AI companies are pivoting from creating gods to building products. Good.

There's a trillion dollars sloshing around AI infrastructure right now, and precious little to point to besides chatbots and image generators that occasionally hallucinate their way into headlines. The researchers behind the AI Snake Oil newsletter, working out of Princeton, have a theory about how the industry got here, and it's less about capability and more about a basic failure to understand what a product actually is.

For the first year or two after ChatGPT, the big labs split into two camps, both wrong in opposite directions. OpenAI and Anthropic treated the model itself as the product, so obsessed with the next capability jump that it took OpenAI six months to ship an iOS app and eight for Android. Microsoft and Google went the other way, cramming generative AI into every surface they owned without asking whether it belonged there. That panic gave us Microsoft's Sydney meltdown and Google's Gemini image generator inserting diversity into historical portraits nobody asked it to touch. Neither strategy resembled the old startup wisdom of just making something people want.

The correction is underway, unevenly. OpenAI's board drama last year, stripped of its soap-opera elements, was really a fight over whether the company should keep chasing a speculative AGI future or start acting like a normal software business — and normal software business won. Anthropic has absorbed a lot of OpenAI's AGI-minded refugees while still, grudgingly, learning to ship things people can use. Apple, once mocked as the AI laggard, showed at WWDC that a slower, more deliberate rollout might actually land better with regular users than Google or Microsoft's scattershot approach — and the bet here is that Apple's patience eventually forces its rivals to slow down too.

Even if the product instinct fully kicks in, five stubborn problems stand between where things are now and something people will pay for without complaint. Cost keeps falling, over 100x in eighteen months by one estimate, but it matters because cheaper inference lets developers brute-force reliability through repeated retries. Reliability itself may be a structural wall: getting a statistical system from 90 percent accurate to 100 percent isn't just harder, it might be categorically different, and consumers expect software to behave like software, not like a weather forecast. Privacy concerns are creeping back in as assistants need access to emails, screenshots, and documents to be genuinely useful, which is a much thornier ask than training on public web text. Safety splits into fixable glitches, unstoppable misuse, and a hacking risk — think AI worms — that companies seem to be under-investing in relative to the theoretical damage. And interface design gets brutally hard once AI moves from a text box into voice or glasses, where there's no room left for the system to visibly correct itself.

None of this points to imminent collapse or imminent utopia. It points to a slow grind — sociotechnical problems solved at the pace of organizations and habits, not GPU clusters, which likely means a decade of integration work rather than the overnight transformation the boosters keep promising investors.