Overnight, everyone is talking about Agent

First of all, it is worth noting the rise of Microsoft last night (which is quite significant for MSFT), and it turns out it's because of the series of Agents released at the London event.

Secondly, the 3.5 new version of Anthropic released last night actually operates computers, which are also agents. You can check out the video that was circulating last night for more details.

Not to mention, the conversation about Old Huang on BG2 also mentioned "We will soon see available agents."

What insights does this have for investment?

1. 2025 will be the year of the agent outbreak. Apple Intelligence and FSD are just around the corner, and the development of software agents for productivity scenarios is in full swing.

2. The reasoning model o1, which has found the most suitable scenario, is also an agent. The feature of o1 is to solve complex tasks of long thinking and multi-step reasoning. Furthermore, the single reasoning cost of o1 is said to be at the "millions" or even "billions" level of improvement, only complex tasks that directly replace human labor can afford the reasoning cost of o1.

3. How much computational power is needed for reasoning under agents? A certain foreign Broker provided a very good calculation yesterday.

Conclusion, training may be difficult to see a several-fold increase in computational power every year, but once the ROI, availability, and stability of reasoning are proven, the computational power for reasoning can still grow exponentially every year, boldly speaking, "exponential growth." Of course, the cost of reasoning is deflationary at the speed of light, but this is a good thing.

This also explains the revenue forecast given in OpenAI's financing materials, which is not impossible if viewed from the perspective of cost addition...

Finally, combining with the new high of NVIDIA mentioned in yesterday's article:

"Recently, there have been many dazzling NVIDIA product lines updated in the galaxy. What does this indicate? From NVL2 to NVL4 to NVL8, 36, 72, 144, 288, and even 512, 576. NVIDIA has laid out a long pipeline for the reasoning wave starting next year. Reasoning scenarios are already very diverse, okay, downstream can take what they need. Isn't this the so-called PMF, product market fit..." This article is from: Information Equality, original title: "Agent Overnight"