Quick summary:
- Data quality determines AI success, and permission-based human data delivers better results than scraped web data, which is mostly bot-generated trash.
- AI amplifies existing skills rather than replacing them and foundational expertise is required to effectively prompt and validate AI results.
- AI can fill service gaps where human demand exceeds supply.
- True personalization can be profitable with AI, which can manage multiple variables simultaneously.
Charlie Silver watched trillion-dollar tech companies build empires on user data while people got nothing in return.
Diederik Greveling saw AI hype cycles crash when the technology couldn’t deliver.
Sandy Skelaney recognized that mental health services were drowning in demand they couldn’t meet.
Steve Toy realized that one-size-fits-all education was broken.
Each built solutions with AI that address real problems and reveal what’s actually working right now. They were among the 8,000 attendees who gathered at Ai4, the world’s largest AI conference, and Greg Matusky sat down with each from the show floor for a special edition of The Disruption Is Now. The guests include:
- Charlie Silver, CEO and Founder, Permission
- Diederik Greveling, CTO, Xebia
- Sandy Skelaney, COO & Head of Partnerships, The Parasol Cooperative
- Steve Toy, CEO, Memrise
Watch now:
Key takeaways:
Paying people for their data can make AI models cleaner
Silver says more than half of the data behind AI training is junk from bots and bad collection. He argues that paying consumers for their data is the best solution.
“Data has value and consumers have yet to see any value from it,” Silver explains. “Meanwhile, big tech companies are worth trillions, all built on the one resource — data — that consumers have.”
Permission works like a talent agent for your personal information. Download the app, create a profile, and grant permission for the company to sell your data. When Nike buys your sports preferences or American Express purchases your shopping habits, you get paid.
He’s seen the lift before. In his work at RealAge, about 60% opted in and advertisers saw ROI jump around 10x when people explicitly asked to hear from them.
Traditional AI paved the way for generative breakthroughs
Greveling has watched AI evolve for 15 years as CTO of a global IT consultancy with 6,500 employees. He’s seen multiple boom-and-bust cycles, starting with neural networks in the 1980s.
“Even in the 80s when neural networks were invented, we had a relatively short cycle where people said this is going to be the future,” Greveling recalls. “It turns out the compute resources needed weren’t able to catch up.”
Before ChatGPT made headlines, Greveling’s team built practical AI solutions for European retailers. They forecast demand for hundreds of thousands of items across multiple stores daily. This “traditional AI” focused on prediction rather than generation.
The shift to generative AI opened new possibilities. As both a CTO and coder, he found AI removed tedious work and let him focus on higher-value tasks.
But he warns against over-dependence. His team built an email system that auto-generates responses to 10,000 daily customer inquiries. “Some users — they still had to validate the answers and they clicked accept quicker when they could have written a better answer on their own.”
Trauma-informed AI addresses critical human needs
As CEO of The Parasol Cooperative, Skelaney built Ruth, a trauma-informed chatbot designed to help people facing abuse, trafficking, technology-facilitated harassment, and other crisis situations.
“Mental health services in general are very overwhelmed,” Skelaney explains. “There’s overwhelming demand and there just aren’t enough workers in this space.” Ruth operates 24/7 in over 90 languages, filling gaps that traditional services can’t address.
The technology reads context in a way that standard chatbots don’t. When someone asks about the tallest bridge in New York after mentioning job loss, most AI systems provide factual information. Ruth recognizes the potential suicide risk and responds with care.
Ruth also handles technology-facilitated abuse, an area where human advocates often lack training. The system can guide someone through digital safety planning if they suspect cyberstalking or social media impersonation.
Skelaney emphasizes that Ruth supplements rather than replaces human support and points out the gaps we need to bridge in healthcare.
AI makes language learning truly personalized
Toy’s company Memrise is a language learning platform that creates an individualized curriculum for each user with AI.
“When we ask users what their goal is, the first answer is always to learn a language,” Toy explains. “But when you do the five whys, you get deeper — to travel, talk to family they married into, or make more money as a taxi driver in Turkey.”
Traditional language education couldn’t accommodate these different goals. Memrise solves this by atomizing dictionaries and importing real-world content. If you’re visiting London and love museums, you get museum vocabulary. The platform then matches YouTube videos to words you’ve learned, filtering content based on your specific knowledge level.
This approach replicates immersive learning without travel. “The best way to learn a language is to go to that country and survive. We can now replicate that experience because of what AI gives us.”
Key moments:
- How the Permission agent gets consumers paid for their data (2:42)
- Garbage in, garbage out and why half the web’s data is bad (7:02)
- GDPR’s bite and shifting ad behavior in Europe (8:41)
- Retail makes hundreds of thousands of forecasts a day (13:18)
- Gen AI speeds coding and writing but invites over-trust (17:05)
- Auto-replying to 10,000 emails a day and the rubber-stamp risk (22:00)
- Ruth spots crisis from context and pivots to safety (27:20)
- Safety by design with escalation and an instant escape feature (28:55)
- Ruth speaks 90+ languages and bridges hotline gaps (30:00)
- Step-by-step digital safety plans for cyberstalking (31:08)
- Memrise filters YouTube to only the words you know (36:18)
- Why AI makes reading, writing, and clear thinking more vital (40:04)
Q&A
Q: Where did the idea for permission-based data originate?
Silver: I once started a company called Real Age, which was the first version of asking permission because Real Age collected hundreds of data points about people’s health and well-being.
And then we would ask permission, would you like to receive products, information about products and services that can make you healthier? And for people who opted in, which was about 60%, the ROI to the advertiser wasn’t just a little bit better. It wasn’t 2 or 3x. It was 10x better.
Q: Are you worried about cognitive decline from AI dependence?
Greveling: You still need critical thinking skills, right? We still need to build this knowledge because whenever I use AI, it’s making like 40 or 60% mistakes sometimes.
So then I sometimes still think that, OK, let me just verify if it’s fine. Because you still need to be self-critical.
Q: What specific safeguards does The Parasol Cooperative have in place?
Skelaney: There’s an escalation protocol. So if somebody is saying that they’re having kind of a crisis or something like that, it definitely cannot under any circumstance be a sycophant and tell you that’s a really great idea, maybe you shouldn’t live anymore. Like that’s something that won’t happen because we’ve made very strict protocols against that. And it will, in addition to that, refer you to emergency services, phone numbers to call.
Q: How does The Parasol Cooperative handle technology-facilitated abuse specifically?
Skelaney: Ruth actually has protocols in place that can help people through a digital safety plan. So you can say, I think I’m being cyber stalked by my ex. Like, how do I know what can I do? And Ruth will take you through a process of assessing your vulnerabilities and then giving you instructions on how to mitigate the risk.
Q: How has generative AI impacted your view of language learning?
Toy: I think it makes reading, writing, and thinking more important. As we interact with these machines, if we can’t get our thoughts down and think clearly, the machines won’t do the things that we want them to do. Many people look at this and go, oh, we don’t have to worry about good grammar and writing exercises and getting our thoughts down on paper, but the opposite is actually true.
Q: What are the main user buckets for Memrise?
Toy: The first is people who are using it for travel and talking to other people and connecting with other people in life, like a new mother in law or a new company in an international location.
And the second bucket is for people who want to make their lives better, because a taxi driver in Turkey that speaks English makes more money. And a waiter in Peru that speaks English makes more money.
