Yet another startup hoping to cash in on the generative AI craze has secured an eye-popping tranche of VC funding.
Called Fixie, the firm, founded by former engineering heads at Apple and Google, aims to connect text-generating models similar to OpenAI’s ChatGPT to an enterprise’s data, systems and workflows. Co-founder and CEO Matt Welsh describes it as the first enterprise-focused platform-as-a-service for building experiences with large language models (LLMs).
“Essentially, Fixie is an infinitely extensible model that enterprises can integrate into their own products and tools,” co-founder and CPO Zach Koch told TechCrunch in an email interview. “The core of Fixie is its LLM-powered agents that can be built by anyone and run anywhere.”
Whether Fixie is the first of its kind is slightly in question, but what isn’t is the founding team’s pedigree.
Welsh was an engineering leader on the Chrome team at Google for nearly a decade before coming to Fixie. Koch was a product director at Shopify and a lead on the Chrome and Android teams. CTO Justin Uberti was one of the original architects of AOL Instant Messenger. As for Fixie’s chief AI officer, Hessam Bagherinezhad, he was a machine learning exec at Apple on products including the iPhone and Apple Watch.
Here’s the ten-thousand-foot view of Fixie platform’s: LLM-powered agents that interface with external systems. Fixie agents can interact with databases, APIs (e.g. GitHub’s), productivity tools (e.g. Google Calendar) and public data sources (e.g. web search engines and social media) to generate and process arbitrary things, like images and text, and then manipulate them in various ways.
With Fixie, a company could, for example, incorporate language model capabilities into their customer support workflows by building agents that take in a customer ticket as input, automatically look up a customer’s purchases, issue a refund if necessary and generate a draft reply to the ticket.
Fixie agents can be implemented in any programming language and hosted on any infrastructure, and each agent can use its own custom-tailored LLM. Fixie supports popular models such as OpenAI’s GPT-4 out of the box, but customers can provide their own models or tap other commercial and open models if they choose.
“Ultimately, we believe that LLMs replace a lot of conventional software, since these models can act as a natural-language-powered ‘problem solving engine,’” Welsh said. “Rather than writing a bunch of gnarly code to interface two systems together, with Fixie, it is a simple matter of wrapping each system in a natural language agent interface and getting those systems to communicate with each other in English. The LLM itself acts as an incredibly powerful symbolic manipulator, requiring no programming to parse, manipulate, and synthesize data. Natural language can act as a lingua franca for diverse computing systems to talk to each other.”
It’s a compelling vision, to be sure — and one that OpenAI embraced recently with the launch of plugins for ChatGPT. In a piece this week, media analyst Ben Thompson wrote about how plugins make ChatGPT more of an aggregator or platform rather than simply a chat interface — similar to how Welsh describes Fixie and its family of agents.
ChatGPT plugins could represent somewhat of an existential threat to Fixie, in fact. But Welsh argues that the Fixie platform offers far more customizability — and freedom — than OpenAI’s take, at least at present.
“New ChatGPT plugins provide a great way to connect OpenAI’s LLM with external APIs. But our focus with Fixie is different,” he said. “Because Fixie is model- and provider-agnostic, enterprises can leverage LLMs of any kind and host agents on their own infrastructure … Fixie handles the underlying LLM interactions as well as details such as user identity, authentication, session management, storage and configuration.”
Welsh sees another rival in Zapier’s Natural Language Actions feature, which lets developers use natural language to move info between apps, products and services. But he doesn’t consider it to be directly competitive, noting that Fixie doesn’t train its own LLMs from scratch but rather enables customers to fine-tune existing LLMs for their agents using either proprietary data or historical data flowing through a given agent.
Indeed, Welsh makes the case that Fixie goes several steps beyond what’s out there by addressing some of the major hurdles in adopting generative AI, namely the high cost of training LLMs and the risks associated with even the best models available today. Fixie allows companies to fine-tune rather than train models themselves, eliminating a cost expenditure, he asserts, and constrains the actions of models to ensure they more reliably perform tasks and answer questions.
Welsh wasn’t hyperbolic to the point that he promised Fixie can completely fix (forgive the wordplay) LLMs’ tendency to make up facts, a problematic phenomenon known as hallucination. (Fixie won’t solve their other problems either, like biases and short memories.) He also conceded that fine-tuning alternatives to the Fixie platform exist, like the open source LangChain and Llama Index. But Welsh stressed that Fixie is designed for users with a range of expertise — in theory lowering the barrier to entry for deploying generative AI.
To wit, Fixie has around 5,000 users in an early access program and says it’s working with “a wide range” of companies on use cases like business automation, customer support, generative AI and graphics. It’ll launch publicly in the coming days, free for personal use, backed by a $17 million investment ($12 million in seed funding, $5 million in pre-seed funding) from Redpoint Ventures, Madrona, Zetta Venture Partners, SignalFire, Bloomberg Beta and Kearny Jackson.
That Fixie found funding easily — and quickly, within the span of the past half year — isn’t surprising, exactly. According to a PitchBook report released this month, VCs have steadily increased their positions in generative AI, from $408 million in 2018 to $4.8 billion in 2021 to $4.5 billion in 2022. Angel and seed deals have grown, as well, with 107 deals and $358.3 million invested in 2022 compared with just 41 and $102.8 million in 2018.
Assuming all goes well, Welsh says Fixie plans to grow its eight-person team to 20 by the end of the year. Before then, it’ll focus mainly on customer acquisition.
“LLMs enable radically new capabilities for software systems of all kinds, but businesses haven’t yet been able to take full advantage of these advancements,” Welsh said. “Everyone has seen the tremendous power of things like ChatGPT, and there is widespread recognition of the huge impact this technology will have on the entire information technology industry. The question is how to best tap into this technology and integrate it with existing and new systems in a way that is secure, scalable, and easy to deploy and manage. That’s where Fixie comes in.”
Fixie wants to make it easier for companies to build on top of language models by Kyle Wiggers originally published on TechCrunch