OthersideAI Raises $2.6M to Automate Email Writing with GPT-3

Sending emails can be a meticulous process, often involving careful crafting and refinement. However, much of this effort is frequently automated. OthersideAI is capitalizing on this trend – backed by a $2.6 million seed investment – by moving beyond simple auto-replies and smart suggestions, utilizing OpenAI’s GPT-3 language model to transform concise bullet points into complete, individualized communications.

GPT-3, or Generative Pre-trained Transformer 3, represents the newest iteration of the artificial intelligence model renowned for generating remarkably persuasive text. It has become commonplace for the AI to author articles, with writers often revealing its involvement at the conclusion, though subtle indicators frequently remain detectable.

Despite limited access, the team at OthersideAI maintains a collaborative, though currently undefined, relationship with OpenAI. This connection originated during their previous project, where they encountered a volume of emails that exceeded their capacity to manage. At that time, GPT-2, the predecessor to GPT-3, was gaining prominence.

“We initially developed a cold email tool using GPT-2, but quickly realized that this could be the core of our business,” explained CEO Matt Shumer. “We then committed to pursuing this direction fully.”

Shumer, along with Jason Kuperberg and Miles Feldstein, created a demonstration that garnered attention when shared on Twitter, subsequently securing access to the latest GPT engine version.

While OpenAI has successfully engineered this impressive language engine, simply deploying it without oversight can lead to unpredictable results, as experienced by users of AI Dungeon. Uncontrolled, GPT-3 can deviate into unusual and nonsensical outputs.

“GPT-3 excels at creating impressive demonstrations, but integrating it into a functional product presents distinct challenges,” Shumer stated. “Our primary task is to manage and refine its creative potential.”

The resulting product converts summaries or bullet points into fully formed emails, as illustrated below:

Image Credits: OthersideAI

If the generated output is unsatisfactory, contains errors, or simply for experimentation, a regeneration button allows for a different version. Providing slight modifications to the initial input helps the system learn and adapt to the user’s preferred style.

The GPT models are trained on extensive datasets of text and phrases, enabling them to generate new content inspired by this data when provided with an initial prompt. In this instance, the system considers not only the user’s bullet points but also context from the email thread and the user’s established preferences.

This approach allows the system to understand nuances and incorporate relevant details. For example, it might include “It was great to sit down for coffee with you” even if the original input only contained “good to meet,” if coffee was previously mentioned. It also learns to emulate the user’s writing style, favoring specific words, phrases, or closing remarks.

The system can also make informed estimations regarding technical and financial specifics, such as when drafting a job offer:

Naturally, one might question the necessity of utilizing AI for such critical communications.

It’s comparable to a car’s maximum speed; while capable of 120 MPH, it’s rarely driven beyond 80 (or perhaps 90). The goal is to ensure the system remains reliable even when operating outside its most common applications. For OthersideAI’s model, this means maintaining robustness for “serious” emails, even if its primary use case is automating routine messages.

Kuperberg noted that the company, currently with nearly 10,000 individuals on its test version waitlist, has received interest from both engineers and developers, as well as sales and support professionals. The potential is clear in support or sales environments where a limited set of scripted responses can be dynamically re-generated, offering a personalized experience without the perception of receiving a standardized “form email.”

The possibility of assisting individuals with typing difficulties was raised – someone who relies on methods like gaze detection to compose emails might find this tool particularly valuable. Shumer mentioned that this application hadn’t initially been considered but has recently attracted attention.

Shumer emphasized the importance of security and data privacy, assuring users that the tool does not engage in data harvesting – a concern given the potential for misuse, as exemplified by Gmail.

They are confident in their security measures, suggesting that Google’s focus lies in selecting appropriate replies, while text generation tools lack the capacity to handle the complex inputs that OthersideAI’s GPT-3-based system manages effectively.

“To emulate a user’s tone, the system requires specific details. It needs human direction, rather than attempting to generate responses independently. It’s about taking instruction,” Shumer clarified.

The $2.6 million seed round was spearheaded by Madrona Venture Group, with contributions from Active Capital, Hustle Fund, Chapter One, and others. These funds will be allocated to expanding the team and developing a fully-fledged product.

Ultimately, their vision extends to a broader system of interconnected AIs capable of securely interacting with one another, providing information and answering questions in a human-like manner with minimal human intervention. While replacing email entirely has been a long-standing goal without success, perhaps it’s time to embrace it while delegating some of the workload to intelligent automation.