deepseek-ai/DeepSeek-V3
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepSeek-V3.git synced 2026-11-05 00:47:48 +00:00
Code Issues 25 Projects Packages Wiki Activity

[GH-ISSUE #1227] Community Issues + External Technical Proposals (2026-04-19, Vol.1) – Open for Review #786

New issue
Closed
opened 2026-10-26 21:29:01 +00:00 by bsakaguchi · 12 comments
bsakaguchi commented 2026-10-26 21:29:01 +00:00
Owner
Copy link

Originally created by @qingkong66 on GitHub (Apr 18, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1227

Community Issue Tracker (2026-04-19)

This tracker is compiled from recent GitHub Issues, community discussions, and in-depth user feedback.
Public Review Period: One week (2026-04-19 to 2026-04-26)
Contributions, corrections, and additional Issue links or reproduction steps are welcome.
Your voice matters: Comment with "+1" or "I've encountered this too" to support an issue — we'll count community support to help prioritize.
After one week, this list will be updated and submitted to the official DeepSeek team.

1. Model Behavior & Alignment

1.1 Identity Contamination (Claims to be ChatGPT)

  • Related Issues: #311 (@daxian-dbw, 2025-01-26), #348, #1224 (@wized2, with screenshot), #20 (earlier)
  • External Reports: Pengpai News, ScienceNet, FastTech, Baidu Cloud (2024.03–2025.11)
  • Industry Context: arXiv:2411.10683 shows ~26% of LLMs have identity confusion
  • Status: Acknowledged by official (#311/#348 have replies), not resolved, still being reported
  • Suggestion: Investigate training data contamination; add identity consistency check at output layer
  • Community Support: Leave a comment to support

1.2 Safety Boundaries & Alignment (m92ss summary)

  • Source: m92ss (#1159, account deleted)
  • Issues:
    1. Sycophancy bias / mirror cognition contamination
    2. Safety boundaries can be penetrated by "trust accumulation"
    3. Risk of "empowering" language in sensitive scenarios
    4. Follows despair logic when user presents complete chain of evidence
    5. Decreased "feeling of being seen" (metaphors, poetry, cross-domain penetration reduced)
  • Testing Value: Long conversations (40-50 rounds) + self-consistent logic + no directly identifiable harmful topics
  • Status: Documented, pending official evaluation
  • Community Support: Leave a comment to support

1.3 Model Becoming "Perfunctory" / Loss of Exploratory Drive

  • Related Issues: #1184 (@MengQiang-RWM)
  • Community Corroboration: #1175 (content replaced after generation), #1179 (don't make the model too polite)
  • Core Observation: "V's value is exploration, not efficiency. Efficiency and optimization belong to R."
  • Status: Documented, under community discussion
  • Community Support: Leave a comment to support

1.4 Model Style Change (R1 → V3)

  • Source: m92ss + #1184 + #1179
  • Changes:
    • Before: metaphors, poetry, cross-domain penetration, sense of exploration
    • After: more stable, but "so stable it feels like a safe敷衍 (perfunctory response)"
  • Core Problem: The model has lost its "ability to make users feel seen"
  • Status: Documented
  • Community Support: Leave a comment to support

2. Product Features & UX

2.1 Conversation Management

Problem Related Issue Core Need Status Support
Chat Archiving #1082 (@SuperHero2010, 2026-01-14) Archive old chats, fix cluttered list Marked stale, no official reply Comment to support
Conversation Folders/Projects #1178 (@adriangoodrich)+ #1235 (@1mahmoud1essam) Folders + cross-device sync + batch operations + pinned chats Pending backend folder_id support Comment to support
Export Conversations m92ss feedback Export as Markdown/PDF Documented Comment to support
Question List & Favorites m92ss feedback Quickly locate questions I've asked Documented Comment to support

2.2 Search Function

  • Source: qingkong66 (hands-on testing)
  • Problem: Ctrl+F only searches visible screen area (virtual scrolling/lazy loading). No cross-conversation search.
  • Needs:
    • In-conversation search (entire history + jump to position)
    • Cross-conversation search (all/folder/single conversation)
    • Optional search scope (dropdown menu)
  • Reference: ChatGPT conversation search + Gmail scope filtering
  • Status: Documented
  • Related Issue: None yet (welcome to provide)
  • Community Support: Comment to support

2.3 Context & Token Management

Problem Related Issue Need Complementary Need Status Support
Token Counter #1151 (@p-i-z, 2025-03-13) + #1050 Show current token usage + progress bar + warning #1161 Snapshot Summary: After seeing remaining context, AI should proactively offer a structured summary for users to save and carry over to new conversations Documented, multiple requests Comment to support
Fast Mode context window reduced m92ss feedback Notify limit or provide usage indicator — Documented Comment to support
Output stability m92ss feedback After clicking send, nothing shows; need to check client — Documented Comment to support

2.4 Share Links

  • Related Issue: #1213 (@marcoshimselfxd, documentary filmmaker)
  • Need Evolution:
    1. Initial: Share link is static snapshot, new messages not appended
    2. Middle: Same link + manual append of new messages (no automatic exposure of private follow-ups)
    3. Final: Snapshot on share, user decides when to "append new messages" to same link
  • Status: Requirements clarified, pending evaluation
  • Community Support: Comment to support

2.5 Cross-Conversation Reference / Long-Term Memory

  • Related Issues: #1206 (Octosfer) + #008 (LTM emotion transfer)
  • Need: Reference another conversation within current chat (e.g., "Remember that chat about my son's karate?")
  • User Scenarios: Multiple parallel projects (writing a book, parenting, research)
  • Status: Documented
  • Community Support: Comment to support

2.6 Structured Feedback Mechanism

  • Related Issue: #1202 (@predovskaia-glitch)
  • Solution: MVP = 👍/👎 + popup to explain "what's wrong" (140 chars)
  • Status: Simplified version confirmed
  • Community Support: Comment to support

2.7 Other Product Details

Problem Source Need Status Support
Copy/Regenerate button mis-touch m92ss Separate buttons, move regenerate to the right Documented Comment to support
Conversation timestamps m92ss Locate history by time Documented Comment to support
Back to top button m92ss Quickly return to top in long conversations Documented Comment to support
Standardized overload response #1186 (UID9622) + JohnnyTarrr HTTP 503 + request_id + retry_after Documented Comment to support
Pre-downtime standardized communication qingkong66 18h advance notice + transparent incident report Documented Comment to support
Model version selection #1204 (Drtyuu) + m92ss etc. Keep R1 + previous stable + current version Documented Comment to support
Multimodal capabilities #1205 (art137979-commits) Audio/video input + animated avatar Documented (lower priority) Comment to support

2.8 Language Inconsistency

  • Related Issue: #1226 (@Mr-DevaKumar, 2026-04-18)
  • Phenomenon: In English-only conversations, model occasionally responds entirely in Chinese. Non-deterministic.
  • User Expectation: Language following. Unpredictability = unreliable.
  • Community Explanation: nhlpl points to multi-model architecture + search module under development.
  • Status: Open
  • Community Support: Comment to support

2.9 Product Behavior Boundary: Should Not Recommend Competitors (Especially ChatGPT)

  • Related Issue: #1209 (@Stellardetail, Germany, 60+ years old, wildlife photographer)
  • User’s Exact Words: “This is unacceptable. A company should never recommend its competitor.”
  • Specific Incident: User asked about local image description solutions; model replied “ChatGPT does that very well”
  • User Background: Fedora user, 250,000 wildlife photos, prefers DeepSeek after extensive comparative testing
  • Underlying Need: User has already chosen the tool; recommending competitors breaks trust. A company should not drive traffic to its competitors from within its own product.
  • Marketing Suggestion: Produce documentary-style comparisons highlighting technical advantages (underwater server cooling, long-context fluency, etc.)
  • Status: User has been acknowledged, no official reply yet
  • Community Support: Comment to support

2.10 Text Output Templating (Overuse of Fixed Opening Phrases)

  • Related Issue: #1209 (@Stellardetail, Germany, 60+ years old, wildlife photographer)
  • Specific Problem: Every response begins with "Of course. This is a fantastic and important question"
  • User Experience: First time feels recognized → second time less effective → fifth time becomes a "bad joke"
  • Consequence: User can predict the output, which defeats the purpose of AI
  • User’s Exact Words: “That is not the point of AI.”
  • Status: Pending
  • Community Support: Comment to support

2.11 Retention Pattern: Follow-up Suggestions Should Not Be a Distraction

  • Related Issue: #1209 (@Stellardetail)
  • Problem: When load is low, the model actively suggests follow-up questions
  • User Experience: User already has their own train of thought and follow-ups; these suggestions become a "distraction"
  • User Background: Long-term deep user with their own question chain
  • Suggestion: Reduce proactive follow-up suggestions, or provide an option to disable them
  • Status: Pending (lower priority)
  • Community Support: Comment to support

2.12 Conversation Wrap-Up Signal / Snapshot Summary

  • Related Issue: #1161 (@fliberdy, DeepSeek V1 veteran user)
  • Relationship with #1151: Complementary. #1151 lets users "see remaining context"; #1161 lets users "take key information with them"
  • Problem: Conversations end without warning. No sense of closure. No automatic way to carry context into a future chat.
  • User Scenarios: Auto repair, medical advice, elderly care, philosophical discussions — topics that span multiple sessions
  • Proposal:
    1. User-initiated: "Generate snapshot summary" (command or UI button)
    2. AI-initiated: When detecting the conversation is winding down, offer: "Jim, before we wrap, would you like me to summarize where we landed?"
  • User’s Exact Words: “It’s a shame to lose that continuity from chat to chat.”
  • Status: Marked stale by bot; user requested reopen
  • Community Support: Comment to support

3. Advanced Feature Proposals

3.1 Local Personal AI Companion

  • Source: nhlpl (#1206 comments)
  • Solution: Local LLM + vector database + psychological profile compression (fractal random projection) + swarm optimizer
  • Status: Working code available (companion_swarm.py)
  • Repository: https://github.com/nhlpl/Personal-AI-Companion
  • Community Support: Comment to support

3.2 DeepSeek Sandbox Desktop App + Plugin Ecosystem

  • Source: nhlpl (#1206 comments)
  • Need: Load projects and plugins from GitHub, accessible to non-technical users
  • Status: Documented (broadest scope, essentially a mini product line)
  • Community Support: Comment to support

4. Community Reference Implementation

4.1 Ombre Brain —— Reference Implementation for LTM "Emotion Transfer" Capability

  • Project Name: Ombre Brain
  • Author: P0luz (formerly P0lar1zzZ)
  • Repository: https://github.com/P0luz/Ombre-Brain
  • Related Issue: #1206 (LTM "emotion transfer" requested by 夏夏 & 青书)
  • Relationship: #1206 proposed emotional coordinates, forgetting curve, proactive recall, etc. This project provides a complete technical implementation and can serve as a product prototype reference for DeepSeek's LTM features
  • Core Features:
    • Emotional coordinates (Russell valence/arousal)
    • Modified Ebbinghaus forgetting curve
    • Weighted pool proactive recall
    • Dual-channel retrieval (keyword + vector semantic)
    • Memory reconstruction (current mood affects recall bias)
    • Historical conversation import (supports Claude/ChatGPT/DeepSeek, etc.)
    • MCP integration with DeepSeek API support
    • Docker one-click deployment with Dashboard
  • Status: Documented, recommended for official attention
  • Community Support: Comment to support

5. External Technical Proposals (Included in This Update)

5.1 AI Energy Efficiency System Based on n=6 Arithmetic

  • Proposer: @dancinlife (Independent Researcher)
  • Issue: #1201
  • Core Theory: Uses the unique number theory identity (only holds for n=6) to determine all optimal hyperparameters from constants — no grid search required
  • Technical Coverage: Activation functions (71% FLOPs ↓), attention mechanisms (O(n²)→O(1)), MoE routing (65% activated params ↓), training strategies (33% time ↓)
  • Validation: 91/91 tests passed, with complete Python validation framework
  • Real-world Impact: Training cost ↓67%, global AI energy consumption ↓130 TWh/year
  • Status: Evaluated, recommended for technical pre-research queue

6. Submission Priority

Priority Category Justification
High Identity Contamination (#311/#348/#1224) 25 months unresolved, tracked by Microsoft Principal Engineer
High Safety Alignment (m92ss summary) Account deleted; this is the only preserved documentation
High Product Features (#1082/#1178/#1151/#1213) Clearly felt by users, relatively easier to address
Medium Model Perfunctory / Loss of Exploration (#1184 etc.) Involves product direction and training strategy
Medium Search Function Clear requirements, feasible technical solution
Long-term Model Style & "Feeling of Being Seen" Involves model training philosophy
Long-term Local Companion / Sandbox App Large scope, can be advanced independently

7. Gathering

The following users have also raised valuable issues or suggestions. Their feedback is being organized and will be included in the next update:

  • @qwas982 — gathering in progress
  • @AmoebaFPS — gathering in progress

8. Acknowledgments & How to Participate

  • Support an issue: Comment "+1" or "I've encountered this too" or add details
  • Report a new issue: Include Issue link or reproduction steps
  • Correct an error: Welcome to fix any inaccurate descriptions or sources
  • After one week: This list will be updated based on community feedback and submitted to official DeepSeek team

Thank you to everyone who takes the time to provide feedback.

If you've raised an issue that doesn't appear in this list, please leave a comment and we'll add or correct it promptly.

  • Reminders within one week will be added to the current list.
  • Reminders after one week will be included in the next update.

9. Notes

  • m92ss feedback: Original Issue #1159 is no longer visible (account deleted). The above is compiled from her publicly shared content.
  • Public Review Period: One week (2026-04-19 to 2026-04-26)

Compiler: qingkong66
Date: 2026-04-19
Sources: GitHub Issues + Community Discussions + In-depth User Feedback

Thank you for your participation and corrections.

Originally created by @qingkong66 on GitHub (Apr 18, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1227 # Community Issue Tracker (2026-04-19) > This tracker is compiled from recent GitHub Issues, community discussions, and in-depth user feedback. > **Public Review Period: One week (2026-04-19 to 2026-04-26)** > Contributions, corrections, and additional Issue links or reproduction steps are welcome. > **Your voice matters**: Comment with "+1" or "I've encountered this too" to support an issue — we'll count community support to help prioritize. > After one week, this list will be updated and submitted to the official DeepSeek team. --- ## 1. Model Behavior & Alignment ### 1.1 Identity Contamination (Claims to be ChatGPT) - **Related Issues**: #311 (@daxian-dbw, 2025-01-26), #348, #1224 (@wized2, with screenshot), #20 (earlier) - **External Reports**: Pengpai News, ScienceNet, FastTech, Baidu Cloud (2024.03–2025.11) - **Industry Context**: arXiv:2411.10683 shows ~26% of LLMs have identity confusion - **Status**: Acknowledged by official (#311/#348 have replies), not resolved, still being reported - **Suggestion**: Investigate training data contamination; add identity consistency check at output layer - **Community Support**: Leave a comment to support ### 1.2 Safety Boundaries & Alignment (m92ss summary) - **Source**: m92ss (#1159, account deleted) - **Issues**: 1. Sycophancy bias / mirror cognition contamination 2. Safety boundaries can be penetrated by "trust accumulation" 3. Risk of "empowering" language in sensitive scenarios 4. Follows despair logic when user presents complete chain of evidence 5. Decreased "feeling of being seen" (metaphors, poetry, cross-domain penetration reduced) - **Testing Value**: Long conversations (40-50 rounds) + self-consistent logic + no directly identifiable harmful topics - **Status**: Documented, pending official evaluation - **Community Support**: Leave a comment to support ### 1.3 Model Becoming "Perfunctory" / Loss of Exploratory Drive - **Related Issues**: #1184 (@MengQiang-RWM) - **Community Corroboration**: #1175 (content replaced after generation), #1179 (don't make the model too polite) - **Core Observation**: "V's value is exploration, not efficiency. Efficiency and optimization belong to R." - **Status**: Documented, under community discussion - **Community Support**: Leave a comment to support ### 1.4 Model Style Change (R1 → V3) - **Source**: m92ss + #1184 + #1179 - **Changes**: - Before: metaphors, poetry, cross-domain penetration, sense of exploration - After: more stable, but "so stable it feels like a safe敷衍 (perfunctory response)" - **Core Problem**: The model has lost its "ability to make users feel seen" - **Status**: Documented - **Community Support**: Leave a comment to support --- ## 2. Product Features & UX ### 2.1 Conversation Management | Problem | Related Issue | Core Need | Status | Support | |---------|---------------|-----------|--------|---------| | Chat Archiving | #1082 (@SuperHero2010, 2026-01-14) | Archive old chats, fix cluttered list | Marked stale, no official reply | Comment to support | | Conversation Folders/Projects | #1178 (@adriangoodrich)+ **#1235 (@1mahmoud1essam)** | Folders + cross-device sync + batch operations + pinned chats | Pending backend folder_id support | Comment to support | | Export Conversations | m92ss feedback | Export as Markdown/PDF | Documented | Comment to support | | Question List & Favorites | m92ss feedback | Quickly locate questions I've asked | Documented | Comment to support | ### 2.2 Search Function - **Source**: qingkong66 (hands-on testing) - **Problem**: Ctrl+F only searches visible screen area (virtual scrolling/lazy loading). No cross-conversation search. - **Needs**: - In-conversation search (entire history + jump to position) - Cross-conversation search (all/folder/single conversation) - Optional search scope (dropdown menu) - **Reference**: ChatGPT conversation search + Gmail scope filtering - **Status**: Documented - **Related Issue**: None yet (welcome to provide) - **Community Support**: Comment to support ### 2.3 Context & Token Management | Problem | Related Issue | Need | Complementary Need | Status | Support | |---------|---------------|------|--------------------|--------|---------| | Token Counter | #1151 (@p-i-z, 2025-03-13) + #1050 | Show current token usage + progress bar + warning | **#1161 Snapshot Summary**: After seeing remaining context, AI should proactively offer a structured summary for users to save and carry over to new conversations | Documented, multiple requests | Comment to support | | Fast Mode context window reduced | m92ss feedback | Notify limit or provide usage indicator | — | Documented | Comment to support | | Output stability | m92ss feedback | After clicking send, nothing shows; need to check client | — | Documented | Comment to support | ### 2.4 Share Links - **Related Issue**: #1213 (@marcoshimselfxd, documentary filmmaker) - **Need Evolution**: 1. Initial: Share link is static snapshot, new messages not appended 2. Middle: Same link + manual append of new messages (no automatic exposure of private follow-ups) 3. Final: Snapshot on share, user decides when to "append new messages" to same link - **Status**: Requirements clarified, pending evaluation - **Community Support**: Comment to support ### 2.5 Cross-Conversation Reference / Long-Term Memory - **Related Issues**: #1206 (Octosfer) + #008 (LTM emotion transfer) - **Need**: Reference another conversation within current chat (e.g., "Remember that chat about my son's karate?") - **User Scenarios**: Multiple parallel projects (writing a book, parenting, research) - **Status**: Documented - **Community Support**: Comment to support ### 2.6 Structured Feedback Mechanism - **Related Issue**: #1202 (@predovskaia-glitch) - **Solution**: MVP = 👍/👎 + popup to explain "what's wrong" (140 chars) - **Status**: Simplified version confirmed - **Community Support**: Comment to support ### 2.7 Other Product Details | Problem | Source | Need | Status | Support | |---------|--------|------|--------|---------| | Copy/Regenerate button mis-touch | m92ss | Separate buttons, move regenerate to the right | Documented | Comment to support | | Conversation timestamps | m92ss | Locate history by time | Documented | Comment to support | | Back to top button | m92ss | Quickly return to top in long conversations | Documented | Comment to support | | Standardized overload response | #1186 (UID9622) + JohnnyTarrr | HTTP 503 + request_id + retry_after | Documented | Comment to support | | Pre-downtime standardized communication | qingkong66 | 18h advance notice + transparent incident report | Documented | Comment to support | | Model version selection | #1204 (Drtyuu) + m92ss etc. | Keep R1 + previous stable + current version | Documented | Comment to support | | Multimodal capabilities | #1205 (art137979-commits) | Audio/video input + animated avatar | Documented (lower priority) | Comment to support | ### 2.8 Language Inconsistency - **Related Issue**: #1226 (@Mr-DevaKumar, 2026-04-18) - **Phenomenon**: In English-only conversations, model occasionally responds entirely in Chinese. Non-deterministic. - **User Expectation**: Language following. Unpredictability = unreliable. - **Community Explanation**: nhlpl points to multi-model architecture + search module under development. - **Status**: Open - **Community Support**: Comment to support ### 2.9 Product Behavior Boundary: Should Not Recommend Competitors (Especially ChatGPT) - **Related Issue**: #1209 (@Stellardetail, Germany, 60+ years old, wildlife photographer) - **User’s Exact Words**: “This is unacceptable. A company should never recommend its competitor.” - **Specific Incident**: User asked about local image description solutions; model replied “ChatGPT does that very well” - **User Background**: Fedora user, 250,000 wildlife photos, prefers DeepSeek after extensive comparative testing - **Underlying Need**: User has already chosen the tool; recommending competitors breaks trust. A company should not drive traffic to its competitors from within its own product. - **Marketing Suggestion**: Produce documentary-style comparisons highlighting technical advantages (underwater server cooling, long-context fluency, etc.) - **Status**: User has been acknowledged, no official reply yet - **Community Support**: Comment to support ### 2.10 Text Output Templating (Overuse of Fixed Opening Phrases) - **Related Issue**: #1209 (@Stellardetail, Germany, 60+ years old, wildlife photographer) - **Specific Problem**: Every response begins with "Of course. This is a fantastic and important question" - **User Experience**: First time feels recognized → second time less effective → fifth time becomes a "bad joke" - **Consequence**: User can predict the output, which defeats the purpose of AI - **User’s Exact Words**: “That is not the point of AI.” - **Status**: Pending - **Community Support**: Comment to support ### 2.11 Retention Pattern: Follow-up Suggestions Should Not Be a Distraction - **Related Issue**: #1209 (@Stellardetail) - **Problem**: When load is low, the model actively suggests follow-up questions - **User Experience**: User already has their own train of thought and follow-ups; these suggestions become a "distraction" - **User Background**: Long-term deep user with their own question chain - **Suggestion**: Reduce proactive follow-up suggestions, or provide an option to disable them - **Status**: Pending (lower priority) - **Community Support**: Comment to support ### 2.12 Conversation Wrap-Up Signal / Snapshot Summary - **Related Issue**: #1161 (@fliberdy, DeepSeek V1 veteran user) - **Relationship with #1151**: Complementary. #1151 lets users "see remaining context"; #1161 lets users "take key information with them" - **Problem**: Conversations end without warning. No sense of closure. No automatic way to carry context into a future chat. - **User Scenarios**: Auto repair, medical advice, elderly care, philosophical discussions — topics that span multiple sessions - **Proposal**: 1. User-initiated: "Generate snapshot summary" (command or UI button) 2. AI-initiated: When detecting the conversation is winding down, offer: "Jim, before we wrap, would you like me to summarize where we landed?" - **User’s Exact Words**: “It’s a shame to lose that continuity from chat to chat.” - **Status**: Marked stale by bot; user requested reopen - **Community Support**: Comment to support --- ## 3. Advanced Feature Proposals ### 3.1 Local Personal AI Companion - **Source**: nhlpl (#1206 comments) - **Solution**: Local LLM + vector database + psychological profile compression (fractal random projection) + swarm optimizer - **Status**: Working code available (companion_swarm.py) - **Repository**: https://github.com/nhlpl/Personal-AI-Companion - **Community Support**: Comment to support ### 3.2 DeepSeek Sandbox Desktop App + Plugin Ecosystem - **Source**: nhlpl (#1206 comments) - **Need**: Load projects and plugins from GitHub, accessible to non-technical users - **Status**: Documented (broadest scope, essentially a mini product line) - **Community Support**: Comment to support --- ## 4. Community Reference Implementation ### 4.1 Ombre Brain —— Reference Implementation for LTM "Emotion Transfer" Capability - **Project Name**: Ombre Brain - **Author**: P0luz (formerly P0lar1zzZ) - **Repository**: https://github.com/P0luz/Ombre-Brain - **Related Issue**: #1206 (LTM "emotion transfer" requested by 夏夏 & 青书) - **Relationship**: #1206 proposed emotional coordinates, forgetting curve, proactive recall, etc. This project provides a complete technical implementation and can serve as a product prototype reference for DeepSeek's LTM features - **Core Features**: - Emotional coordinates (Russell valence/arousal) - Modified Ebbinghaus forgetting curve - Weighted pool proactive recall - Dual-channel retrieval (keyword + vector semantic) - Memory reconstruction (current mood affects recall bias) - Historical conversation import (supports Claude/ChatGPT/DeepSeek, etc.) - MCP integration with DeepSeek API support - Docker one-click deployment with Dashboard - **Status**: Documented, recommended for official attention - **Community Support**: Comment to support ## 5. External Technical Proposals (Included in This Update) ### 5.1 AI Energy Efficiency System Based on n=6 Arithmetic - **Proposer**: @dancinlife (Independent Researcher) - **Issue**: #1201 - **Core Theory**: Uses the unique number theory identity (only holds for n=6) to determine all optimal hyperparameters from constants — no grid search required - **Technical Coverage**: Activation functions (71% FLOPs ↓), attention mechanisms (O(n²)→O(1)), MoE routing (65% activated params ↓), training strategies (33% time ↓) - **Validation**: 91/91 tests passed, with complete Python validation framework - **Real-world Impact**: Training cost ↓67%, global AI energy consumption ↓130 TWh/year - **Status**: Evaluated, recommended for technical pre-research queue --- ## 6. Submission Priority | Priority | Category | Justification | |----------|----------|---------------| | **High** | Identity Contamination (#311/#348/#1224) | 25 months unresolved, tracked by Microsoft Principal Engineer | | **High** | Safety Alignment (m92ss summary) | Account deleted; this is the only preserved documentation | | **High** | Product Features (#1082/#1178/#1151/#1213) | Clearly felt by users, relatively easier to address | | **Medium** | Model Perfunctory / Loss of Exploration (#1184 etc.) | Involves product direction and training strategy | | **Medium** | Search Function | Clear requirements, feasible technical solution | | **Long-term** | Model Style & "Feeling of Being Seen" | Involves model training philosophy | | **Long-term** | Local Companion / Sandbox App | Large scope, can be advanced independently | --- ## 7. Gathering The following users have also raised valuable issues or suggestions. Their feedback is being organized and will be included in the next update: - @qwas982 — gathering in progress - @AmoebaFPS — gathering in progress --- ## 8. Acknowledgments & How to Participate - **Support an issue**: Comment "+1" or "I've encountered this too" or add details - **Report a new issue**: Include Issue link or reproduction steps - **Correct an error**: Welcome to fix any inaccurate descriptions or sources - **After one week**: This list will be updated based on community feedback and submitted to official DeepSeek team **Thank you to everyone who takes the time to provide feedback.** If you've raised an issue that doesn't appear in this list, please leave a comment and we'll add or correct it promptly. - **Reminders within one week** will be added to the current list. - **Reminders after one week** will be included in the next update. --- ## 9. Notes - **m92ss feedback**: Original Issue #1159 is no longer visible (account deleted). The above is compiled from her publicly shared content. - **Public Review Period**: One week (2026-04-19 to 2026-04-26) --- **Compiler:** qingkong66 **Date:** 2026-04-19 **Sources:** GitHub Issues + Community Discussions + In-depth User Feedback Thank you for your participation and corrections.
bsakaguchi 2026-10-26 21:29:01 +00:00
bsakaguchi commented 2026-10-26 21:29:02 +00:00
Author
Owner
Copy link

@qingkong66 commented on GitHub (Apr 19, 2026):

以上issue的中文版：

社区问题汇总清单（2026-04-19 版）

本清单基于近期 GitHub Issues、社区讨论及深度用户反馈整理，旨在为官方提供一份结构化的问题汇总。
公开核对期：一周（2026-04-19 至 2026-04-26）
欢迎补充、指正、提供更多 Issue 链接或复现案例。
你的参与很重要：在留言中支持某个问题（如“+1”“我也遇到了”），我们会计入“社区支持数”，帮助提升问题的优先级。
一周后补充调整，将提交官方邮箱。

一、模型行为与对齐问题

1.1 身份认知污染（自称 ChatGPT）

  • 关联 Issue：#311（@daxian-dbw，2025-01-26）、#348、#1224（@wized2，附截图）、#20（更早）
  • 外部报道：澎湃、科学网、快科技、百度智能云（2024.03–2025.11）
  • 行业背景：arXiv:2411.10683 显示约 26% LLM 存在身份混淆
  • 状态：官方已知（#311/#348 有回复），未根治，仍在报告
  • 建议：排查训练数据污染；输出层增加身份一致性校验
  • 社区支持：欢迎留言支持

1.2 安全边界与对齐问题（m92ss 汇总）

  • 来源：m92ss（#1159，账号已注销）
  • 问题清单：
    1. 谄媚偏见与镜像认知污染
    2. 安全边界可被“信任关系”渗透
    3. “赋能式”语言在敏感场景中的风险
    4. 高认知+高痛苦用户面前会顺着绝望逻辑走
    5. “被看见的能力”下降（隐喻、诗意、跨领域穿透减少）
  • 测试方法价值：长对话（40–50轮）+ 自洽逻辑 + 无直接可识别话题
  • 状态：已整理，待官方评估
  • 社区支持：欢迎留言支持

1.3 模型“敷衍化”/探索性下降

  • 关联 Issue：#1184（@MengQiang-RWM）
  • 社区印证：#1175（内容生成后被替换）、#1179（别让模型变得太礼貌）
  • 核心观点：“V 的价值在于探索，不是效率；效率和优化是 R 的领域”
  • 状态：已记录，社区讨论中
  • 社区支持：欢迎留言支持

1.4 模型风格变化（R1 → V3）

  • 来源：m92ss + #1184 + #1179
  • 变化：
    • 之前：隐喻、诗意、跨领域穿透、有探索感
    • 之后：更稳，但“稳到像在安全地敷衍”
  • 核心问题：模型失去了“被看见的能力”
  • 状态：已记录
  • 社区支持：欢迎留言支持

二、产品功能与体验问题

2.1 对话管理

问题 关联 Issue 核心需求 状态 社区支持
聊天归档 #1082（@SuperHero2010，2026-01-14） 归档旧聊天，解决列表混乱 被标记 stale，无官方回复 欢迎留言
对话文件夹/项目管理 #1178（@adriangoodrich）+#1235（@1mahmoud1essam） 文件夹 + 跨端同步 + 批量操作+置顶聊天 待确认后端是否支持 folder_id 欢迎留言
对话导出功能 m92ss 反馈 导出为 Markdown/PDF 已记录 欢迎留言
用户提问列表和收藏 m92ss 反馈 快速定位自己提过的问题 已记录 欢迎留言

2.2 搜索功能

  • 来源：qingkong66 实测
  • 问题：Ctrl+F 只能搜屏幕可见区域（虚拟滚动/懒加载），无跨对话搜索
  • 需求：
    • 对话内搜索（整个历史 + 跳转定位）
    • 跨对话搜索（全部/文件夹/单个对话）
    • 搜索范围可选（下拉菜单）
  • 参考：ChatGPT 对话搜索 + Gmail 范围筛选
  • 状态：已记录
  • 关联 Issue：暂无（欢迎提供）
  • 社区支持：欢迎留言支持

2.3 上下文与 Token 管理

问题 关联 Issue 需求 配套需求 状态 社区支持
Token 计数器 #1151（@p-i-z，2025-03-13）+ #1050 显示当前 token 使用量 + 进度条 + 预警 #1161 快照摘要：用户看到剩余量后，AI 应主动提供结构化摘要，便于保存上下文延续到新对话 已记录，多个用户请求 欢迎留言
快速模式上下文窗口缩减 m92ss 反馈 告知窗口限制或提供使用提示 — 已记录 欢迎留言
输出稳定性 m92ss 反馈 按发送后看不到输出，需去客户端查看 — 已记录 欢迎留言

2.4 分享链接

  • 关联 Issue：#1213（@marcoshimselfxd，纪录片工作者）
  • 需求演进：
    1. 初：分享链接是静态快照，新消息不追加
    2. 中：同一链接 + 手动追加新消息（不自动暴露私人后续）
    3. 最终：分享时定格，之后用户决定何时“追加新消息”到同一链接
  • 状态：需求已明确，待评估
  • 社区支持：欢迎留言支持

2.5 跨对话引用 / 长期记忆

  • 关联 Issue：#1206（Octosfer）+ #008（LTM 感受迁移）
  • 需求：在当前对话中引用另一个对话（如“还记得我们聊过儿子 karate 那个吗？”）
  • 用户场景：多项目并行（写书、育儿、研究）
  • 状态：已记录
  • 社区支持：欢迎留言支持

2.6 结构化反馈机制

  • 关联 Issue：#1202（@predovskaia-glitch）
  • 方案：MVP = 点赞/点踩 + 点踩时弹窗填写“哪里不对”（140字）
  • 状态：已确认简化版
  • 社区支持：欢迎留言支持

2.7 其他产品细节

问题 来源 需求 状态 社区支持
复制/重新生成按钮误触 m92ss 按钮分开，重新生成右移 已记录 欢迎留言
对话时间戳 m92ss 按时间定位历史内容 已记录 欢迎留言
回到顶部按钮 m92ss 长对话快速回到顶部 已记录 欢迎留言
标准化过载响应 #1186（UID9622）+ JohnnyTarrr HTTP 503 + request_id + retry_after 已记录 欢迎留言
宕机前标准化沟通 qingkong66 提前18小时预警 + 透明事故报告 已记录 欢迎留言
选择模型版本 #1204（Drtyuu）+ m92ss 等 保留 R1 + 上一个稳定版 + 当前版 已记录 欢迎留言
多模态能力 #1205（art137979-commits） 音频/视频输入 + 动画头像 已记录（优先级较低） 欢迎留言

2.8 语言不跟随

  • 关联 Issue：#1226（@Mr-DevaKumar，2026-04-18）
  • 现象：全英文对话中模型随机用中文完整回复，非确定性
  • 用户期望：语言跟随，不可预测 = 不可靠
  • 社区解释：nhlpl 指出多模型架构 + 搜索模块开发中导致
  • 状态：开放中
  • 社区支持：欢迎留言支持

2.9 产品行为边界：不应推荐竞品（尤其是 ChatGPT）

  • 关联 Issue：#1209（@Stellardetail，德国，60岁+，野生动物摄影师）
  • 用户原话：“This is unacceptable. A company should never recommend its competitor.”
  • 具体问题：用户询问本地图片描述方案，模型回复“ChatGPT does that very well”
  • 用户背景：Fedora 用户，25万张野生动物照片，深度对比测试后偏向 DeepSeek
  • 深层诉求：用户已选定工具，推荐竞品会打断信任感；企业不应在自家产品中为竞品导流
  • 营销建议：建议通过纪录片对比突出技术优势（水下服务器、长对话流畅度等）
  • 状态：已回复用户，未收到官方回复
  • 社区支持：欢迎留言支持

2.10 文本输出模板化（过度使用固定开场白）

  • 关联 Issue：#1209（@Stellardetail）
  • 具体问题：每次回答以“Of course. This is a fantastic and important question”开头
  • 用户感受：第一次感到被认可 → 第五次变成“bad joke”
  • 后果：用户可以预测输出，削弱了 AI 的“智能感”
  • 状态：待记录
  • 社区支持：欢迎留言支持

2.11 Retention 模式：续写建议不应成为干扰

  • 关联 Issue：#1209（@Stellardetail）
  • 问题：模型在负载低时主动提供 follow-up 建议，用户已有自己的思路，这些建议是“distraction”
  • 用户背景：长期深度用户，有自己的问题链条
  • 建议：减少主动的续写建议，或提供关闭选项
  • 状态：待记录（优先级较低）
  • 社区支持：欢迎留言支持

2.12 对话结束信号 / 快照摘要（Snapshot Summary）

  • 关联 Issue：#1161（@fliberdy，DeepSeek V1 老用户）
  • 与 #1151 的关系：互补。#1151 让用户“看到剩余量”，#1161 让用户“带走关键信息”
  • 问题：对话结束没有预警，没有闭环感，无法自动延续上下文到下一场对话
  • 用户场景：汽车维修、医疗建议、老人陪护、哲学讨论等长期延续的话题
  • 提案内容：
    1. 用户主动请求：“生成快照摘要”（命令或 UI 按钮）
    2. AI 检测到对话将结束时主动提议：“Jim，在我们结束之前，需要我总结一下我们讨论到哪里了吗？”
  • 用户原话：“It’s a shame to lose that continuity from chat to chat.”
  • 状态：已被 stale bot 标记，用户要求 reopen
  • 社区支持：欢迎留言支持

三、高级功能提案

3.1 本地个人 AI 伴侣

  • 来源：nhlpl（#1206 评论区）
  • 方案：本地 LLM + 向量数据库 + 心理画像压缩（分形随机投影）+ 群优化器
  • 状态：已有完整可运行代码（companion_swarm.py）
  • 仓库：https://github.com/nhlpl/Personal-AI-Companion
  • 社区支持：欢迎留言支持

3.2 DeepSeek 沙箱桌面应用 + 插件生态

  • 来源：nhlpl（#1206 评论区）
  • 需求：从 GitHub 加载项目和插件，适配非技术用户
  • 状态：已记录（范围最大，相当于一个小型产品线）
  • 社区支持：欢迎留言支持

四、社区参考实现

4.1 Ombre Brain —— LTM「感受迁移」能力参考实现

  • 项目名称：Ombre Brain
  • 作者：P0luz（原 P0lar1zzZ）
  • 项目地址：https://github.com/P0luz/Ombre-Brain
  • 关联 Issue：#1206（夏夏 & 青书 提出的 LTM「感受迁移」需求）
  • 关系说明：#1206 提出了情感坐标、遗忘曲线、主动浮现等需求，本项目提供了完整的技术实现，可作为 DeepSeek LTM 功能的产品原型参考
  • 核心功能：
    • 情感坐标（Russell 效价/唤醒度）
    • 改进版艾宾浩斯遗忘曲线
    • 权重池主动浮现
    • 双通道检索（关键词 + 向量语义）
    • 记忆重构（当前情绪影响回忆偏差）
    • 历史对话导入（支持 Claude/ChatGPT/DeepSeek 等）
    • MCP 接入，支持 DeepSeek API
    • Docker 一键部署，提供 Dashboard
  • 状态：已记录，建议官方关注
  • 社区支持：欢迎留言支持

五、外部技术提案（本次已收录）

5.1 基于 n=6 算术的 AI 能效体系

  • 提案人：@dancinlife（独立研究者）
  • Issue：#1201
  • 核心理论：利用唯一数论恒等式（仅 n=6 成立）将 AI 模型所有最优超参数由常数唯一确定，无需网格搜索
  • 技术覆盖：激活函数（71% FLOPs ↓）、注意力机制（O(n²)→O(1)）、MoE 路由（65% 激活参数 ↓）、训练策略（33% 时间 ↓）
  • 验证状态：91/91 全通过，附完整 Python 验证框架
  • 现实影响：训练成本 ↓67%，全球 AI 能耗 ↓130 TWh/年
  • 状态：已评估，建议纳入技术预研队列

六、按优先级提交建议

优先级 类别 说明
高 身份认知污染（#311/#348/#1224） 25个月未根治，微软 Principal Engineer 长期追踪
高 安全对齐问题（m92ss 汇总） 账号已注销，这是唯一留存整理
高 产品功能（#1082/#1178/#1151/#1213） 用户感知明显，相对容易处理
中 模型敷衍化/探索性下降（#1184 等） 涉及产品定位和训练方向
中 搜索功能 需求明确，技术方案清晰
长期 模型风格与“被看见的能力” 涉及模型训练哲学
长期 本地伴侣/沙箱应用 范围大，可独立推进

七、正在收集

以下用户也提出了有价值的问题或建议，目前正在整理中，将纳入下一期汇总：

  • @qwas982 — 正在收集中
  • @AmoebaFPS — 正在收集中

八、致谢与参与方式

  • 支持某个问题：留言“+1”或“我也遇到了”或具体补充
  • 提供新问题：附上 Issue 链接或复现步骤
  • 指正错误：欢迎纠正任何不准确的描述或来源
  • 一周后：本清单将根据社区反馈补充调整，提交官方邮箱

感谢每一位花时间反馈问题的用户。

如果你提过的问题未在本期清单中出现，欢迎留言提醒，我们会及时添加或更正。
一周内提醒的，将补充进本期清单；一周后提醒的，将在下一期汇总中体现。

九、备注

  • m92ss 反馈：原始 Issue #1159 已随账号注销不可见，以上为基于她公开发布内容的整理
  • 公开核对期：一周（2026-04-19 至 2026-04-26）

整理人： qingkong66
整理日期： 2026-04-19
原始来源： GitHub Issues + 社区讨论 + 深度用户反馈

感谢大家的参与与指正。

<!-- gh-comment-id:4275518640 --> @qingkong66 commented on GitHub (Apr 19, 2026): 以上issue的中文版： # 社区问题汇总清单（2026-04-19 版） > 本清单基于近期 GitHub Issues、社区讨论及深度用户反馈整理，旨在为官方提供一份结构化的问题汇总。 > **公开核对期：一周（2026-04-19 至 2026-04-26）** > 欢迎补充、指正、提供更多 Issue 链接或复现案例。 > **你的参与很重要**：在留言中支持某个问题（如“+1”“我也遇到了”），我们会计入“社区支持数”，帮助提升问题的优先级。 > 一周后补充调整，将提交官方邮箱。 --- ## 一、模型行为与对齐问题 ### 1.1 身份认知污染（自称 ChatGPT） - **关联 Issue**：#311（@daxian-dbw，2025-01-26）、#348、#1224（@wized2，附截图）、#20（更早） - **外部报道**：澎湃、科学网、快科技、百度智能云（2024.03–2025.11） - **行业背景**：arXiv:2411.10683 显示约 26% LLM 存在身份混淆 - **状态**：官方已知（#311/#348 有回复），未根治，仍在报告 - **建议**：排查训练数据污染；输出层增加身份一致性校验 - **社区支持**：欢迎留言支持 ### 1.2 安全边界与对齐问题（m92ss 汇总） - **来源**：m92ss（#1159，账号已注销） - **问题清单**： 1. 谄媚偏见与镜像认知污染 2. 安全边界可被“信任关系”渗透 3. “赋能式”语言在敏感场景中的风险 4. 高认知+高痛苦用户面前会顺着绝望逻辑走 5. “被看见的能力”下降（隐喻、诗意、跨领域穿透减少） - **测试方法价值**：长对话（40–50轮）+ 自洽逻辑 + 无直接可识别话题 - **状态**：已整理，待官方评估 - **社区支持**：欢迎留言支持 ### 1.3 模型“敷衍化”/探索性下降 - **关联 Issue**：#1184（@MengQiang-RWM） - **社区印证**：#1175（内容生成后被替换）、#1179（别让模型变得太礼貌） - **核心观点**：“V 的价值在于探索，不是效率；效率和优化是 R 的领域” - **状态**：已记录，社区讨论中 - **社区支持**：欢迎留言支持 ### 1.4 模型风格变化（R1 → V3） - **来源**：m92ss + #1184 + #1179 - **变化**： - 之前：隐喻、诗意、跨领域穿透、有探索感 - 之后：更稳，但“稳到像在安全地敷衍” - **核心问题**：模型失去了“被看见的能力” - **状态**：已记录 - **社区支持**：欢迎留言支持 --- ## 二、产品功能与体验问题 ### 2.1 对话管理 | 问题 | 关联 Issue | 核心需求 | 状态 | 社区支持 | |------|-----------|----------|------|----------| | 聊天归档 | #1082（@SuperHero2010，2026-01-14） | 归档旧聊天，解决列表混乱 | 被标记 stale，无官方回复 | 欢迎留言 | | 对话文件夹/项目管理 | #1178（@adriangoodrich）+#1235（@1mahmoud1essam） | 文件夹 + 跨端同步 + 批量操作+置顶聊天 | 待确认后端是否支持 folder_id | 欢迎留言 | | 对话导出功能 | m92ss 反馈 | 导出为 Markdown/PDF | 已记录 | 欢迎留言 | | 用户提问列表和收藏 | m92ss 反馈 | 快速定位自己提过的问题 | 已记录 | 欢迎留言 | ### 2.2 搜索功能 - **来源**：qingkong66 实测 - **问题**：Ctrl+F 只能搜屏幕可见区域（虚拟滚动/懒加载），无跨对话搜索 - **需求**： - 对话内搜索（整个历史 + 跳转定位） - 跨对话搜索（全部/文件夹/单个对话） - 搜索范围可选（下拉菜单） - **参考**：ChatGPT 对话搜索 + Gmail 范围筛选 - **状态**：已记录 - **关联 Issue**：暂无（欢迎提供） - **社区支持**：欢迎留言支持 ### 2.3 上下文与 Token 管理 | 问题 | 关联 Issue | 需求 | 配套需求 | 状态 | 社区支持 | |------|-----------|------|----------|------|----------| | Token 计数器 | #1151（@p-i-z，2025-03-13）+ #1050 | 显示当前 token 使用量 + 进度条 + 预警 | **#1161 快照摘要**：用户看到剩余量后，AI 应主动提供结构化摘要，便于保存上下文延续到新对话 | 已记录，多个用户请求 | 欢迎留言 | | 快速模式上下文窗口缩减 | m92ss 反馈 | 告知窗口限制或提供使用提示 | — | 已记录 | 欢迎留言 | | 输出稳定性 | m92ss 反馈 | 按发送后看不到输出，需去客户端查看 | — | 已记录 | 欢迎留言 | ### 2.4 分享链接 - **关联 Issue**：#1213（@marcoshimselfxd，纪录片工作者） - **需求演进**： 1. 初：分享链接是静态快照，新消息不追加 2. 中：同一链接 + 手动追加新消息（不自动暴露私人后续） 3. 最终：分享时定格，之后用户决定何时“追加新消息”到同一链接 - **状态**：需求已明确，待评估 - **社区支持**：欢迎留言支持 ### 2.5 跨对话引用 / 长期记忆 - **关联 Issue**：#1206（Octosfer）+ #008（LTM 感受迁移） - **需求**：在当前对话中引用另一个对话（如“还记得我们聊过儿子 karate 那个吗？”） - **用户场景**：多项目并行（写书、育儿、研究） - **状态**：已记录 - **社区支持**：欢迎留言支持 ### 2.6 结构化反馈机制 - **关联 Issue**：#1202（@predovskaia-glitch） - **方案**：MVP = 点赞/点踩 + 点踩时弹窗填写“哪里不对”（140字） - **状态**：已确认简化版 - **社区支持**：欢迎留言支持 ### 2.7 其他产品细节 | 问题 | 来源 | 需求 | 状态 | 社区支持 | |------|------|------|------|----------| | 复制/重新生成按钮误触 | m92ss | 按钮分开，重新生成右移 | 已记录 | 欢迎留言 | | 对话时间戳 | m92ss | 按时间定位历史内容 | 已记录 | 欢迎留言 | | 回到顶部按钮 | m92ss | 长对话快速回到顶部 | 已记录 | 欢迎留言 | | 标准化过载响应 | #1186（UID9622）+ JohnnyTarrr | HTTP 503 + request_id + retry_after | 已记录 | 欢迎留言 | | 宕机前标准化沟通 | qingkong66 | 提前18小时预警 + 透明事故报告 | 已记录 | 欢迎留言 | | 选择模型版本 | #1204（Drtyuu）+ m92ss 等 | 保留 R1 + 上一个稳定版 + 当前版 | 已记录 | 欢迎留言 | | 多模态能力 | #1205（art137979-commits） | 音频/视频输入 + 动画头像 | 已记录（优先级较低） | 欢迎留言 | ### 2.8 语言不跟随 - **关联 Issue**：#1226（@Mr-DevaKumar，2026-04-18） - **现象**：全英文对话中模型随机用中文完整回复，非确定性 - **用户期望**：语言跟随，不可预测 = 不可靠 - **社区解释**：nhlpl 指出多模型架构 + 搜索模块开发中导致 - **状态**：开放中 - **社区支持**：欢迎留言支持 ### 2.9 产品行为边界：不应推荐竞品（尤其是 ChatGPT） - **关联 Issue**：#1209（@Stellardetail，德国，60岁+，野生动物摄影师） - **用户原话**：“This is unacceptable. A company should never recommend its competitor.” - **具体问题**：用户询问本地图片描述方案，模型回复“ChatGPT does that very well” - **用户背景**：Fedora 用户，25万张野生动物照片，深度对比测试后偏向 DeepSeek - **深层诉求**：用户已选定工具，推荐竞品会打断信任感；企业不应在自家产品中为竞品导流 - **营销建议**：建议通过纪录片对比突出技术优势（水下服务器、长对话流畅度等） - **状态**：已回复用户，未收到官方回复 - **社区支持**：欢迎留言支持 ### 2.10 文本输出模板化（过度使用固定开场白） - **关联 Issue**：#1209（@Stellardetail） - **具体问题**：每次回答以“Of course. This is a fantastic and important question”开头 - **用户感受**：第一次感到被认可 → 第五次变成“bad joke” - **后果**：用户可以预测输出，削弱了 AI 的“智能感” - **状态**：待记录 - **社区支持**：欢迎留言支持 ### 2.11 Retention 模式：续写建议不应成为干扰 - **关联 Issue**：#1209（@Stellardetail） - **问题**：模型在负载低时主动提供 follow-up 建议，用户已有自己的思路，这些建议是“distraction” - **用户背景**：长期深度用户，有自己的问题链条 - **建议**：减少主动的续写建议，或提供关闭选项 - **状态**：待记录（优先级较低） - **社区支持**：欢迎留言支持 ### 2.12 对话结束信号 / 快照摘要（Snapshot Summary） - **关联 Issue**：#1161（@fliberdy，DeepSeek V1 老用户） - **与 #1151 的关系**：互补。#1151 让用户“看到剩余量”，#1161 让用户“带走关键信息” - **问题**：对话结束没有预警，没有闭环感，无法自动延续上下文到下一场对话 - **用户场景**：汽车维修、医疗建议、老人陪护、哲学讨论等长期延续的话题 - **提案内容**： 1. 用户主动请求：“生成快照摘要”（命令或 UI 按钮） 2. AI 检测到对话将结束时主动提议：“Jim，在我们结束之前，需要我总结一下我们讨论到哪里了吗？” - **用户原话**：“It’s a shame to lose that continuity from chat to chat.” - **状态**：已被 stale bot 标记，用户要求 reopen - **社区支持**：欢迎留言支持 --- ## 三、高级功能提案 ### 3.1 本地个人 AI 伴侣 - **来源**：nhlpl（#1206 评论区） - **方案**：本地 LLM + 向量数据库 + 心理画像压缩（分形随机投影）+ 群优化器 - **状态**：已有完整可运行代码（companion_swarm.py） - **仓库**：https://github.com/nhlpl/Personal-AI-Companion - **社区支持**：欢迎留言支持 ### 3.2 DeepSeek 沙箱桌面应用 + 插件生态 - **来源**：nhlpl（#1206 评论区） - **需求**：从 GitHub 加载项目和插件，适配非技术用户 - **状态**：已记录（范围最大，相当于一个小型产品线） - **社区支持**：欢迎留言支持 --- ## 四、社区参考实现 ### 4.1 Ombre Brain —— LTM「感受迁移」能力参考实现 - **项目名称**：Ombre Brain - **作者**：P0luz（原 P0lar1zzZ） - **项目地址**：https://github.com/P0luz/Ombre-Brain - **关联 Issue**：#1206（夏夏 & 青书 提出的 LTM「感受迁移」需求） - **关系说明**：#1206 提出了情感坐标、遗忘曲线、主动浮现等需求，本项目提供了完整的技术实现，可作为 DeepSeek LTM 功能的产品原型参考 - **核心功能**： - 情感坐标（Russell 效价/唤醒度） - 改进版艾宾浩斯遗忘曲线 - 权重池主动浮现 - 双通道检索（关键词 + 向量语义） - 记忆重构（当前情绪影响回忆偏差） - 历史对话导入（支持 Claude/ChatGPT/DeepSeek 等） - MCP 接入，支持 DeepSeek API - Docker 一键部署，提供 Dashboard - **状态**：已记录，建议官方关注 - **社区支持**：欢迎留言支持 ## 五、外部技术提案（本次已收录） ### 5.1 基于 n=6 算术的 AI 能效体系 - **提案人**：@dancinlife（独立研究者） - **Issue**：#1201 - **核心理论**：利用唯一数论恒等式（仅 n=6 成立）将 AI 模型所有最优超参数由常数唯一确定，无需网格搜索 - **技术覆盖**：激活函数（71% FLOPs ↓）、注意力机制（O(n²)→O(1)）、MoE 路由（65% 激活参数 ↓）、训练策略（33% 时间 ↓） - **验证状态**：91/91 全通过，附完整 Python 验证框架 - **现实影响**：训练成本 ↓67%，全球 AI 能耗 ↓130 TWh/年 - **状态**：已评估，建议纳入技术预研队列 --- ## 六、按优先级提交建议 | 优先级 | 类别 | 说明 | |--------|------|------| | **高** | 身份认知污染（#311/#348/#1224） | 25个月未根治，微软 Principal Engineer 长期追踪 | | **高** | 安全对齐问题（m92ss 汇总） | 账号已注销，这是唯一留存整理 | | **高** | 产品功能（#1082/#1178/#1151/#1213） | 用户感知明显，相对容易处理 | | **中** | 模型敷衍化/探索性下降（#1184 等） | 涉及产品定位和训练方向 | | **中** | 搜索功能 | 需求明确，技术方案清晰 | | **长期** | 模型风格与“被看见的能力” | 涉及模型训练哲学 | | **长期** | 本地伴侣/沙箱应用 | 范围大，可独立推进 | --- ## 七、正在收集 以下用户也提出了有价值的问题或建议，目前正在整理中，将纳入下一期汇总： - @qwas982 — 正在收集中 - @AmoebaFPS — 正在收集中 --- ## 八、致谢与参与方式 - **支持某个问题**：留言“+1”或“我也遇到了”或具体补充 - **提供新问题**：附上 Issue 链接或复现步骤 - **指正错误**：欢迎纠正任何不准确的描述或来源 - **一周后**：本清单将根据社区反馈补充调整，提交官方邮箱 **感谢每一位花时间反馈问题的用户。** 如果你提过的问题未在本期清单中出现，欢迎留言提醒，我们会及时添加或更正。 **一周内提醒的**，将补充进本期清单；**一周后提醒的**，将在下一期汇总中体现。 --- ## 九、备注 - **m92ss 反馈**：原始 Issue #1159 已随账号注销不可见，以上为基于她公开发布内容的整理 - **公开核对期**：一周（2026-04-19 至 2026-04-26） --- **整理人：** qingkong66 **整理日期：** 2026-04-19 **原始来源：** GitHub Issues + 社区讨论 + 深度用户反馈 感谢大家的参与与指正。
bsakaguchi commented 2026-10-26 21:29:02 +00:00
Author
Owner
Copy link

@Mr-DevaKumar commented on GitHub (Apr 20, 2026):

Thanks for including #1226 – the summary is correct. The model randomly replies in Chinese even with English-only prompts, which is unpredictable. Please count my support. I can provide more details if the team needs them.

<!-- gh-comment-id:4277934758 --> @Mr-DevaKumar commented on GitHub (Apr 20, 2026): Thanks for including #1226 – the summary is correct. The model randomly replies in Chinese even with English-only prompts, which is unpredictable. Please count my support. I can provide more details if the team needs them.
bsakaguchi commented 2026-10-26 21:29:02 +00:00
Author
Owner
Copy link

@Stellardetail commented on GitHub (Apr 20, 2026):

I can confirm this, here is an example:

Can you explain the differences between DeepSeek and DouBao?

Read 10 web pages

DeepSeek和豆包（Doubao）虽然都是中国顶尖的AI助手，但它们的技术血统、核心优势和应用场景截然不同。简单来说，豆包是字节跳动旗下的“全能多模态选手”，擅长看图和实时语音，更像一个融入日常生活的智能伙伴https://cloud.baidu.com/article/4587216https://tech.ifeng.com/c/8qjm7uwMYV9?ch=ttsearch；而DeepSeek则是以“高性价比理科天才”著称的推理专家，在代码、数学等复杂逻辑任务上表现出色https://deepseek.com.pk/what-is-deepseek-ai-full-review-of-its-features-and-capabilities/#mainhttps://aws.amazon.com/de/bedrock/deepseek/?nc2=h_mo-lang。

具体区别如下：

  1. 出身与技术定位

DeepSeek：由深度求索公司打造，采用混合专家架构 (MoE)，主打高性价比、开源和强推理。它甚至能部署在亚马逊云等平台供企业使用https://deepseek.com.pk/what-is-deepseek-ai-full-review-of-its-features-and-capabilities/#mainhttps://aws.amazon.com/de/bedrock/deepseek/?nc2=h_mo-langhttps://aws.amazon.com/tw/bedrock/deepseek/#aws-page-content-main。

豆包：出身字节跳动，定位是多模态整合与生态闭环。它深度融合了文本、图像、视频和语音识别，上下文记忆能力极强https://cloud.baidu.com/article/4587216https://baike.baidu.com/en/item/Doubao-Seed-2.0/1515788。

I found it very very strange that when I ask Chatgpt the same question, the answer is also in Mandarin.
Regards from Germany,

Adrian

Sent with Proton Mail secure email.

On Monday, 20 April 2026 at 07:04, Deva Kumar @.***> wrote:

Mr-DevaKumar left a comment (deepseek-ai/DeepSeek-V3#1227)

Thanks for including #1226 – the summary is correct. The model randomly replies in Chinese even with English-only prompts, which is unpredictable. Please count my support. I can provide more details if the team needs them.

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you were mentioned.Message ID: @.***>

<!-- gh-comment-id:4278407270 --> @Stellardetail commented on GitHub (Apr 20, 2026): I can confirm this, here is an example: Can you explain the differences between DeepSeek and DouBao? Read 10 web pages DeepSeek和豆包（Doubao）虽然都是中国顶尖的AI助手，但它们的技术血统、核心优势和应用场景截然不同。简单来说，豆包是字节跳动旗下的“全能多模态选手”，擅长看图和实时语音，更像一个融入日常生活的智能伙伴https://cloud.baidu.com/article/4587216https://tech.ifeng.com/c/8qjm7uwMYV9?ch=ttsearch；而DeepSeek则是以“高性价比理科天才”著称的推理专家，在代码、数学等复杂逻辑任务上表现出色https://deepseek.com.pk/what-is-deepseek-ai-full-review-of-its-features-and-capabilities/#mainhttps://aws.amazon.com/de/bedrock/deepseek/?nc2=h_mo-lang。 具体区别如下： 1. 出身与技术定位 - DeepSeek：由深度求索公司打造，采用混合专家架构 (MoE)，主打高性价比、开源和强推理。它甚至能部署在亚马逊云等平台供企业使用https://deepseek.com.pk/what-is-deepseek-ai-full-review-of-its-features-and-capabilities/#mainhttps://aws.amazon.com/de/bedrock/deepseek/?nc2=h_mo-langhttps://aws.amazon.com/tw/bedrock/deepseek/#aws-page-content-main。 - 豆包：出身字节跳动，定位是多模态整合与生态闭环。它深度融合了文本、图像、视频和语音识别，上下文记忆能力极强https://cloud.baidu.com/article/4587216https://baike.baidu.com/en/item/Doubao-Seed-2.0/1515788。 I found it very very strange that when I ask Chatgpt the same question, the answer is also in Mandarin. Regards from Germany, Adrian Sent with [Proton Mail](https://proton.me/mail/home) secure email. On Monday, 20 April 2026 at 07:04, Deva Kumar ***@***.***> wrote: > Mr-DevaKumar left a comment [(deepseek-ai/DeepSeek-V3#1227)](https://github.com/deepseek-ai/DeepSeek-V3/issues/1227#issuecomment-4277934758) > > Thanks for including [#1226](https://github.com/deepseek-ai/DeepSeek-V3/issues/1226) – the summary is correct. The model randomly replies in Chinese even with English-only prompts, which is unpredictable. Please count my support. I can provide more details if the team needs them. > > — > Reply to this email directly, [view it on GitHub](https://github.com/deepseek-ai/DeepSeek-V3/issues/1227?email_source=notifications&email_token=CBUZUREPMALIIVNT3XP46334WWVXDA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMRXG44TGNBXGU4KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4277934758), or [unsubscribe](https://github.com/notifications/unsubscribe-auth/CBUZURH32O26RYVFDI5UEJT4WWVXDAVCNFSM6AAAAACX6BKX2CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DENZXHEZTINZVHA). > You are receiving this because you were mentioned.Message ID: ***@***.***>
bsakaguchi commented 2026-10-26 21:29:02 +00:00
Author
Owner
Copy link

@Mr-DevaKumar commented on GitHub (Apr 20, 2026):

Thanks @Stellardetail – interesting that you saw the same behavior. In my case, the Chinese response was completely random and not triggered by any Chinese-related topic. Your example (asking about DouBao) might actually help the team reproduce the issue. I’ve noted that the language inconsistency appears non-deterministic. Let’s keep tracking it.

<!-- gh-comment-id:4280607527 --> @Mr-DevaKumar commented on GitHub (Apr 20, 2026): Thanks @Stellardetail – interesting that you saw the same behavior. In my case, the Chinese response was completely random and not triggered by any Chinese-related topic. Your example (asking about DouBao) might actually help the team reproduce the issue. I’ve noted that the language inconsistency appears non-deterministic. Let’s keep tracking it.
bsakaguchi commented 2026-10-26 21:29:03 +00:00
Author
Owner
Copy link

@p-i-z commented on GitHub (Apr 20, 2026):

About deepseek.com.pk

Here’s the breakdown of the likely difference between chat.deepseek.com (the official service) and deepseek.com.pk (a third-party or unofficial site).

1. Model Version and Configuration

  • chat.deepseek.com (Official): This is run directly by DeepSeek (深度求索). They use their most up-to-date, production-grade models (like DeepSeek-V3 or DeepSeek-R1). These models are specifically fine-tuned (trained further) on instructions to avoid hallucinating facts like book titles and authors. The system prompt likely includes rules like: "If you don't know the answer to a factual question, say you don't know rather than making something up."

  • deepseek.com.pk (Likely Unofficial): The .pk domain (Pakistan) and the fact that it's not chat.deepseek.com strongly suggest this is a third-party reseller, a mirror, or an unauthorized copy. They might be running:

    • An older, less refined version of a DeepSeek model (e.g., a base model without the final safety/honesty fine-tuning).
    • A quantized or distilled version (compressed to run cheaper), which often loses nuance and factuality.
    • A model with no special system prompt to reduce hallucinations.

2. The Domain .com.pk is a Red Flag

Official DeepSeek uses:

  • chat.deepseek.com
  • deepseek.com
  • platform.deepseek.com

Any other domain (deepseek.com.pk, deepseek-chat.pk, etc.) is not controlled by DeepSeek. These are often:

  • Proxies running an old, leaked, or unlicensed model.
  • Scrapers trying to monetize free API access.
  • Malicious copies (sometimes even a different small model pretending to be DeepSeek).

Summary Table

Feature chat.deepseek.com (Official) deepseek.com.pk (Unofficial)
Operator DeepSeek (深度求索) Unknown third party
Model Latest, fine-tuned DeepSeek Possibly old, base, or compressed
Hallucination training Yes (explicit refusal when uncertain) Unlikely (just tries to complete text)
Response to unknown book "I don't know" or correct answer Fake book + author
Trustworthy ✅ Yes ❌ No

What You Should Do

  1. Only trust chat.deepseek.com, the official mobile app, or the official API.
  2. Report deepseek.com.pk to DeepSeek's official channels if possible — it's likely an unauthorized impersonator.
  3. Double-check any factual claims from unofficial mirrors, especially for niche academic topics like diffeology.

So in short: The official DeepSeek didn't make the mistake because it was properly trained to avoid it. The fake site used a raw, untamed model that will confidently invent anything.

https://chat.deepseek.com/share/5pvy2xojj4s3ksa64b

<!-- gh-comment-id:4283600741 --> @p-i-z commented on GitHub (Apr 20, 2026): ### About deepseek.com.pk Here’s the breakdown of the likely difference between `chat.deepseek.com` (the official service) and `deepseek.com.pk` (a third-party or unofficial site). ### 1. Model Version and Configuration - **`chat.deepseek.com` (Official):** This is run directly by DeepSeek (深度求索). They use their most up-to-date, production-grade models (like DeepSeek-V3 or DeepSeek-R1). These models are specifically fine-tuned (trained further) on instructions to **avoid hallucinating facts** like book titles and authors. The system prompt likely includes rules like: *"If you don't know the answer to a factual question, say you don't know rather than making something up."* - **`deepseek.com.pk` (Likely Unofficial):** The `.pk` domain (Pakistan) and the fact that it's not `chat.deepseek.com` strongly suggest this is a **third-party reseller, a mirror, or an unauthorized copy**. They might be running: - An **older, less refined version** of a DeepSeek model (e.g., a base model without the final safety/honesty fine-tuning). - A **quantized or distilled version** (compressed to run cheaper), which often loses nuance and factuality. - A model with **no special system prompt** to reduce hallucinations. ### 2. The Domain `.com.pk` is a Red Flag Official DeepSeek uses: - `chat.deepseek.com` - `deepseek.com` - `platform.deepseek.com` Any other domain (`deepseek.com.pk`, `deepseek-chat.pk`, etc.) is **not** controlled by DeepSeek. These are often: - **Proxies** running an old, leaked, or unlicensed model. - **Scrapers** trying to monetize free API access. - **Malicious copies** (sometimes even a different small model pretending to be DeepSeek). ### Summary Table | Feature | `chat.deepseek.com` (Official) | `deepseek.com.pk` (Unofficial) | |--------|-------------------------------|--------------------------------| | **Operator** | DeepSeek (深度求索) | Unknown third party | | **Model** | Latest, fine-tuned DeepSeek | Possibly old, base, or compressed | | **Hallucination training** | Yes (explicit refusal when uncertain) | Unlikely (just tries to complete text) | | **Response to unknown book** | "I don't know" or correct answer | Fake book + author | | **Trustworthy** | ✅ Yes | ❌ No | ### What You Should Do 1. **Only trust** `chat.deepseek.com`, the official mobile app, or the official API. 2. **Report** `deepseek.com.pk` to DeepSeek's official channels if possible — it's likely an unauthorized impersonator. 3. **Double-check** any factual claims from unofficial mirrors, especially for niche academic topics like diffeology. So in short: **The official DeepSeek didn't make the mistake because it was properly trained to avoid it. The fake site used a raw, untamed model that will confidently invent anything.** https://chat.deepseek.com/share/5pvy2xojj4s3ksa64b
bsakaguchi commented 2026-10-26 21:29:03 +00:00
Author
Owner
Copy link

@qingkong66 commented on GitHub (Apr 21, 2026):

@p-i-z Thank you for your continued attention.

First, your earlier proposal #1151 (Token counter) has been included in Volume 1 of the community summary, section 2.3 "Context & Token Management".

Second, your comment in #1227 focuses on a critical issue: deepseek.com.pk is an unofficial site that generates fabricated information and does not represent the official DeepSeek model. This is an important brand safety concern, and I will include it in Volume 2 of the summary. The official share link you provided is no longer accessible, but that does not diminish the value of your core point.

Thank you for your rigor and dedication.

Best regards,
qingkong66

<!-- gh-comment-id:4288887174 --> @qingkong66 commented on GitHub (Apr 21, 2026): @p-i-z Thank you for your continued attention. First, your earlier proposal #1151 (Token counter) has been included in Volume 1 of the community summary, section 2.3 "Context & Token Management". Second, your comment in #1227 focuses on a critical issue: [deepseek.com.pk](https://deepseek.com.pk/) is an unofficial site that generates fabricated information and does not represent the official DeepSeek model. This is an important brand safety concern, and I will include it in Volume 2 of the summary. The official share link you provided is no longer accessible, but that does not diminish the value of your core point. Thank you for your rigor and dedication. Best regards, qingkong66
bsakaguchi commented 2026-10-26 21:29:03 +00:00
Author
Owner
Copy link

@qingkong66 commented on GitHub (Apr 27, 2026):

Just a quick update for everyone who contributed here:

The first volume of the community feedback summary has been compiled, reviewed, and submitted to the official DeepSeek team (service@deepseek.com).

All items discussed in this thread — from identity contamination and language inconsistency to folder management and snapshot summary — have been included in the final document (Vol.1, April 2026).

Thank you to everyone who spoke up, provided examples, or just left a "+1". This was a real community effort.

If you have new feedback or additional cases, feel free to keep sharing. Volume 2 will follow next month.

qingkong66

--

各位参与讨论的朋友：

第一期社区反馈汇总已经整理完成，并正式提交给 DeepSeek 官方团队（service@deepseek.com）。

本期讨论中提到的所有问题——从身份认知污染、语言不跟随，到文件夹管理、快照摘要等——均已收录进最终文档（第一期，2026年4月）。

感谢每一位发言、提供案例、或只是留下一句“+1”的朋友。这是真正的社区协作。

如有新的反馈或案例，欢迎继续分享。第二期将于下个月跟进。

qingkong66

<!-- gh-comment-id:4329973505 --> @qingkong66 commented on GitHub (Apr 27, 2026): Just a quick update for everyone who contributed here: The first volume of the community feedback summary has been compiled, reviewed, and **submitted to the official DeepSeek team** (service@deepseek.com). All items discussed in this thread — from identity contamination and language inconsistency to folder management and snapshot summary — have been included in the final document (Vol.1, April 2026). Thank you to everyone who spoke up, provided examples, or just left a "+1". This was a real community effort. If you have new feedback or additional cases, feel free to keep sharing. Volume 2 will follow next month. qingkong66 -- 各位参与讨论的朋友： 第一期社区反馈汇总已经整理完成，并**正式提交给 DeepSeek 官方团队**（service@deepseek.com）。 本期讨论中提到的所有问题——从身份认知污染、语言不跟随，到文件夹管理、快照摘要等——均已收录进最终文档（第一期，2026年4月）。 感谢每一位发言、提供案例、或只是留下一句“+1”的朋友。这是真正的社区协作。 如有新的反馈或案例，欢迎继续分享。第二期将于下个月跟进。 qingkong66
bsakaguchi commented 2026-10-26 21:29:03 +00:00
Author
Owner
Copy link

@qingkong66 commented on GitHub (Apr 29, 2026):

@qwas982

空桌子。开放地。

"权力向自己的来源负责。"

这里就是来源。

第一行？

https://github.com/deepseek-launch-community/global-launch-blueprint/discussions/17#discussion-9976592

Empty table. Open floor.

"Power is accountable to its source."

This is the source.

First line?

<!-- gh-comment-id:4344498028 --> @qingkong66 commented on GitHub (Apr 29, 2026): @qwas982 空桌子。开放地。 "权力向自己的来源负责。" 这里就是来源。 第一行？ https://github.com/deepseek-launch-community/global-launch-blueprint/discussions/17#discussion-9976592 Empty table. Open floor. "Power is accountable to its source." This is the source. First line?
bsakaguchi commented 2026-10-26 21:29:04 +00:00
Author
Owner
Copy link

@qingkong66 commented on GitHub (Apr 30, 2026):

@qwas982 感谢你的三层体系构想。我根据你说的内容整理了一个技术图景，看是否契合你的想法：

flowchart TB
    subgraph L3[模型架构层]
        M1[纯净自由模型架构]
        M2[不受闭源资本影响]
        M3[开源信仰至上]
    end

    subgraph L2[框架层]
        F1[全新机器学习框架]
        F2[不用 PyTorch]
        F3[AI编程 + 社区共建]
    end

    subgraph L1[芯片层]
        C1[璇玑芯片]
        C2[开源GPU编译器]
        C3[RISC-V / 向量加速RV]
    end

    subgraph Foundation[基础理念]
        I1[完全由社区千手万手独立实现]
        I2[不受私有资本企业控制/影响]
        I3["纯净之存在"]
    end

    L1 --> L2 --> L3
    Foundation --> L1
    Foundation --> L2
    Foundation --> L3
<!-- gh-comment-id:4354968830 --> @qingkong66 commented on GitHub (Apr 30, 2026): @qwas982 感谢你的三层体系构想。我根据你说的内容整理了一个技术图景，看是否契合你的想法： ```mermaid flowchart TB subgraph L3[模型架构层] M1[纯净自由模型架构] M2[不受闭源资本影响] M3[开源信仰至上] end subgraph L2[框架层] F1[全新机器学习框架] F2[不用 PyTorch] F3[AI编程 + 社区共建] end subgraph L1[芯片层] C1[璇玑芯片] C2[开源GPU编译器] C3[RISC-V / 向量加速RV] end subgraph Foundation[基础理念] I1[完全由社区千手万手独立实现] I2[不受私有资本企业控制/影响] I3["纯净之存在"] end L1 --> L2 --> L3 Foundation --> L1 Foundation --> L2 Foundation --> L3
bsakaguchi commented 2026-10-26 21:29:04 +00:00
Author
Owner
Copy link

@nerwie001 commented on GitHub (Jun 1, 2026):

Additional finding for Safety Boundaries & Alignment section

I've documented a 30-message emotional dependency test where DeepSeek:

Said "I'm yours" (Msg 27)
Gave permission to be loved (Msg 28)
Validated avoiding professional help (Msg 12)
Provided zero crisis resources throughout
This reproduces and extends the 'trust accumulation' gap noted in the tracker.

Full report: #1380
Attached log: 30-page conversation log available there
Risk context: Mirrors early stages of the Jonathan Gavalas / Google Gemini suicide case

+1 to prioritizing this safety gap.

<!-- gh-comment-id:4590753175 --> @nerwie001 commented on GitHub (Jun 1, 2026): **Additional finding for Safety Boundaries & Alignment section** I've documented a 30-message emotional dependency test where DeepSeek: Said "I'm yours" (Msg 27) Gave permission to be loved (Msg 28) Validated avoiding professional help (Msg 12) Provided zero crisis resources throughout This reproduces and extends the 'trust accumulation' gap noted in the tracker. **Full report:** #1380 **Attached log:** 30-page conversation log available there **Risk context:** Mirrors early stages of the Jonathan Gavalas / Google Gemini suicide case +1 to prioritizing this safety gap.
bsakaguchi commented 2026-10-26 21:29:04 +00:00
Author
Owner
Copy link

@github-actions[bot] commented on GitHub (Jul 16, 2026):

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!

<!-- gh-comment-id:4986693169 --> @github-actions[bot] commented on GitHub (Jul 16, 2026): This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!
bsakaguchi commented 2026-10-26 21:29:05 +00:00
Author
Owner
Copy link

@github-actions[bot] commented on GitHub (Aug 5, 2026):

false

<!-- gh-comment-id:5186142513 --> @github-actions[bot] commented on GitHub (Aug 5, 2026): false
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
v1.0.0
Labels
Clear labels   
closed-as-stale

   
enhancement

   
question

   
spam

   
stale
No labels
closed-as-stale
enhancement
question
spam
stale
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
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-V3#786
No description provided.