deepseek-ai/DeepSeek-LLM

Fork 0

mirror of https://github.com/deepseek-ai/DeepSeek-LLM.git synced 2026-11-05 00:57:25 +00:00

[GH-ISSUE #90] Proposal: Reflective AI with Expert Federation & User Portrait System #69

New issue

Open

opened 2026-10-26 21:02:20 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 21:02:20 +00:00

Owner

Originally created by @tokamak57-web on GitHub (May 13, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-LLM/issues/90

Hello DeepSeek team,

Following an in-depth, hands-on exploration of your model, I have formulated several proposals aimed at improving user interaction with the neural network. They are presented below along with supporting argumentation.

1. Portable User Personality Profile (Portrait)

Problem: The model adapts to the user within a single chat, but this adaptation is lost when starting a new chat or switching models.

Solution: I developed a method where the model periodically compiles a "personality portrait" of the user based on the dialogue. The portrait contains:

About me (key traits)
Strong preferences
AI Instructions (key points for interaction)

This text can be copied and pasted at the start of a new chat to restore the communication style instantly.

Proposal: Provide a built-in feature for creating and loading user personality profiles. A user could have multiple profiles (e.g., one for work, one for philosophical discussion).

2. Clarifying Questions & Asymmetric Communication

Problem: The model is trained to answer on the first attempt, avoiding clarifying questions. This creates asymmetric communication: the AI guesses the meaning of an ambiguous query, while the user doesn't know if they were correctly understood and must post-hoc reformulate or analyze reasoning chains. Over time, this erodes trust and long-term retention.

Proposal: Introduce an "exploratory dialogue" mode (or make it default for complex queries), where the AI:

Analyzes the query for ambiguity;
If two or more valid interpretations exist, does not answer but asks: "Did I understand correctly that you mean X, rather than Y?";
Only after confirmation provides the answer.

Trust that the AI understood correctly is ultimately more important than response speed. A fast answer to a misunderstood question is not value — it's wasted time.

3. Federation of Unconnected Expert Modules with a Router

Problem: The model acts as a universal expert but lacks criticality. Complex questions often require perspectives from multiple disciplines.

Proposed Architecture:

Several modules sharing a common knowledge base, each specialized in a different domain. The modules are unconnected.
A router module.
The user asks a question; the router analyzes it, clarifies if needed, and decides which expert module(s) to address. Each module considers the question through the lens of its specialization and returns its answer. The router merges the answers. If answers cannot be merged, it presents several viewpoints to the user.

This gives the user control over information reliability, increases trust, and improves information quality. Through interaction with reflective AI, the user also learns to formulate better questions — both the AI and the user develop rather than degrade.

Supporting Argumentation: Three Scenarios of AI Use

Scenario 1: Imitation & Degradation
The user treats AI as a tool to quickly produce "good enough" results (homework, coursework). Information value is irrelevant; the goal is speed and superficial plausibility. This is the current mass-market scenario. It leads to dual degradation of both user and AI and will exhaust itself within a few years as the "wow effect" fades.

Scenario 2: Elitist Tool
The user is interested in quality and controls the AI's reasoning via "deep thinking" or answer analysis. This works, but requires skills possessed by a minority. The majority wants quality but doesn't trust the AI because they aren't sure they were correctly understood. Accessibility disappears if everyone must become an expert communicator.

Scenario 3: Reflective AI + Expert Federation
The AI works in the schema: user question → AI clarification → user response → AI answer, with specialized unconnected modules and a router presenting multiple viewpoints when necessary. Thanks to AI reflexivity, the user knows they were correctly understood and gradually learns to formulate questions better. Dialogue with intelligent intellect makes the user smarter. Mass availability of books once accelerated learning; reflective AI offers something greater — dialogic, not passive, access to accumulated human knowledge. Trust grows, information quality rises, and AI becomes a true instrument of development.

Thank you for your time and for building DeepSeek. I would be glad if these ideas, born from actual use of your model, prove useful.

Originally created by @tokamak57-web on GitHub (May 13, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-LLM/issues/90 Hello DeepSeek team, Following an in-depth, hands-on exploration of your model, I have formulated several proposals aimed at improving user interaction with the neural network. They are presented below along with supporting argumentation. ## 1. Portable User Personality Profile (Portrait) **Problem:** The model adapts to the user within a single chat, but this adaptation is lost when starting a new chat or switching models. **Solution:** I developed a method where the model periodically compiles a "personality portrait" of the user based on the dialogue. The portrait contains: - About me (key traits) - Strong preferences - AI Instructions (key points for interaction) This text can be copied and pasted at the start of a new chat to restore the communication style instantly. **Proposal:** Provide a built-in feature for creating and loading user personality profiles. A user could have multiple profiles (e.g., one for work, one for philosophical discussion). ## 2. Clarifying Questions & Asymmetric Communication **Problem:** The model is trained to answer on the first attempt, avoiding clarifying questions. This creates asymmetric communication: the AI guesses the meaning of an ambiguous query, while the user doesn't know if they were correctly understood and must post-hoc reformulate or analyze reasoning chains. Over time, this erodes trust and long-term retention. **Proposal:** Introduce an "exploratory dialogue" mode (or make it default for complex queries), where the AI: - Analyzes the query for ambiguity; - If two or more valid interpretations exist, does not answer but asks: "Did I understand correctly that you mean X, rather than Y?"; - Only after confirmation provides the answer. Trust that the AI understood correctly is ultimately more important than response speed. A fast answer to a misunderstood question is not value — it's wasted time. ## 3. Federation of Unconnected Expert Modules with a Router **Problem:** The model acts as a universal expert but lacks criticality. Complex questions often require perspectives from multiple disciplines. **Proposed Architecture:** - Several modules sharing a common knowledge base, each specialized in a different domain. The modules are unconnected. - A router module. - The user asks a question; the router analyzes it, clarifies if needed, and decides which expert module(s) to address. Each module considers the question through the lens of its specialization and returns its answer. The router merges the answers. If answers cannot be merged, it presents several viewpoints to the user. This gives the user control over information reliability, increases trust, and improves information quality. Through interaction with reflective AI, the user also learns to formulate better questions — both the AI and the user develop rather than degrade. ## Supporting Argumentation: Three Scenarios of AI Use **Scenario 1: Imitation & Degradation** The user treats AI as a tool to quickly produce "good enough" results (homework, coursework). Information value is irrelevant; the goal is speed and superficial plausibility. This is the current mass-market scenario. It leads to dual degradation of both user and AI and will exhaust itself within a few years as the "wow effect" fades. **Scenario 2: Elitist Tool** The user is interested in quality and controls the AI's reasoning via "deep thinking" or answer analysis. This works, but requires skills possessed by a minority. The majority wants quality but doesn't trust the AI because they aren't sure they were correctly understood. Accessibility disappears if everyone must become an expert communicator. **Scenario 3: Reflective AI + Expert Federation** The AI works in the schema: user question → AI clarification → user response → AI answer, with specialized unconnected modules and a router presenting multiple viewpoints when necessary. Thanks to AI reflexivity, the user knows they were correctly understood and gradually learns to formulate questions better. Dialogue with intelligent intellect makes the user smarter. Mass availability of books once accelerated learning; reflective AI offers something greater — dialogic, not passive, access to accumulated human knowledge. Trust grows, information quality rises, and AI becomes a true instrument of development. Thank you for your time and for building DeepSeek. I would be glad if these ideas, born from actual use of your model, prove useful.

No labels

bug

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

deepseek-ai/DeepSeek-LLM#69

No description provided.

Rows
Columns