents, create smart filters, generate catalog using images only, category and product page listing on webstores and many more. Join our dev sessions everyday at 6pm to deep dive into groclake by registering here: https://lnkd.in/gbX7YWGH

The largest Agentic AI Hackathon in India is live now with 500+ teams already registered! o More than Rs. 50k worth prizes o Paid internships from Plotch.ai o Certificates and recommendation letters Register here: https://lnkd.in/gGfaFsCZ If you are a developer, you do not want to miss this opportunity. Agentic AI is the next revolution in AI after GenAI and is expected to be much bigger in size. Prepare for future today!

Market share of OpenAI is going down! Interesting trend of market share of OpenAI (2023 vs 2024). Their share has come down from 50% to 34% while Anthropic has gained market share from 12% to 24% and Google has jumped from 7% to 12%.

Here we propose building AgNet: Network of Agents in India at billion unit scale. The key components can be borrowed from Open Network For Digital Commerce (ONDC) architecture to build on the learnings of already scaled network in India. > Agent Registry: This registry holds name, description, tags, id, endpoint, public key and organization name of participating agent > The api format between agents we propose a generic conversational format to take advantage of LLM. Hence agents talk to each other using human language format. This format we are internally calling at Plotch.ai as opengroc format. This avoids headache of poiint-to-point integration amongst agents. > All agents will need to be rated to build trust in the network > Rating, discovery, register and de-register apis can be opengroc format too (or can be structured format given they can be easily defined) Eric Broda, Pramod Varma and sujith nair are key inspiration behind this and most of the thoughts are borrowed (other than opengroc format). We at Plotch.ai are just trying to stitch this as a possibility. We can all build this network together.

Menlo Ventures defines the modern AI stack consisting of four layers: > Layer 1 (Compute + Foundation): OpenAI, Unsloth AI, Lamini, Predibase, Modal, Baseten, Fireworks AI, Modular, CentML, Clarifai, DatologyAI, Scale AI, Mercor, Cleanlab, Together AI, Lightning AI, Mosaic > Layer 2 (Data): Databricks, MongoDB, Pinecone, Neon, Reducto, Indexify, unstructured.io, Datavolo, Voyage AI, Superlinked > Layer 3 (Orchestration): Langchain, Llamaindex, CrewAI, Braintrust, Patronus AI > Layer 4 (Tooling): Anon, Arcade AI, Composio, Browserbase, Tiny Fish The tooling layer is the emergent layer which did not exist last year. Menlo Ventures report did not mention any Indian company in this stack (to my knowledge). I have added some prominent ones here to highlight them: > Layer 1: Sarvam, E2E Cloud and Simplismart > Layer 3: Tune AI > Layer 4: Plotch.ai What are the companies based in India which fall in these four layers of AI stack (which I missed)?