Success Diary #17: GPT4 Understanding

GPT is underestimated and the age of intelligence will eventually come. Individuals will also be greatly empowered

What GPT can do 🔗︎

• It can teach, it can solve problems.
• Selflessness , patience
• Ability to analogise , predict , step by step outputs

Ability to:

• Understanding complex ideas: language translation, translation of tone and style, cross-domain translation, GTP4 can understand complex ideas (e.g. understand human jokes)
• Spatial comprehension: gpt+ combined with stable diffision outputs images that meet expectations
• Visual ability: visual ability from text, drawing miniatures from text
• 3d modelling skills
• Code comprehension: combining tools to realise intent through multiple steps (solution provided)
• Mathematical ability: correct solution
• Interacting with the world: calling api’s and sending emails, browsing the web
• Physical interaction: can’t actually see or perform actions, but can interface through language and can perform needs to understand environment, tasks, actions and feedback.
• Interacting with humans: reasoning about others mental states strong

Flaws of GPT 🔗︎

Lack of planning in text generation, poorly done mental arithmetic: limitations of autoregressive models Models are cured once trained, cannot learn quickly or from experience (induction, solipsistic reasoning)

How to interact with GPT 🔗︎

1 Structuring of unstructured data 2 one shot: knowledge + known; solution

What are the future opportunities 🔗︎

Opportunities for AI: rapid digitisation, reducing the cost of physical modelling (unstructured -> structured). Everything can be done from the assistant’s point of view, from the point of view of solving problems, not from the point of view of so-called cheapness.

Scenario characteristics: 🔗︎

Broad, high-frequency, fast feedback, clear purpose function, low to medium decision dimension

Why Jitterbug beats Racer: let the lady post videos quickly via video + audio, not via algorithms.

Why GPT is so great. 🔗︎

GPT is opportunity based on probability and features, not on markers. For example: I’m in Beijing Tian (Anmen) Tian (gas) (pattern recognition or called compression, with a multi-layer structure, language carries all the wisdom of mankind)

Currently there are 2 categories:

1, fine-tune: tuning (without marking)

2、Hypertext model: knowledge base No tuning

long-chain ?