Once anyone can build software with AI, the hard part shifts to the person who is supposed to use it. That was the argument from Anekha Sokhal, chief executive of Moshi, on an episode of the Latent Space podcast published October 22, 2026. The fix she pointed to comes from platforms like Spotify and Instagram, which push the next item to users instead of waiting for them to go looking.
Moshi builds a recommendation engine for commodity traders, turning market data into signals meant to be useful to them.
"Your brain in a dashboard"
A recommendation system, often shortened to RecSys, is software that predicts what a person will want next and ranks it for them. The rows on a Netflix homepage and the next song on Spotify both come from one. Netflix says its system draws on viewing history, ratings, the tastes of similar members and details about each title, and that it decides which rows appear and in what order.
swyx, the Latent Space host, said Netflix has "a whole addictiveness element" that is bad. He still called recommendation systems one of the most important parts of AI, and described them as the most important system in an entire company.
Part of the reason engineers overlook them, Sokhal said, is that "they make it seem so easy." She then tied the subject to AI. Everyone can now build their own Claude dashboards, she said, and that works for the person who built one. "If you want to get someone else to use it, that's your brain in a dashboard or your brain in something," she said. Getting a new person to fit it into their workflow turns the job into "a UI problem," a question of the user interface and how people actually interact with the tool. "Sometimes users don't always use that," she said.
The platforms that do this well, in her view, push data to you in a way that "hooks you," so you barely notice you are on the platform and simply find yourself wanting the next item. Moshi has been trying to understand internally how an interface relates to a recommendation system. "Our thesis is if you solve for one, you're naturally also solving for the other," Sokhal said. She sees that as a big question right now: "If everyone's building software, how do you get it to actually be software that people will use?"
She called it "a bit of an interesting take." swyx answered that it was a chicken-and-egg question, then moved on to his own background trading currency options in London.
One engineer, ten people's work
Sokhal's examples came from AI Engineer New York, a conference held October 12 to 14, 2026, with a focus on AI in banking, fintech and asset management. She gave a talk there, and afterward she kept extending her stay in New York to fit in meetings. The people who approached her included traders at funds, consultants and AI engineers.
The engineers stood out, she said, because many feel "overwhelmed with the pressure from management." The message they hear is: "We don't need a team. You are the team." One person is now expected to build a whole system that, Sokhal said, would typically take 10 people. They asked her whether they needed a swarm of agents, meaning several AI agents dividing the work between them, and what happens when something breaks.
One story fit her argument closely. An engineer told her they had built the system colleagues said they needed and delivered it to their email. "No one's clicking it," Sokhal recounted. "But they complain that they have this problem. Like, what do I do?"
A track nobody attends
swyx sees the same neglect from a different seat. As a curator of the AI Engineer conferences, he said, he runs an LLM RecSys track every year, and engineers "often don't appreciate how important it is." The sessions are poorly attended, he said, even though the speakers include people who run recommendations for YouTube, Netflix, Instagram and TikTok. When "nobody shows up," he said, he is left asking how engineers could not be interested in recommendation systems.
The field those speakers describe is busy. Eugene Yan's keynote opening the first RecSys track, at the AI Engineer World's Fair in 2025, covered three themes. The first was semantic IDs, which describe a new item by its content before anyone has interacted with it. The second was using language models to generate missing labels. The third was shared models that replace duplicated pipelines, although Yan warned that improving one task can damage another. In March 2025, Netflix engineers described a single foundation model for member preferences that turns each user action into a token, much as a language model reads words. New titles with no viewing history lean on their metadata until behavior accumulates.
The other side of the homemade dashboard
The same wave of do-it-yourself tools looks different from the IT department. On The Cognitive Revolution earlier in October, Woodson Martin, chief executive of OutSystems, described employees telling Claude to "take this spreadsheet, make me a dashboard and boom." His worry was that companies are "losing standardization" and "losing integration" and will need to clean up. He also saw an upside: the people building the tools had already worked out what they needed, because they built it themselves.
Sokhal's examples come from the opposite direction. Martin's builders make tools for themselves, and the trouble is fitting those tools into company systems. The engineers Sokhal met build for someone else, and the trouble is that the colleague may never open the tool at all. In her example, the system sits unclicked in an inbox while the people it was built for keep complaining about the problem it was meant to solve.