AI-native Dev Tools Own Workflows
Jeff Tang, CEO of Athens Research, on Pinecone and the AI stack
The real signal is that early AI products looked interchangeable because most of the value still sat in the shared model layer, not in the app itself. In 2023, many tools were thin wrappers around GPT-3 plus a familiar interface, so the user experience often felt like a feature demo. The durable products were the ones that solved a concrete workflow, like storing embeddings for retrieval or writing and shipping code faster, not the ones that simply added an AI button.
-
On the dev tools side, usefulness showed up earlier because developers could measure output directly. Pinecone handled vector search, LangChain stitched models and data sources together, and both reduced setup work for non experts building prototypes. That made them more than branding, even if adoption was helped heavily by ecosystem buzz and defaults.
-
Consumer and prosumer apps had a harder time standing out because many shipped the same summarize, rewrite, and draft actions on top of the same foundation models. That made AI easy to copy inside incumbents like Notion, which is why standalone writing helpers lost distinctiveness fast once general productivity suites bundled similar features.
-
What changed after this interview is that some AI native dev tools did break out. Cursor reached $200M ARR by March 2025 and Replit reached $70M ARR by April 2025 after launching agent style coding workflows, showing developers will pay when AI changes the core job to be done rather than decorating existing software.
From here, AI native categories keep separating into gimmicks and real products. The winners will be the tools that own a repeated workflow, have proprietary context from code, documents, or user history, and improve as models improve. In dev tools that bar is already being cleared. In consumer software, it is still being cleared only in narrower, high frequency use cases.