The reality was much more difficult than he expected. I recently had breakfast with Bornstein and his CTO, Maria Belousova, to learn about their startup, Daydream, which was funded with $50 million from VCs like Google Ventures. The conversation took an unexpected turn as the women told me about the surprising difficulty of translating the magic of AI systems into something that people actually find useful. Tea
His story helps explain something. My first newsletter of 2025 announced that it would be the year of the AI app. Although there are actually many such apps, they haven’t changed the world the way I thought they would. Since ChatGPT launched in late 2022, people have been amazed by the tricks AI can perform, but study after study has shown that the technology has yet to deliver a significant increase in productivity. (One exception: coding.) A study published in August found that 19 out of 20 AI enterprise pilot projects delivered no measurable value. I think productivity is going to increase, but it’s taking longer than people expect. Hearing the stories of startups like Daydream gives some hope that persistence and patience can actually make those successes possible.
fashionista fail
Bornstein’s original appeal to VCs seemed obvious: Use AI to solve tricky fashion problems by matching customers with the right apparel they’d be happy to pay for. (Daydream would cut down.) You’d think setup would be simple – just connect to the API for a model like ChatGPT and you’re ready to go, right? Umm, no. Signing up over 265 partners with access to over 2 million products from boutique shops to retail giants was the easy part. It turns out that fulfilling even a simple request like “I need a dress for a wedding in Paris” is incredibly complex. Are you the bride, the mother-in-law or a guest? what season is it? How formal is the wedding? What statement do you want to make? Even when those questions are resolved, different AI models still have different views on such things. “We found that the model sometimes left out one or two elements of the question due to the lack of consistency and reliability of the model, and because of hallucinations,” says Bornstein. A user in Daydream’s long-running beta test would say something like, “I’m a rectangle, but I need a dress to look like an hourglass.” The model will respond by showing dresses with geometric patterns.
Ultimately, Bornstein understood he had to do two things: postpone the app’s planned fall 2024 launch (though it’s available now, Daydream is still technically in beta until 2026) and upgrade his tech team. In December 2024 they hired Belousova, former CTO of Grubhub, who in turn brought a team of top engineers. Daydream’s secret weapon in the fierce talent war is the chance to work on an interesting problem. “Fashion is an interesting space because it’s about taste, personalization and visual data,” Belousova says. “It’s an interesting problem that hasn’t been solved.”
Furthermore, Daydream needs to address this problem twice-By first interpreting what the customer says and then matching their sometimes bizarre criteria with goods on the catalog side. with input like I need a revenge dress for the Bat Mitzvah where my ex-husband is attending with his new wife, That understanding is important. “We have this notion of shopper terminology and merchant terminology in Daydream, right?” Bornstein says. “Merchants talk categories and specialties, and buyers say things like, ‘I’m going to this event, it’s going to be rooftop, and I’m going to be with my boyfriend.’ How do you actually combine these two vocabularies into something at run time? And sometimes it has to be repeated several times in a conversation.” Daydreaming revealed that language is not enough. “We’re using visual models, so we really understand the products in a much more nuanced way,” she says. The customer can share a specific color or show a necklace they would wear.
Bornstein says Daydream’s subsequent reconstruction has yielded better results. (Although when I tried it on, I was shown beige athletic-fit trousers in addition to what I asked for in the request for black tuxedo pants. Hey, it’s a beta.) “We decided to go to a group of multiple models from the same call,” says Bornstein. “Everyone has a special call. We have one for color, one for fabric, one for weather, one for location.” For example, Daydream has found that for its purposes, OpenAI models are really good at understanding the world from a clothing perspective. Google’s Gemini is less so, but it’s fast and accurate.
<a href