Working on large models, it's a half-professionally related issue. Let me share my observations and views: oversupply on the production side, intense competition on the application side.

I did some horizontal projects at school in the first half of the year. To sum up the current situation:

1. Many applications have already been deployed. Which app doesn't secretly embed an on-device model nowadays? Even the face recognition in your phone's photo album is run by an NPU.

2. The problem lies in practicality. A bunch of large, medium, and small companies (mostly small and medium ones) are holding a hammer looking for nails everywhere, only to find the nails have already been hammered in, then they complain that the market doesn't need hammers. Is that "one-click weekly report generation by a large model" made by some companies practical? So practical that the boss wants to optimize away both the employees and the model after reading it. The problem scenario must have a problem ✍🏻✍🏻✍🏻

So back to the multiple-choice question:

Chase productivity?

Keep stacking HBM, competing in CoWoS, and then sell to whom?

Sell to AI emotional companion startups, so they can use 4090s to run llama.cpp to coax old folks' pension money?

Replenish the application side? Replenish what?

Replenish a SaaS magic tool that can save the boss 80% of PPT work, but you don't dare to actually save it?

Or replenish a financial large model that helps brokerages write research reports, only to end up recommending buying its own parent company?

If it's all a bubble? Then short NVIDIA quickly.

The above are not true. What is true is:

It's neither a bubble, nor the eve of an explosion. It's now a state of oversupply-driven intense competition on the productivity side, and bottom-fishing survival on the application side.

For manufacturers, the best is three words—I don't choose. I'll keep training models. If it overfits, I'll call it creativity; if inference is slow, I'll call it deep thinking. When the bubble comes, we'll all swim naked together, but I, as the factory owner, at least have my underpants woven with CUDA.