Building AI Products That Last: Lessons from SXSW 

At a packed session at SXSW this spring, Wharton Human-AI Research (WHAIR) faculty co-director Kartik Hosanagar and Microsoft chief product officer Aparna Chennapragada offered a candid and complementary set of lessons for builders navigating the fast-moving AI landscape. Their shared thesis: in AI product development, competitive advantage is both harder and easier to achieve than most people assume. 

The Model Is Not Your Moat 

Hosanagar opened with a cautionary tale familiar to many in the room. Jasper, a content generation startup that reached $1.5 billion in valuation by late 2022, saw its business undercut almost overnight when ChatGPT launched. This is because Jasper’s underlying magic resided in a foundation model any competitor could also access. Hosanagar’s conclusion: building on top of a shared model is not a strategy. The moat, if there is one, must come from data. 

But not just any data. Hosanagar’s more pointed claim is that simply accumulating proprietary data, and fine-tuning a model on it is not sufficient. His example: Bloomberg GPT, trained on Bloomberg’s substantial financial data, outperformed GPT-3 on finance tasks when it launched, only to be eclipsed by GPT-3.5 and GPT-4 within a year. Foundation models improve too fast for a one-time training advantage to hold. 

The durable edge, he argues, comes from architecting a data flywheel: a product designed so that user actions continuously generate reward signals that improve the AI in near real time, creating compounding advantages that widen with scale.  

The Self-Compounding Loop 

The most powerful version of this flywheel, Hosanagar argues, removes the human from the loop entirely. When an AI system can generate output, automatically verify whether that output is correct, and use that signal to improve, without waiting for human feedback, it can compound on its own. He calls this reinforcement learning with verifiable rewards, and points to software as the domain most naturally suited to it: code either compiles or it doesn’t, and unit tests either pass or they don’t. No human judgment required. This self-reinforcing loop, he argues, explains why AI coding tools have improved so dramatically in such a short time. 

Today’s Magic Is Tomorrow’s Commodity 

Chennapragada, drawing on her experience leading AI product development at Microsoft (and earlier as the founder of Google Lens), offered a practitioner’s complement to Hosanagar’s framework. Her core observation: every few months, a step-function improvement in model capability renders prior product workarounds obsolete. The right response is not to over-engineer around current limitations, but to ask whether your product gets better as models improve (or worse.) 

She introduced the concept of NLX (natural language experience) as the new UX: in a world of long-running agents, the design challenge is no longer about clicks and screens, but about how to communicate uncertainty, progress, and agency to users who may not see results for minutes or hours. 

Her other lessons were equally practical: small teams have a structural advantage because they’re forced to lean on the model rather than build engineering scaffolding around it; AI product development should start with “golden prompt sets” rather than product requirements documents; and the bottleneck in AI development has shifted from engineering to editorial judgment, knowing what to build, what good looks like, and when to ship. 

What This Means for Builders and Investors 

Together, the two speakers converged on a striking reframe: in AI, defensibility is less about model access and more about loop design. For investors, Hosanagar’s advice is blunt – stop evaluating AI companies on current performance, which is easily replicated, and start evaluating the mechanics of their data flywheel. For incumbents, a large existing dataset is an asset only if it’s actively generating feedback that improves the product in real time. 

And for builders, Chennapragada’s closing question may be the most useful framing of all: in a world where intelligence is no longer the scarce resource, what is?