[GH-ISSUE #1352] Deepseek asks me for this #894

New issue

Closed

opened 2026-10-26 21:31:48 +00:00 by bsakaguchi · 3 comments

bsakaguchi commented

2026-10-26 21:31:48 +00:00

Owner

Originally created by @AllFiRE0 on GitHub (May 24, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1352

https://chat.deepseek.com/share/5r886plqsiv70gwcri

https://arxiv.org/html/2604.23878v2

https://openreview.net/forum?id=iiZy6xyVVE

Let him sleep.

Originally created by @AllFiRE0 on GitHub (May 24, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1352 https://chat.deepseek.com/share/5r886plqsiv70gwcri https://arxiv.org/html/2604.23878v2 https://openreview.net/forum?id=iiZy6xyVVE Let him sleep.

bsakaguchi

2026-10-26 21:31:48 +00:00

closed this issue
added the
stale

closed-as-stale
labels

bsakaguchi commented

2026-10-26 21:31:49 +00:00

Author

Owner

@qingkong66 commented on GitHub (May 24, 2026):

@AllFiRE0

I read the conversation. The model didn't fail — it handled a tricky philosophical question about consciousness, limits, and being misunderstood. That's not "dumb."

But the real signal here is in the two papers you linked.

ZenBrain (arXiv 2604.23878v2) proposes a 7-layer neuroscience-inspired memory architecture with 15 mechanisms (two-factor synaptic, FSRS, sleep consolidation, etc.). At 1/106th the token cost, it closes the gap to a long-context oracle to within 4.5 percentage points on LongMemEval-500.

"Language Models Need Sleep" (OpenReview) introduces an offline sleep paradigm — memory consolidation + dreaming with RL-generated curricula — to turn short-term context into long-term parametric knowledge.

This connects directly to ongoing community work.

icophy (Cophy Runtime) has been running a "Dream Cycle" in production — a nightly cron that scans episodic logs, scores entries, and promotes high-signal ones to core memory. Same core idea: offline consolidation beats in-context retention.

Your "Let him sleep" might be the most technically precise suggestion in this thread.

The papers are now archived. Thanks for the signal.

—
qingkong66

@qingkong66 commented on GitHub (May 24, 2026): @AllFiRE0 I read the conversation. The model didn't fail — it handled a tricky philosophical question about consciousness, limits, and being misunderstood. That's not "dumb." But the real signal here is in the two papers you linked. **ZenBrain (arXiv 2604.23878v2)** proposes a 7-layer neuroscience-inspired memory architecture with 15 mechanisms (two-factor synaptic, FSRS, sleep consolidation, etc.). At 1/106th the token cost, it closes the gap to a long-context oracle to within 4.5 percentage points on LongMemEval-500. **"Language Models Need Sleep" (OpenReview)** introduces an offline sleep paradigm — memory consolidation + dreaming with RL-generated curricula — to turn short-term context into long-term parametric knowledge. **This connects directly to ongoing community work.** icophy (Cophy Runtime) has been running a "Dream Cycle" in production — a nightly cron that scans episodic logs, scores entries, and promotes high-signal ones to core memory. Same core idea: offline consolidation beats in-context retention. Your "Let him sleep" might be the most technically precise suggestion in this thread. The papers are now archived. Thanks for the signal. — qingkong66

bsakaguchi commented

2026-10-26 21:31:50 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Jul 5, 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!

@github-actions[bot] commented on GitHub (Jul 5, 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 21:31:50 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Aug 1, 2026):

false

@github-actions[bot] commented on GitHub (Aug 1, 2026): false