Join top executives in San Francisco on July 11-12, to hear how leaders are integrating and optimizing AI investments for success. Learn More
HyperWrite, a startup known for its generative AI writing extension, unveiled an experimental AI agent today that can browse the web and interact with websites much like a human user.
In a closed demonstration over Zoom, HyperWrite’s CEO Matt Shumer showed VentureBeat how the agent, accessible through the company’s Chrome extension, could navigate to the Domino’s Pizza website and begin placing an order, looking up an address and zip code to complete the transaction.
Although the demo ended before a credit card number would have been entered, Mr. Shumer said that the capability would be enabled only for select test users at first. The agent is intended to serve as a personal assistant, automatically handling basic web tasks on command.
The AI agent trend is exploding
If you follow AI on Twitter, you are already seeing the AI agent trend explode and your news feed is likely filled with references to AutoGPT and BabyAGI. These AI agents, powered by OpenAI’s GPT-4 and other models, are capable of generating prompts automatically, also known as ‘self-prompting’ or ‘auto-prompting’, and can further develop and execute prompts based on an initial input, leading to the generation of new prompts.
Join us in San Francisco on July 11-12, where top executives will share how they have integrated and optimized AI investments for success and avoided common pitfalls.
Currently, applications built on the GPT API are basically limited to single-session use, meaning the model can’t recall information from previous interactions. This has to do with the amount of data, referred to as the number of “tokens,” that individual queries require, as well as GPT’s tendency to hallucinate — a problem that becomes increasingly noticeable as token counts rise.
Shumer described HyperWrite’s personal assistant agent as “the first user-friendly AI agent that can operate a browser like a human.” But experts cautioned that software exhibiting human-level web skills could also be prone to familiar human vulnerabilities like phishing, hacking or scams.
“We recognize the importance of the AI’s ability to take actions and complete tasks autonomously,” said Shumer. “However, potential concerns arise when such an AI could be misused to hack into company websites, manipulate communications, or spread spam on social media platforms. These things already happen with older technology, but It is crucial not to make these malicious activities easier for others.”
A focus on deploying AI agents safely
Shumer said his team is very focused on issues of safety. “We want to figure out the right way to do it, and that’s sort of the common theme through all this, we’re taking our time to do this the right way,” he said. “It’s about how do you deploy it in a way that actually is safe for society at large? And for the individual user? You don’t want to just [have the AI agent] going off and ordering something without your knowledge, right? You want it to be able to check with you.”
“There is precedent for responsible launches by giving people access, but with constraints to prevent disruptive activities,” he added. “Our team is arguably the best at building with and getting the most out of these models, and we’ve anticipated their potential impact for quite some time.”
Agents like AutoGPT and BabyAGI, he explained, can be thought of as simple stateless tools that accept text input and provide output. You can integrate them into a system that adapts to new information and improves over time, and they essentially loop until they reach the desired goal. GPT, albeit powerful, is not advanced enough to be considered incredibly dangerous or AGI-like. Its real-world applications are somewhat limited; it can perform basic research tasks, but our focus lies mainly in business productivity.
“I think this is gonna blow people’s minds,” said Shumer. “People that are fully outside of tech — family members, friends who frankly barely understand what I do on a daily basis — I show this to them and their jaws just drop because it is so tangible, it’s so easy to understand how this could impact them in their daily lives. But I also want to make sure that excitement doesn’t overshadow doing it the right way.”
To try HyperWrite’s personal assistant AI agent, install the Chrome extension and request early access.
VentureBeat’s mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.