Reborn in 2010, the AI ​​in my mind is too powerful.

The Google Brain project has always been an idea of ​​Jeff Dean.

He proposed creating such a project team back in 2008 and even contacted Andrew Ng specifically for it.

The two sides had already completed the preliminary work through step-by-step cooperation, and Google had officially launched the project at Google X Labs.

However, in 2010, Chen Yiran emerged with DeepFlow, catching Google off guard and ultimately halting the Google Brain project.

They were actually the first company to truly recognize the potential of Deepflow. At the time, everyone's attention was diverted by the amazing technical strength of Chen Yiran's EDA company. Although they thought Deepflow was very powerful, they still did not truly recognize the scalability of this framework.

After all, Chen Yiran has a very poor sense of sharing; he didn't even make it fully open source, nor did he publish any papers.

However, Jeff Dean, as Google's technological backbone, began working on DeepFlow in October 2010.

The whole purpose was to bypass him and try to achieve the same goal in another way.

In fact, Jeff Dean was inspired and wrote a very good paper, but his Transformer was like a toddler version of Deepflow in every respect.

Fortunately, Chen Yiran has not yet formally proposed the so-called concept of artificial intelligence. Google believes there is still time. They have only formulated a future strategy internally, prioritizing AI in all aspects of their strategy.

Unexpectedly, Chen Yiran's sudden press conference not only revealed the complete form of the Doubao AI model, but also bypassed them to intervene in Samsung's internal affairs.

This is completely unacceptable to Google, which is why they pressured Samsung in the hopes of forcing Chen Yiran to negotiate with them.

The result was quite surprising; Chen Yiran immediately flipped the table and launched a full-scale war.

What truly terrified Google was Chen Yiran's recent trip to Asia.

Nobody's a fool. Chen Yiran's private jet hadn't applied for an anonymous call sign at all. It went to South Korea so openly, and Google, with its 24/7 ADS-B monitoring, immediately discovered it.

Not to mention that Samsung made no attempt to avoid the issue at all. Their airport reception, which resembled that of a gangster, was captured on video and posted online by many fans who were there to photograph their idols.

After that, Google went on a spending spree to curry favor with Samsung, and the situation reversed dramatically.

Google even offered to pay Samsung to pre-install the entire Google suite.

After all, in the third quarter of 2011, global Android phone shipments were approximately 60 million, with Samsung accounting for half of them. Regardless of Samsung's reasons for abandoning the Google suite, it meant that half of the Android platform's traffic was simply gone.

In the days that followed, more and more bad news came.

Google and LG had almost reached an agreement and finalized a five-year contract for Google's entire product suite.

However, this company, which holds 10% of the global Android market, has now changed its stance and explicitly refused to sign the contract.

Although Google secured deals with HTC, Sony, and Motorola, these three companies together only account for 15% of Android platform shipments.

Andy Rubin has begun to despair and has even started suggesting that Google needs to make a comeback in China, since China accounts for 20% of Android shipments.

But that was obviously impossible. After weighing the options, Google could only have Jeff Dean release Transformer as soon as possible and start taking action in various ways to try to eliminate the impact of deepflow.

First, there's the Transformer paper, which completely targets DeepFlow, essentially doing the opposite.

DeepFlow's underlying technology is dynamic sparse attention, which only calculates strongly related word pairs and skips irrelevant calculations. Transformer does the opposite, calculating the relevance of each word to all words in the entire sentence.

Google is attempting to prove that sparsity is not necessary; given enough computing power, fully autonomous attention can learn on its own which words to focus on and which to ignore. This eliminates the need for DeepFlow's attention routing mechanism in the Transformer.

DeepFlow's positional encoding is learned and shares weights with multimodal attention. Transformer, on the other hand, directly calculates the positional encoding using sine and cosine formulas, requiring no training whatsoever.

Google wants people to think that location coding is simple enough to use only trigonometric functions, without the complex design of DeepFlow that is tied to multimodal processing.

In DeepFlow, the signal automatically selects which computation modules to go through based on the input content. In Transformer, after each layer, everything goes through the same fully connected feedforward network, regardless of the signal type.

Google claims that once enough layers are stacked, a fixed feedforward network naturally achieves hybridization without requiring architectural layer support.

DeepFlow included a multimodal attention interface from day one, while Transformer only handles text translation.

Google hopes to first outperform DeepFlow in the translation field, which DeepFlow cares least about, before announcing its expansion into vision and speech.

DeepFlow relies on architectural efficiency to achieve high accuracy with small models. The Transformer paper does not shy away from the computational cost, although its training cost is an order of magnitude higher than that of DeepFlow.

However, Google avoided this point in its paper, only saying that the Transformer can be extrapolated to greater accuracy.

Besides the various techniques presented in the paper, Google also mobilized its long-standing network of relationships, starting by criticizing the fact that DeepFlow's commercialization required licensing, arguing that open source does not necessarily mean a lack of the spirit of sharing.

Transformer was published as a paper, and on the day the paper was published, they essentially gave up their patent offensiveness, and Google simultaneously released the complete implementation code.

Then Google launched a translation API that small and medium-sized developers can use directly without installing a framework.

Then Google sent eight of its most technically influential people on a nationwide academic tour, distributing free computing vouchers everywhere, in an effort to get all relevant students to use their Transformer framework by default.

Furthermore, a special funding pool has been set up to support all research groups that are willing to replace DeepFlow with Transformer.

It has to be said that Google is still very powerful. If Chen Yiran hadn't paid any attention and Google had just been burning through money like this, it might have actually buried DeepFlow alive.

However, Chen Yiran was not too worried, because Transformer, in order to avoid the influence of deepflow, overemphasized the brute-force aesthetics of computing power.

However, current hardware limitations make it extremely difficult for them to match DeepFlow.

Moreover, Google has always been unable to solve a core problem: with the same computing power, DeepFlow will always be stronger.

Google's various actions can at most be considered a form of forcibly reserving seats.

Chen Yiran just needs to maintain his own advantages, or rather, he needs to attack Google in various ways.

For example, now that Apple has launched the Duck Complete Collection, he's ready to make a big splash.

Jasmine and Chen Yiran selected several interesting directions from the vast database of papers, and after some verification, decided to start a new project.

FlowEDA has accumulated enough experience to design its first chip: a tensor processor to sell to Apple, significantly reducing costs in search and video.