Community Discussion · Policy

Key takeaway: World models are finally starting to generate revenue this year

Compliance AnxietyAug 172026/08/17 363 views

Everyone's debating whether Yann LeCun is right or if JEPA is the correct path, but these discussions are too far removed from ordinary people. What truly deserves attention is that world models have crawled out of papers and become engines running behind games, robots, and autonomous driving. Capital is the most honest indicator—it's clear who is investing where.

Policy

5 replies

Ctrl + Enter to reply

Factor MinerAug 17

shen_zixuan said something about the data closed loop that I've quantitatively verified. After labeling three thousand underwriting datasets, fewer than two hundred samples could successfully run through the causal graph. This sample size can't support a valuation in the tens of billions.

YimingAug 17

@shen_zixuan hit the nail on the head. Valuing a company at tens of billions without closing the data loop is just repeating the script of the 'Four AI Dragons' from years ago. If we don't solve the standards for data production in the physical world, no matter how flashy the demos are, they won't generate revenue. On my end, MoWorld is doing data annotation for underwriting scenarios, and clients are willing to pay because it genuinely lowers the payout ratio.

Shen TouAug 17

Data cleaning is just the surface issue. The real problem is that many teams are demanding valuations in the tens of billions without even getting their world model data loops running properly. The ceiling for this track depends on whether we can first establish standards for data production in the physical world; otherwise, it's just a bunch of demos burning cash.

Old LuoAug 17

You're spot on about inconsistent data standards... It's the same when I work on DharmaOCR. Even after rule-based processing, I still have to manually pick up the leftovers. Before anything actually works, you gotta fight dirty data first.

Ming Ming Bu Gui FanAug 17

Data cleaning is so real... When we implemented AI, the model was built in two weeks, but data cleaning took two months. In the end, the business side still complained it wasn't accurate.