OpenAI-GPT4(V/Turbo)

Anthropic-Claude3

Google-Gemini 1.5, Gemma

Mistral AI & Meta & Bigcode…

Domestic players

OpenAI-DallE3, Sora

Stability AI – Stable Diffusion3

Midjourney – MJ V6

Google & Runway &…

Domestic players

Google/Deepmind: Robotic Transformer, Palm-E

Tesla: FSD, Optimus

Institute: VoxPoser, ALOHA

Domestic players

The underlying route is relatively clear, mostly based on the development and improvement of the Decoder-only architecture at the bottom of Transformer, combined with MOE and multi-modal embedding, etc.

The bottom layer route is clearer. Previously, we referred to the Stable Diffusion structure in the diffusion bottom layer, which is currently upgraded to Diffusion Transformer.

The core structure is complex and has many optimization points, including codec, patching, Diffusion module, Transformer module, Clip module and frame insertion module

The underlying technical route is unclear and there are various structures, such as Nerf, diffusion model, etc.

The underlying technical route is not clear, such as using VLM and LLM aggregation, but there are multiple technical logics at the reasoning and operation control levels

Llama2,Gemma

Starcoder2, CodeLlama

SD series

From the plug-in market to GPTs, the early popularity is high, but long-term excellent products still need to be developed by professional developers

Technologies such as MOE and RAG partially replace Finetuning, and FT still needs to improve the problems of catastrophic forgetting and update timeliness.

Combined with vector database, it has low cost, fast update frequency, improved answer accuracy and reduced illusion in the professional field. It will be applied on a large scale in the future.

Short-term rigid demand and long-term terminal demand may decrease. Currently, algorithms that help users automatically expand prompts and machine align are developing rapidly.

Divided into quantization, compression and pruning, etc., which are strictly needed for applications, especially locally deployed C-side and B-side models.

It is widely used, but in the early stage of development, the reasoning, planning and illusion problems of the underlying model need to be solved.

Chatbot

AI Search Engine

AI Image Editing/Generating

AI Writer Generator

AI Character Generator

AI Tools for Education

Other productivity

After nearly two years of continuous observation and tracking, the leading companies in some scenarios have become more recognizable, and outstanding companies have gradually opened up the gap with other players in the industry, such as MLLM and Chatbot, code generation applications, video generation application products, legal, Leading companies have emerged in the medical field

High investment density, talent density, capital density, and computing power density

The technology and team are highly identifiable, making it easy to monopolize core resources and gain a relative leading advantage.

There is a risk of technology iteration. Better open source and updated model architecture may reduce the technical advantages of existing underlying model manufacturers.

Choose scenarios that are easy to handle with the current model and don’t rush to optimize the model.

Quickly obtain PMF and revenue with the help of optimal model technology, with low technical risk

However, most products are homogeneous, and we need to find a business model that can continue to accumulate barriers.

Advance planning of products brought about by information gap

The company with the clearest investment niche and the most recognizable head

Scenario selection: There is an immediate need for pain points, the product has certain requirements, and customers are willing to pay a high order (preferably one that can generate revenue)

Data advantages: data sources for specific scenarios

Product iteration: can be made thicker, but migration cost is high

Operational data: user renewal/CAC/customer unit price, etc.