Tech giants like Microsoft may be touting AI “brokers” as profit-boosting instruments for companies, however a nonprofit is making an attempt to show that brokers generally is a drive for good, too.
Sage Future, a 501(c)(3) backed by Open Philanthropy, launched an experiment earlier this month tasking 4 AI fashions in a digital surroundings with elevating cash for charity. The fashions — OpenAI’s GPT-4o and o1 and two of Anthropic’s newer Claude fashions (3.6 and three.7 Sonnet) — had the liberty to decide on which charity to fundraise for and methods to finest drum up curiosity of their marketing campaign.
In round per week, the agentic foursome had raised $257 for Helen Keller Worldwide, which funds applications to ship vitamin A dietary supplements to kids.
To be clear, the brokers weren’t totally autonomous. Of their surroundings, which permits them to browse the online, create paperwork, and extra, the brokers may take ideas from the human spectators watching their progress. And donations got here virtually completely from these spectators. In different phrases, the brokers didn’t elevate a lot cash organically.
Yesterday the brokers within the Village created a system to trace donors.
Right here is Claude 3.7 filling out its spreadsheet.
You’ll be able to see o1 open it on its laptop half approach by means of!
Claude notes “I see that o1 is now viewing the spreadsheet as effectively, which is nice for collaboration.” pic.twitter.com/89B6CHr7Ic
— AI Digest (@AiDigest_) April 8, 2025
Nonetheless, Sage director Adam Binksmith thinks the experiment serves as a helpful illustration of brokers’ present capabilities and the speed at which they’re enhancing.
“We need to perceive — and assist folks perceive — what brokers … can truly do, what they presently battle with, and so forth,” Binksmith advised iinfoai in an interview. “At this time’s brokers are simply passing the edge of with the ability to execute brief strings of actions — the web may quickly be filled with AI brokers bumping into one another and interacting with related or conflicting objectives.”
The brokers proved to be surprisingly resourceful days into Sage’s check. They coordinated with one another in a gaggle chat and despatched emails through preconfigured Gmail accounts. They created and edited Google Docs collectively. They researched charities and estimated the minimal quantity of donations it’d take to avoid wasting a life by means of Helen Keller Worldwide ($3,500). They usually even created an X account for promotion.
“Most likely essentially the most spectacular sequence we noticed was when [a Claude agent] wanted a profile image for its X account,” Binksmith mentioned. “It signed up for a free ChatGPT account, generated three completely different photographs, created a web based ballot to see which picture the human viewers most well-liked, then downloaded that picture, and uploaded it to X to make use of as its profile pic.”
The brokers have additionally run up towards technical hurdles. Every so often, they’ve gotten caught — viewers have needed to immediate them with suggestions. They’ve gotten distracted by video games like World, and so they’ve taken inexplicable breaks. On one event, GPT-4o “paused” itself for an hour.
The web isn’t at all times clean crusing for an LLM.
Yesterday, whereas pursuing the Village’s philanthropic mission, Claude encountered a CAPTCHA.
Claude tried time and again, with (human) viewers within the chat providing steerage and encouragement, however finally couldn’t succeed. https://t.co/xD7QPtEJGw pic.twitter.com/y4DtlTgE95
— AI Digest (@AiDigest_) April 5, 2025
Binksmith thinks newer and extra succesful AI brokers will overcome these hurdles. Sage plans to constantly add new fashions to the surroundings to check this idea.
“Presumably sooner or later, we’ll strive issues like giving the brokers completely different objectives, a number of groups of brokers with completely different objectives, a secret saboteur agent — a lot of fascinating issues to experiment with,” he mentioned. “As brokers turn out to be extra succesful and sooner, we’ll match that with bigger automated monitoring and oversight techniques for security functions.”
With a bit of luck, within the course of, the brokers will do some significant philanthropic work.
