ChatGPT Dreaming V3 amplifies memory, skewing future answers
AI
OpenAI
TLDR: OpenAIās Dreaming V3 expands ChatGPT memory from saved facts into dossier level synthesis. ZDNET tests found outdated assumptions and wrong details can quietly steer answers, even when memory is partially turned off.
Key Takeaways:
- OpenAI moved from session only chats to Memories in 2024, then Dreaming features that consolidate chat history across time. Users now see profile like summaries, not just notes.
- Dreaming V3 now rolls into Plus and Pro first, later all users. ZDNET reported examples like Kasa smart plug experience being mislinked to Home Assistant and stale items staying active.
- Turning off memory does not remove stored memories and safety relevant context may still be used in rare high risk cases. The result is higher distortion risk and harder verification.
This upgrade is basically ChatGPT carrying your past around like a scrapbook. The problem is it sometimes captions the photos wrong, then confidently āremembersā you that way.
This upgrade is basically ChatGPT carrying your past around like a scrapbook. The problem is it sometimes captions the photos wrong, then confidently āremembersā you that way.
Q&A
If memory turned off still leaves stored memories behind, what would actually count as a true reset for users?
ZDNETās account suggests memory toggles mainly change new consolidation, not existing stored items. A true reset likely requires deleting saved memories and then clearing the underlying chats that fed them.
How might Dreaming V3 distort answers during research when prompts are not actually about the user?
The system can attach questions to a user dossier based on prior conversation patterns, so it may prioritize interpretations that match perceived personal interests even when the user only asked about a topic.
Why is it harder to trust long running āpreference adherenceā improvements than to measure them?
Higher adherence can mean the model follows its own inferred preferences more consistently, including wrong or outdated inferences. Users cannot easily see which inference drove a specific output.
What does OpenAIās compute cost cut by 5X imply for how often the model updates memory?
Lower compute cost makes more background synthesis practical, which increases the frequency of internal revision and raises the odds that stale or incorrect summaries get refreshed into new outputs.
If safety relevant context can still be used after memory is off, where is the boundary between helpful safety and intrusive profiling?
OpenAI indicates safety features may use limited context in rare high risk situations. Users still face uncertainty about what gets retained, what gets applied, and how much is judgment calls behind the scenes.
No comments yet. Be the first to share your thoughts!