journalfront

Google's Gradient Ventures Funds Cake's Open-Source AI Platform

2024-12-04

In a significant development today, a fresh company is stepping into the spotlight with the backing of Google's AI-focused venture fund. This company aims to assist businesses in compiling their open-source AI infrastructure and reducing engineering overheads. Cake, the emerging entity, integrates and secures over 100 components crucial for enterprises. These include data source adapters like Apache Hadoop, data ingestion tools such as Apache Kafka, data labelling platforms like Label Studio, vector and graph databases like Milvus or Neo4j, generative AI APIs and related tools like Anthropic, and many more categories. This gives a clear indication of why Cake is named as it is - it takes the diverse "layers" of the AI stack and integrates them into a more manageable and production-ready format suitable for business operations.

Founding and Early Years

Cake was founded in 2022 in New York by Misha Herscu (CEO) and Skyler Thomas (CTO). Last year, it launched and is already collaborating with customers like AI bioscience startup Altis Labs and data intelligence insurtech Ping. However, until now, the company hadn't been making much public noise. Since its inception, Cake has raised $13 million. This includes $3 million in pre-seed funding during its formative years and a recent $10 million seed round led by Google's Gradient Ventures. "We haven't been overly secretive; we've been focused on building and working with customers," Herscu explained in an interview with TechCrunch last week.Previously, Herscu founded McCoy Medical Technologies, an AI company focused on machine learning infrastructure for radiology, which he sold in 2017 to IT vendor TeraRecon. Later, he joined New York VC firm Primary Venture Partners as an "operator in residence" and engaged in extensive customer discovery calls. "I did over 200 customer discovery calls, asking about their biggest pain points and bottlenecks. The main issue wasn't a single part of the stack like setting up a vector database or data pipeline. It was the plethora of different components across a rich ecosystem. How do you integrate everything reliably and make it production-ready?" Herscu said. This "big picture problem" is where his new business comes into play.Cake is all about making sense of the numerous open-source components that make up the modern AI stack and providing bundled, managed open-source AI infrastructure for small teams. It's not about building a business around a single open-source project like many others have done. Instead, it's about assembling and serving a curated selection of open-source projects across the entire stack and ensuring smooth operation.For example, a large financial services company with millions of complex financial data documents might want to perform RAG (retrieval augmented generation) against these files to enhance the quality of natural-language query responses. If an off-the-shelf product doesn't meet the requirements or is not compliant, the company would have to build its own system by installing and integrating multiple different components. Cake can handle this time-consuming task.In another scenario, a hospital might need to build a secure system for analyzing CT scan images, or an e-commerce company might want to upgrade its recommendation engine. These are all potential use-cases for Cake. "We do cover a wide range, but our sweet spot is definitely when companies need more than a simple, off-the-shelf product," Herscu said.

Team Expertise and Parallel Development

Cake's CTO Thomas has an extensive background. He previously worked at IBM as a chief architect and more recently was a distinguished engineer and director of strategy at Hewlett Packard Enterprise, which acquired a previous company he worked at called MapR. Thomas has worked on hundreds of projects with both large and small customers and noticed a common trend - almost all of them were using open-source tools in some form, often directly from research labs. Yet, using them in the enterprise wasn't straightforward."It takes a tremendous amount of time for even the largest enterprises to incorporate what's emerging from the labs into their operations," Thomas told TechCrunch. "A lot of it is because much of it isn't ready for enterprise use - it may lack authentication and authorization, and enterprises have to handle these aspects themselves."There are similarities to what Cake is aiming for. In Europe, Finnish Aiven, a $2 billion unicorn, is doing something similar but with a focus on data infrastructure. The most obvious comparison might be Red Hat, which IBM acquired for $34 billion and is renowned for its enterprise-grade Linux operating system (RHEL)."In the early days of Linux, there were thousands of open-source packages that everyone wanted to use but weren't integrated or secure," Thomas said. "There was no proper support model, and that's where the Red Hats of the world made Linux safe for enterprises. We want to do a similar thing for AI today."While there are plans to introduce a hosted version of Cake in the future, for now, companies have to run it in their own environments. For many, this isn't a problem as data privacy regulations prevent them from sending data outside their systems. But a hosted version might be attractive to organizations with lower compliance requirements."It is actually more convenient for us if we can control the cloud," Herscu added.Aside from lead investor Gradient, Cake's seed round saw participation from its pre-seed investor Primary Venture Partners, as well as Alumni Ventures, Friends & Family Capital, Correlation Ventures, and Firestreak Ventures. The previously unannounced $10 million seed round, which closed in April, not only reflects the founders' backgrounds but also the company's progress. Herscu said the company is already looking towards its next financing round, with tentative plans to raise again around the middle of 2025. "From a traction perspective, we already look like a Series A company. We were able to achieve this relatively quickly. When we go for the Series A, it might seem more like a Series B," Herscu said.

When Nicolás Camhi, Matías Pérez Pefaur, and Diego Chahuán launched Vambe last year, their initial focus was on building a CRM for debt collection. However, they soon discovered that their customers were more intrigued by the WhatsApp AI agents Vambe had developed to assist in the debt collection process. As Camhi shared with TechCrunch, “Our customers were already asking us, ‘Hey, could you ask the AI so that when you go out after a debt, could you offer this person, I don’t know, “x” product or “”x” service?’” This realization led them to a significant pivot. In March 2024, Vambe shifted its focus to these AI agents, aiming to empower small and medium-sized businesses with automated communication tools to close sales on platforms like WhatsApp.

Customer Base and Revenue Growth

Vambe’s customer base spans a wide range, from small mom-and-pop businesses with just five employees, such as carpet cleaners, to large retail companies with thousands of workers. Before the pivot, the company had an annual recurring revenue (ARR) of about $20,000. But since March, this number has witnessed a remarkable surge. In November, Vambe closed with an ARR of $1 million. This rapid growth has positioned Vambe as a force to be reckoned with in the market.

Funding and Investor Interest

To capitalize on this growth, Vambe recently raised a $3.85 million seed round. Led by Brazil-based VC firm Monashees, the round also saw participation from Mexico-based investor Nazca and U.S.-based M13. M13’s partner Brent Murri highlighted that the firm started building its thesis on the Latin American region two years ago but had been waiting for the right opportunity. Meeting Vambe’s team at the Berkeley Skydek accelerator earlier this year was the turning point. Murri was particularly impressed by Vambe’s tech. He gave an example of building a fake business and using Vambe to set up an AI agent, demonstrating how quickly and efficiently it could be done. He also noted that M13 is bullish on Vambe’s potential, stating that in the U.S., the space for AI agents for sales and marketing is crowded, but in Latin America, the adoption by SMBs and enterprises, as well as consumer willingness to adopt AI products, is actually higher than in other parts of the world.

Cultural Factors and Business Dynamics

The reason for this affinity towards sales tech in Latin America is the conversational nature of business here. Camhi explained, “It’s kind of a cultural thing. Here in LatAm, we really like to talk. People with companies don’t just engage and buy something directly from a web page. They try to reach out. They want someone to help them. They want to understand pricing. They want to understand delivery. And all of that kind of communication is something super hard to scale.” This cultural aspect presents a unique opportunity for Vambe to thrive.

Competitors in the LatAm Market

Vambe is not the only company building WhatsApp AI applications for businesses in Latin America. Mercately, based in Ecuador, is also focused on developing the back-end tech that companies need to communicate with customers and sell directly to them through WhatsApp. In the U.S., there are several competitors as well. Bret Taylor’s Sierra, which recently raised a $175 million round, and ElevenLabs, which has raised more than $100 million in total venture capital, are among the notable players. Voiceflow, a smaller entrant, has also raised more than $39 million in VC.

Future Plans and Ambitions

Beyond its ambitions in Mexico, Vambe is looking to expand its reach to Spanish-speaking businesses in the U.S. It plans to continue building out its team and developing its technology. As Camhi emphasized, “We are putting extremely advanced technology in the hands of businesses that don’t even know how to prompt. They are really increasing their sales and reducing their costs. I think that is super, super important.”

DeepMind, the AI research arm of Google, has made a remarkable breakthrough by presenting a model capable of creating an "endless" variety of playable 3D worlds. This model, named Genie 2, is the successor to DeepMind's earlier Genie released this year. It has the astonishing ability to generate an interactive, real-time scene from a single image and text description, such as "A cute humanoid robot in the woods". In this regard, it shares similarities with models being developed by Fei-Fei Li's company, World Labs, and the Israeli startup Decart.

DeepMind's Genie 2: Unleashing the Potential of 3D Worlds

Model's Training and Capabilities

Trained on videos, Genie 2 can simulate object interactions, animations, lighting, physics, reflections, and the behavior of "NPCs". Many of its simulations resemble AAA video games, and the reason might be that its training data includes playthroughs of popular titles. However, like many AI labs, DeepMind has not disclosed many details about its data sourcing methods due to competitive or other reasons.This model can generate a "vast diversity of rich 3D worlds", and users can take actions like jumping and swimming by using a mouse or keyboard. It can generate consistent worlds with different perspectives, such as first-person and isometric views, for up to a minute, with the majority lasting 10-20 seconds. Genie 2 responds intelligently to actions taken by pressing keys on a keyboard, identifying the character and moving it correctly. For example, it can figure out that arrow keys should move a robot and not trees or clouds.

Comparison with Other Models

Most models like Genie 2, which are world models, can simulate games and 3D environments but often face issues such as artifacting, consistency, and hallucination. For instance, Decart's Minecraft simulator, Oasis, has a low resolution and quickly "forgets" the layout of levels. However, Genie 2 can remember parts of a simulated scene that are not in view and render them accurately when they become visible again. World Labs' models also possess this ability.Although games created with Genie 2 might not be overly fun as they erase progress every minute or so, DeepMind positions the model as a research and creative tool. It can turn concept art and drawings into fully interactive environments and help researchers generate evaluation tasks that agents have not seen during training.

Future Implications and Research Focus

While Genie 2 is still in the early stages, DeepMind believes it will be a key component in developing AI agents of the future. Google has been pouring increasing resources into world model research, which is expected to be the next big thing in generative AI. In October, DeepMind hired Tim Brooks, who was leading the development of OpenAI's Sora video generator, to work on video generation technologies and world simulators. Two years ago, the lab also poached Tim Rocktäschel, known for his "open-endedness" experiments with video games like Nethack, from Meta.This research holds great potential for various fields, from game development to artificial intelligence. It opens up new possibilities for creating immersive and interactive experiences and evaluating AI agents in diverse environments. As DeepMind continues to advance this technology, it will likely have a significant impact on the future of both the gaming and AI industries.