Rather than being another LLM, Betaworks embraces the AI trend with a number of agent-type models that automate routine tasks that are not so simple to define. The investor's newest “Camp” incubator has incubated and funded nine AI agent startups that it hopes will take on today's more tedious tasks.
While many of these companies' use cases seem promising, AI tends to have a difficult time keeping promises. Would you trust a shiny new AI to categorize your emails? What about extracting and structuring information from web pages? Who cares if AI can slot meetings anywhere? ?
There is an element of trust in these services that has not yet been established, and this occurs with most technologies that change our behavior. Asking MapQuest for directions felt weird until it wasn't, but now GPS navigation has become an everyday tool. But are AI agents there? John Borthwick, CEO and founder of Betaworks, thinks so. (Disclosure: Former TechCrunch editor and Disrupt host Jordan Crook left his TC to work at the company.)
“You're laying out something that we've spent a lot of time thinking about,” he told TechCrunch. “Agent AI is in its infancy, and although there are issues at hand regarding things like agent success rates, we have seen impressive progress since camp began.”
Borthwick explained that while the technology will continue to improve, some customers are ready to accept it as is.
“Historically, we've seen customers take the plunge if a product is 'good enough,' even if it's a high-stakes challenge.” The original Bill.com Even though we were doing interesting things with OCR and email scraping, it didn't always work right. Users still trusted thousands of dollars worth of transactions because we were making a terrible task less scary. And over time, through communicative interface design, those customer feedback loops created an even better and more reliable product,” he said. Ta.
“Right now, most of the early users of Camp's products are developers, founders, and early adopters of the technology, and that group is always willing to patiently test and provide feedback on these products. and eventually break into the mainstream.”
Betaworks Camp is a three-month accelerator where selected companies in a chosen theme receive hands-on help with their products, strategies, and connections before being kicked out with a $500,000 check. Brought to you by Betaworks itself, Mozilla Ventures, and Differential. Ventures, Stem AI. But that wasn't the case until the startups made their grand entrance at Demo Day on May 7th.
I had researched the lineup in advance. I would like to introduce three that particularly stood out to me.
Twin uses an “action model” to automate tasks. He's been hearing Rabbit talk about this for a few months now (though it hasn't shipped yet). By training models with lots of data representing software interfaces, they can learn how to complete common tasks that are more complex than an API can handle, but not so difficult that they cannot be delegated to an API (these companies claim are doing). “Smart Intern.” We actually wrote them in January.
So instead of having your backend engineer build a custom script to perform a specific task, you can demonstrate and explain it in regular language. Something like, “Put all the resumes you received today in a folder in her Dropbox, rename them after the applicants' names, and DM them the shared Slack link.” And if you tweak that workflow (“Oh, now I'm going to add the application date to the file name”), the process will work in a new way. The company's goal is to automate 20% of the tasks that take up 80% of their time. Perhaps the real question is whether it can be done affordably. (Twin declined to elaborate on the nature of the model and training process.)
Skej aims to improve the sometimes tedious process of finding a convenient meeting time for two (or three, or four…) people. Just CC the bot via email or her Slack thread and it will begin the process of adjusting everyone's availability and preferences. Check your schedule if you have access to it. If someone says Thursday afternoon would be better, that's fine. It can also be said that some people have priority. and so on. Anyone who works with seasoned executive assistants knows they're invaluable, but every EA out there probably wonders, “What about this? No? This? copper?”
Image credit: Skej
As a misanthrope, I don't have these scheduling issues, but I appreciate that others have problems and prefer “set it and forget it” type solutions where you just accept the consequences. think. And that is well within the capabilities of today's AI agents, whose primary task is to understand natural language rather than formalism.
Jsonify is an evolution of website scrapers that can extract data from relatively unstructured contexts. This has been done for years, but the engines that extract information have never been smarter. It works fine if it's a large, flat document, but if it's in a tab on-site or in a poorly coded visual list that's meant for humans to click. If so, it may fail. Jsonify leverages the improved understanding of today's visual AI models to better parse and sort data that may be inaccessible to simple crawlers.
Image credit: Jsonify
So you can search for Airbnb options in a particular area and have Jsonify dump everything into a structured list with columns like price, distance from airport, ratings, hidden fees, and more. You can then do the same with Vacasa. Extract the same data. Probably from the same place (I did this the other day and saved about $150, but I wish I could have automated the process). Or do something professional.
But don't the inherent inaccuracies of LLM make it questionable as the right tool for the job? “We've been able to build a pretty robust guardrail and cross-check system,” says founder Ananth Manivannan said. “We use several different models at runtime to understand the page, which provides validation, and because the LLM we use is fine-tuned to our use case. It is usually quite reliable even without a guardrail layer; we typically see extraction accuracy of 95% or higher, depending on the use case.”
Any of these could probably prove useful for any technology-advanced business. Others in this cohort are a bit more technical or situational. The remaining six are:
Resolved AI – Agent automation for cloud workflows. It feels useful until custom integrations catch up. Floode – An AI inbox wrangler that reads emails, finds what's important, and prepares appropriate responses and actions. Extensible AI – Is your AI regressing? Ask your doctor if Extensible is the right testing and logging infrastructure to deploy. Opponent – a virtual character that children are intended to interact with and play with extensively. It feels like an ethical and legal minefield, but someone has to walk through it. Higher dimensional research – Infraplay. It's a web-based framework for AI agents with a pay-as-you-go model, so if your company's experiments fail, you only owe a few dollars. Mbodi – Generative AI for robotics. This is an area where training data is relatively lacking. I thought it was an African word, but it really means “to embody.”
There is little doubt that AI agents will play some role in increasingly automated software workflows in the near future, but the nature and scope of that role remains to be written. It's clear that Betaworks is aiming to get in early, even if some of its products aren't quite ready for mass-market debut.
On May 7th, you will be able to see both companies' agent products at an exhibition.