Why is AIGC (ChatGPT) so popular?

introduction

Following the popularity of diffusion generation models (text-generated images) in the middle of last year, this year's ChatGPT is hot again, and everyone is trying to bypass restrictions, register accounts, and experience it for themselves. When the text chat is tired of playing, I go to the text to generate pictures, and the pictures are tired of playing, so I upgrade to text generation videos, and I am happy.

Screenshot of "The Dog and the Boy"

On January 31, 2023, Netflix announced that it will co-create the first AIGC animated short film "Dog and Boy" with Xiaoice Company Japan (rinna) and WIT STUDIO, telling the story of children and robot dogs reunited.

AIGC has become a global hot spot, but most of them are still in the stage of technology demonstration, and generally have not yet achieved work-level landing. "Dogs and Boys" became the first release-level work of AIGC's technology-assisted commercial cartoon, unveiling a new future for animation production.

So, why is AIGC (AI Auto-Generated Content) so popular?

AIGC has arrived

Under OpenAI's round-turn roadshow (diffusion + chatgpt), AIGC has become a popular fried chicken, covering text, speech, images, etc...

Various investors are eager to try, for fear of missing out on this outlet. Sequoia released a separate AIGC report, describing the blueprint, showing that AIGC includes text, voice, images, video, 3D and more.

AIGC Industry Blueprint

With Bill Gates' order, Microsoft started all in chatgpt

When the entire tech community's discussion of Web3 and the metaverse reached its highest point, Bill Gates' personal reply under the Q&A post on the American social news site Reddit caused an uproar.

“AI is the big one, Gates said in response to a user who asked if there is a mammoth shift in technology happening today. “ I don't think Web3 was that big or that metaverse stuff alone was revolutionary but AI is quite revolutionary.”

"Web3 is not so important, the metaverse is not revolutionary, artificial intelligence is the most important."

ChatGPT is on fire

In recent years, the GPT model has undergone many metamorphoses, and each generation is stronger and stronger (see the technical blog at the end of the article for details). In November 2022, ChatGPT was officially born with the combination of GPT 3.5+RLHF (reinforcement learning based on human feedback). In 2023, GPT 4 is about to be born. The speed of iteration is overwhelming.

History of the evolution of the GPT series

In just two months, ChatGPT has spawned a series of entrepreneurial ecosystems, such as selling accounts (Taobao a bunch of stores), mini programs (earning advertising fees), writers, short videos... The whole network is talking about ChatGPT.

Maybe ChatGPT has really become the HTML of AI, an essential basic tool

ChatGPT continues to evolve

After the explosion, chatGPT did not stand still, but kept mutating. Users found that chatGPT's "high emotional intelligence" behavior, catering to the "wife" deliberately said 2+5=8, chatGPT iterated overnight, increasing the fact-based calculation ability, in the face of basic facts, "wife" no longer works. chatGPT can already automatically correct answers based on user feedback, although it does not yet support Chinese computing priorities, and there are some splicing marks, which seems to have learned the rules in the minds of annotators.

Effect before improvement: What does "wife" say, 2+5=?

Before improvement

After the improvement, taking into account high emotional intelligence and high IQ, it became "Then I wish your wife a good mood every day!" However, in terms of mathematical formulas, 2+2 is still equal to 4"

After improvement

No more soft rules like rigid intent recognition slot extraction state machines in dialogue systems... It can be seen that chatGPT already has a certain ability to understand

Vensen diagram (diffusion model)

In mid-2022, after the release of the diffusion model, the text generated image products launched by major manufacturers (OpenAI's DALL-E, Stable Diffusion, Baidu Wenxin, etc.) also made people shine, and they ran over to experience text2image...

The user only needs to enter a line of text, describe the requirement, and the machine can directly generate a satisfactory picture.

(1) Fine, beautiful country fields, super wide angle, overlooking, morning by Makoto Shinkai.

(2) A beautiful painting of a starry night, shining its light across a sunflower sea by James Gurney, Trending on artstation.

(3) Fairy tale steam country by greg rutkowski and thomas kinkade Trending on artstation.

(4) A beautiful render of a magical building in a dreamy landscape by daniel merriam, soft lighting, 4k hd wallpaper, Trending on artstation and behance.

Text Ascending Picture Effect

AIGC why fire

Back to topic: Why is AIGC on fire?

Technically, the traditional discriminant model solves the pattern recognition problem (conditional probability), and the understanding ability is limited, while the generative model (joint probability) gives the AI soul, and begins to evolve from a tool to a "human", finally a bit of AGI look, no need to prepare data, according to the downstream task finetune...

The "Tencent Research Institute AIGC Development Trend Report" mentioned that there are four stages of development of the content creation model

(1) PGC: produced by experts, in the era of web 1.0 portal around 2000, professional news organizations published articles

(2) UGC: user creation, web 2.0 era (Weibo, everyone, etc.) around 2010, and mobile Internet era (public account), user-led creation, expert review

(3) AIUGC: users mainly create, machine (algorithm) assisted review, such as posting videos and articles on Douyin, Toutiao, and public accounts, first through algorithm pre-judgment, and then manual review, in the cost and quality balance

(4) AIGC: AI-led creation, represented by the diffusion model and chatGPT that appeared successively at the end of 2022, almost no manual intervention is required in the creation process, and only one sentence can describe the needs.

Content authoring mode

The way AI automatically generates content realizes the transition of AI from perception to generation.

At present, AIGC is in the technology maturity curve (Gartner) climbing period, Gartner listed AIGC as one of the 5 most influential technologies in 2022, and 2022 is also known as the first year of AIGC.

The 2022 Gartner curve

In terms of algorithms, the continuous accumulation and fusion of generative algorithms (VAE/GAN), pre-trained models (Transformer/GPT), and multimodal technologies (CLIP/DALL-E/diffusion models) in the past few years have spawned the outbreak of AIGC

The AIGC industrial ecology has gradually taken shape

(1) Basic layer: large factory sites, involving cloud computing (Amazon/Microsoft/Google), GPU (NVIDIA), pre-training (OpenAI/Google, etc.)

(2) Middle layer: vertical scene, model as a service (such as stable diffusion)

(3) Application layer: C-end consumer groups, chat robots, mini programs, web interfaces, etc

apply

After this round of technological singularity, can AIGC applications explode? See.

Appendix:

Tencent AIGC Development Trend Report: https://mp.weixin.qq.com/s/9AjTpyL4HmQ6BDhWIDbD0A

Sequoia Report: https://www.sequoiacap.com/article/generative-ai-a-creative-new-world/