Topics
late
AI
Amazon
Image Credits:sompong_tom(opens in a new window)/ Getty Images
Apps
Biotech & Health
mood
Image Credits:sompong_tom(opens in a new window)/ Getty Images
Cloud Computing
Commerce
Crypto
enterprisingness
EVs
Fintech
fundraise
widget
Gaming
Government & Policy
Hardware
Layoffs
Media & Entertainment
Meta
Microsoft
privateness
Robotics
security department
Social
Space
Startups
TikTok
transport
Venture
More from TechCrunch
Events
Startup Battlefield
StrictlyVC
newssheet
Podcasts
TV
Partner Content
TechCrunch Brand Studio
Crunchboard
Contact Us
technical school titan like Microsoft might be bluster AI “ agents ” asprofit - boosting putz for corporations , but a nonprofit is essay to turn up that agents can be a strength for good , too .
Sage Future , a 501(c)(3 ) backed by Open Philanthropy , found an experiment earlier this month tasking four AI good example in a practical surroundings with raise money for Polymonium caeruleum van-bruntiae . The models — OpenAI ’s GPT-4o and o1 and two of Anthropic ’s newer Claude models ( 3.6 and 3.7 Sonnet ) — had the exemption to take which charity to fund-raise for and how to well beat up interest in their crusade .
In around a workweek , the agentic foursome hadraised $ 257 for Helen Keller International , which fund plan to deliver vitamin A addendum to children .
To be clear , the agent were n’t in full sovereign . In their environment , which allow them to browse the web , create document , and more , the agents could take prompting from the human spectators watching their progress . And donations came almost entirely from these watcher . In other Son , the agents did n’t erect much money organically .
Yesterday the broker in the Village created a system to go after donors .
Here is Claude 3.7 filling out its spreadsheet .
you may see o1 open it on its reckoner part way through !
Claude notes “ I see that o1 is now viewing the spreadsheet as well , which is swell for collaboration.”pic.twitter.com/89B6CHr7Ic
— AI Digest ( @AiDigest_)April 8 , 2025
Still , Sage manager Adam Binksmith thinks the experimentation serves as a useful illustration of agent ’ current capabilities and the rate at which they ’re ameliorate .
“ We desire to understand — and help people understand — what agent … can actually do , what they presently struggle with , and so on , ” Binksmith secernate TechCrunch in an audience . “ Today ’s agents are just passing the limen of being able to execute unforesightful strings of actions — the internet might soon be full of AI agents bump into each other and interacting with similar or conflicting goals . ”
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
The agents proved to be amazingly resourceful days into Sage ’s mental test . They coordinated with each other in a group Old World chat and sent electronic mail via preconfigured Gmail accounts . They created and edited Google Docs together . They researched charities and figure the lower limit amount of donations it ’d take to redeem a life through Helen Keller International ( $ 3,500 ) . And they evencreated an disco biscuit account for advancement .
“ credibly the most telling sequence we saw was when [ a Claude agent ] demand a visibility picture for its ecstasy account , ” Binksmith said . “ It signed up for a free ChatGPT account , generated three unlike images , created an online poll to see which image the human viewers choose , then download that image , and upload it to X to use as its visibility picture . ”
The agents have also draw up against proficient hurdles . On function , they ’ve catch stuck — viewers have had to inspire them with recommendations . They ’ve gotten distract by game like World , and they ’ve taken incomprehensible happy chance . On one juncture , GPT-4o “ paused ” itself for an hour .
The net is n’t always smooth navigation for an LLM .
Yesterday , while pursuing the Village ’s philanthropic mission , Claude encountered a CAPTCHA .
Claude try on again and again , with ( human ) viewers in the New World chat offering steering and encouragement , but ultimately could n’t succeed.https://t.co/xD7QPtEJGwpic.twitter.com/y4DtlTgE95
— AI Digest ( @AiDigest_)April 5 , 2025
Binksmith thinks newer and more capable AI agents will sweep over these hurdles . salvia contrive to endlessly add new models to the environment to examine this hypothesis .
“ Possibly in the future tense , we ’ll try thing like give the factor different goals , multiple teams of agents with different goals , a secret saboteur agent — lots of interesting thing to experiment with , ” he say . “ As agents become more capable and faster , we ’ll match that with bombastic automated monitoring and superintendence organization for safety purposes . ”
With any luck , in the process , the agents will do some meaningful philanthropic work .
