If you cheat, money will follow

Now that the crux of the "model collapse" problem has been found.

The next step is about to begin.

"We can start the backup language model first and use new ideas to cultivate the language model." Suzkevi said:

“This time, we have to feed the model with real data created by humans.”

"Detect and filter data quality to remove duplicates, errors, and biased content."

“And pay attention to enhancing the diversity of data, improving and enriching the content of generated data.”

"In addition, there is a need to enhance robustness technology, improve the anti-interference ability of language models, and automatically identify and generate data."

"I think as artificial intelligence continues to advance in the next few years, finding real data will be one of the main challenges for language models."

"Robustness" refers to the characteristic of a control system maintaining certain performance under certain parameter perturbations.

For example: Computer software does not freeze or crash in the event of input errors, disk failures, network overloads or intentional attacks, which is the robustness of the software.

"We have an advantage in this regard. Our partner Goose Factory's APP itself is a large-scale human natural language database." Li Fei said.

"However, this is not enough. We also need more human classic books, image data, news data, etc." Hinton said:

"These are the essence of human natural language."

"In addition, we recommend that Goose Factory appropriately restrict the permissions of various application program interfaces to extract data..." Krichevsky said:

"Not only do they download data, they also generate a lot of generative data that impacts our language models."

"I will coordinate all these with Goose Factory, and we will work hard to train and make up for the progress lost during this period." Chang Le said.

…………

That night, Chang Le informed Brother Xiao Ma, Mr. Ren and others about the "model crash" situation.

They were all surprised.

Surprisingly, the model is so similar to humans even when it creates problems.

The next morning.

Brother Ma showed up at the mountainside villa.

"Can I understand that with the development and large-scale promotion of language models, the data created by humans will become even more scarce in the future." Xiao Ma asked.

"It can be understood that the natural inertia of human beings will promote the proliferation of data generated by language models." Chang Le nodded:

"At least at this stage, there is still a gap between the data generated by the model and human data."

"Lack of authenticity and diversity of characteristics, repetition, bias, fallacies, etc."

Chang Le poured a cup of tea for Brother Ma:

“When this data is used for language model training, it will cause the model to ignore and forget the characteristics and information of the original data. This is the essence of model collapse.”

"Then this is a big business opportunity." Xiao Ma started to get excited.

"Yes, it is indeed a big business opportunity. At present, other language models at home and abroad have not yet advanced to this step, and have not yet realized the preciousness of human-generated data..." Chang Le said with a smile:

"At this time, you can invest or acquire some new media companies first, and the price will not be too high."

“Moreover, I also suggest that social apps appropriately cut off or limit the permissions of some application program interfaces to extract data. Their large-scale proliferation will accelerate the proliferation of data generated by language models.”

"I understand, I will arrange it right away." Xiao Ma nodded.

Jiang Xia came over with fruit: "Mr. Ma, eat some fruit."

"Thank you."

"I'll leave after lunch, so I won't be in a hurry." Chang Le said.

"Okay, I won't be polite, haha." Xiao Ma smiled.

"It should be so. In fact, investment and acquisition of some news media and online communities can only be regarded as a good supplement at this stage." Chang Le said.

"I know that even if our company's social APP itself is a huge database, it will inevitably be contaminated by data generated by language models in the future." Xiao Ma said.

"Yes, so it is necessary to build and expand an online library. For example, your company's reading APP is very good, covering novels, economics, biography, science, society, history, etc. This is a natural treasure house." Chang Le said.

"You reminded me that we will further expand the scale and scope of this APP." Xiao Ma said:

"In addition, our novel websites will also increase support and lean closer to traditional literature."

"We need to hurry up, because based on the lessons learned from this crash, once our backup model is activated, it may be purer and cleaner than before, and the time to market may be further accelerated." Chang Le said.

"No problem, I will personally follow up on this. Do you have any good suggestions on community forums?" Xiao Ma asked.

"Suggestion? In fact, it doesn't need to be large, as long as the real activity is high enough, like HOOP?" Chang Le said.

"HOOP? Haha, Mr. Chang, I didn't expect you to pay attention to the places where straight men gather?" Xiao Ma laughed.

past life.

Goose Factory has always looked down upon HOOP.

They believe that the user group of this website is too distinctive, too concentrated, too narrow, and not broad enough.

"I'm under 30 years old this year. It's normal to pay attention." Chang Le also chuckled.

HOOP, to be precise, should not be called a place where straight men gather, but a place where green men gather.

why?

Because some media believe that straight men across the country who are afraid of being "green", are currently being "green", or have been "green" gather at HOOP.

They use pure, direct and simple language to describe without any concealment the process of being afraid of "green", being "green", and already "green".

The general format is: "It's finally my turn, am I being cheated?"

Although HOOP officials never admit this point of view.

They define themselves as a true emotional community that integrates basketball, football, e-sports, sports, film and television, automobiles, digital, emotions, etc.

However, what is really hot and discussed is the cuckold posts on the pedestrian street.

Once a post has the following style, there will be many responses.

"JRS guys, let's see if I've been cheated on?"

"Usually, she doesn't go to bed until 1 a.m., but last night she went to bed at 11 p.m. I must have been cheated on."

"It turned out that she only ate seafood pot when eating hot pot. Now she likes to eat mandarin duck pot. I may have been cheated."

"She didn't let me see her phone. I must have been cheated on."

When straight men see this kind of post, they usually click on it.

Then, use your own hearsay and emotional experience to help the poster analyze it in detail.

Finally, I told the poster categorically and conclusively:

"Based on the experience of people who have been there, you should have been cheated."

Chang Le boldly guessed that after being influenced by HOOP's green man culture, the language model may develop special abilities in identifying "cuckolds".

To enhance the insight of the vast number of straight men who may become "green men".

aside.

Jiang Xia made suggestions.

"I think Red Book is also good. There are a lot of people there."

"Red book? That won't work." Chang Le quickly shook his head.

"Why not? The language model is allowed to train straight male culture, but not straight female culture? You are engaging in sexism." Jiang Xia glared at Chang Le.

Changle is unmoved and insists on his own opinion:

"This is not a question of straight women or not, but a question of IQ."

Jiang Xia: "..."

Chang Le said: "Only those with hemorrhoids would agree with the concept that buying cosmetics and bags is investing in oneself."

"With a salary of 3000 yuan a month, he is actually keen to discuss lipsticks worth thousands of yuan, bags worth tens of thousands, cars worth millions, and flats worth tens of millions..."

"If you crowdfund and take a photo, you will think that you are rich and beautiful, and you will be immersed in the illusory world you created all day long."

"It's not as exaggerated as you said. Although some people can't tell the difference between reality and fiction, there are still some who are very rational and the experiences they share are very effective." Jiang Xia said.

"Are the people you mentioned highly active?"

"..."