GateUser-2124263a

The recent development logic is: coding agents trigger democratization of development, which leads to a explosion in productivity enhancement tools, causing token consumption to increase at a geometric rate, which in turn makes large model companies profit handsomely, followed by a surge in data center business, then the semiconductor industry (essentially the mining equipment manufacturing sector) continues to develop at ultra-high speed, and finally basic infrastructure like power grids benefit tremendously—completing the loop.
I estimate the Federal Reserve sees the same situation, which is

If blockchain is attempting to democratize finance, then AI agents are democratizing development.
Using AI for development is literally more addictive than playing games. You don't want to sleep at all—it's so rewarding.

Ever since I learned how to play with sub agents, I felt like one Mac mini isn't enough anymore. Now I finally understand why some people buy machines 10 at a time.

Sub agents really do consume tokens like crazy. Running a few sub agents in parallel, and in just one day we burned through the entire daily Codex Plus usage.
Opening 5 Plus accounts costs $100
One Pro account costs $200
Not sure which one is more cost-effective?

I manually farm shrimp, then let the shrimp use Claude code, and then let Claude code use Codex. This should be the composability of AI, right 😂

It seems that the current contradiction is:
The growing demand of the people for shrimp farming versus the limited availability of cheap computing power.

We won big.
— Emperor Chuan's speech in the State of the Union address

Will Meta regret not acquiring Manus earlier?

If you treat the AI agent as the user, then skills become apps, opening up many more development possibilities and enabling functional composability.

I feel that in another year, I will become an AI Pilot.

Everyone, tell me, has President Xi received the invitation to join the Peace Committee from King Chuan?

After reading the new paper from DeepSeek, it seems that memory prices will continue to rise, which is positive for Micron.
The main idea of the paper is to separate the long-term memory part of the LLM from the model and store it in memory, thereby freeing up VRAM. This engineering approach is called n-gram.
I'm going to hold onto my DDR5 memory. Doubling it again this year is no problem.

The value of Manus mode boils down to three points:
1. It is the first company in human history to have over 80 million "employees";
2. Its essence is a set of "artificial intelligence operating systems";
3. The technological model it represents will ultimately be successful, which will mean a 0.5 level leap in human civilization.
Huxiu

Today I tested an AI agent for video production. The 60s video quality is acceptable and basically meets expectations, reaching a commercially acceptable level, but it consumed quite a few tokens, costing a full $20.
It's still a matter of severe computational power shortage. If the cost could be reduced to $0.2, it would be perfect.

The biggest expectation now is next month's DeepSeek V4, combined with Claude Code. Wouldn't that be amazing?

Looking forward to the evaluation of GLM-4.7 used in Claude code.

I believe Santa Rally is coming for sure.

Today was my first time trying NotebookLM, and I was a bit blown away.