Just announced the new AI company!

  It's the first time in the history of the world.

AI is back in business!

A month after leaving India, Jiang Qing officially announced today the founding of a brand new AI company called Intent Lab.

There's only one word on the official website: create a "production-level system" by using "your intentions".

It is also a well-known recipe that "Old Partner" White Junje joined as co-founder.

In the comments section, the big cattle in the industry have lined up to congratulate.

As soon as the company debuted, Jiang Qing delivered its first surprise response.

Intent Lab was created by an "AI team" called "Fleet" that was able to independently complete full-link development tasks.

From architectural design, code writing, running tests, to code review and QA acceptance.

These complex links, which required the collaboration of the entire engineering team, are now taken over by this "fleet" autonomously.

With such an AI fleet, humans have only one thing left to do: make it clear what they want.

Fleet rebuilt the entire TRT-LLM.

On the two Grace Blackwell nodes, the same open-source model's throughput increased from 102 tokens per second to 647 tokens per second, a total of 6.3 times.

The key is that everything from performance analysis, optimization to deployment is done by the AI Agent itself.

The first shot was fired by a man who had just left INVEDA for a month.

In a line of hints, Fleet builds a database engine from scratch.

With no existing documentation, no reference code, the Fleet itself went all the way until all six million SQLite compatibility tests were passed.

Fleet also wrote an extremely fast file system, assuming that the users of this infrastructure were not humans but agents.

The file system is notoriously difficult, and an instantaneous state damage can take up to three days to check.

Fleet uses formal verification to pinpoint, locate, repair, and correct the bugs left behind by a coding agent.

Three things, three areas that are completely out of bounds.

But the point that Jiajing emphasizes is that they come from the same set of systems, the same method.

It is not a stack of three special tools, but a proof of a general ability in three directions.

In order to understand why Jiang Qing is now in the trenches, we need to dial back the timeline.

With a background in Qingdao and Dr. Berkeley, he gained fame for his development of the Coffee Framework.

As a benchmark for early-stage deep learning frameworks, Caffe has been widely adopted by giants such as Microsoft, Ingrid, and Adobe.

As a result, the industry has been criticized for saying that "physics can't get around Newton, deep learning can't get around Jiajing".

After completing his Ph.D. in 2013, Jiang joined Google Brain as one of the core authors of TensorFlow and participated in the creation of the first neural network worm, GoogLeNet, which surpassed human performance in image classification tasks.

Three years later, he was transferred to Facebook (now Meta).

In November of the same year, Caffe2Go was launched as a pioneer in end-to-end AI; in April 2017, Caffe2 was officially made open source;

He then worked on designing a prototype for the open model format, the ONNX industry standard that was later launched by Microsoft, Amazon, and other giants.

In May 2018, Caffe2 merged with PyTorch 0.4 to formally promote the landmark PyTorch 1.0.

Most of the AI tool chains you use today have his name engraved on the ground.

In March 2023, Jiang Qing announced his graduation from Ali University.

Just four months later, the AI startup Lepton AI surfaced with a $11 million angel wheel and a team of only 20 people.

One of Lepton's greatest achievements was to create a conversational search engine with less than 500 lines of code and fire it into GitHub.

This is exactly what brought it into the eyes of the Inuit.

In April, Invidia purchased Lepton AI for about $700 million. Jiang Qing became Vice President of Invidia Systems Software, leading DGX Cloud Lepton.

And that's the end we all saw.

A month after the acquisition, the Lepton platform ceased operations, prohibiting new user registration and official account cancellation. When the acquisition was announced, the Lepton core platform was open by 2026, and not a word was released until the end of June.

According to SemiAnalysis, the open source plan was ultimately rejected.

A man who has been open-source for twenty years has collided with a business model that can't open-source. On June 29, news of his departure came to light, just a year away from the acquisition.

A month later today, he gave his answer: Intent Lab.

In the last paragraph of his tweet, Jiang Qing wrote what he really bet on:

In the last 50 years, the most rational approach in the software industry has been to create a product and sell it to as many people as possible, even if everyone's needs are different.

If this layer is actually done, the logic will change.

There will be a lot more software in the world, and it will be much more customized.

And if the marginal cost of making software really collapses, "selling to everyone" is no longer the best option.

A company of fifty people can have a set of systems that are completely self-sustaining; a warehouse manager can have a set of scheduling tools that only serve that warehouse; a person who doesn't write code can say "what I want" directly.

Software will change from "commodities" to "tools".

If this road really goes down, who will be replaced first?