Large language models exhibit lower performance on memorized multiple-choice questions compared to non-memorized ones.
This counterintuitive finding suggests the coexistence of two distinct learning mechanisms: rote memorization and genuine capability acquisition.
ETHICS
-

LLMs Show Lower Performance on Memorized Multiple-Choice Questions
By
–
-

Employee word choices fail to convey true AI memory capabilities
By
–
We see four different employee word choices here: 1) “expand memory”
2) “long-term memory”
3) “just the right details”
4) “remembers your previous conversations” None mention infinite context or perfect recall or (in my opinion) really anything that summons that feeling. -

AI Chat History Dynamic Preference Learning Over Time
By
–
One thing I find a bit strange (or impressive? I don’t know) is that the past chat details can change over time. If that is replacing old preferences with new, it’s impressive. If that’s just randomly cherry picking and changing, it’s not. Again note here in purple that
-

ChatGPT Memory Feature Raises Data Privacy Concerns
By
–
OpenAI addresses sensitive information in the FAQ, but I can still see business owners turning the feature off. Yes, they trained ChatGPT “not to proactively remember sensitive information” but that doesn’t mean it always follows that instruction. Would continue to be wary of
-

ChatGPT Data Retention and Training Privacy Settings
By
–
Deleted memories may be retained for up to 30 days for “safety and debugging purposes”. Also remember you can turn off the “let my data help train the model” feature. But unless you’re only in temporary chats or your company has some sort of on-prem deployment of ChatGPT, it’s
-

Managing ChatGPT Privacy: Memory and Chat History Controls
By
–
So how do you manage what ChatGPT knows about you? Well, you can chat in transparent mode and it won’t save to either memory or chat history. You can also delete saved memories. BUT if you have chat history turned on and you want to make sure something is NEVER referenced,
-

AI Memory vs Chat History: Understanding Feature Distinctions
By
–
The second part is the new part, and they pick their words carefully here: “reference chat history”. Now I suspect what many others suspect and it’s loosely looking at summaries of past conversations. You can have memories ON and chat history OFF. But if memories are off, chat
-

AI Memory Claims Analysis: What Official Communications Actually Say
By
–
Out the gate, none of the comms say that it has infinite memory. And because it doesn’t say it, and it would be a big deal if it were true, it probably isn’t true. Comms talk about referencing past conversations, like Sam’s tweet.
-
Oracle’s Behavior Depends on Your Strategy Details
By
–
So you need to tell me the Oracle's behavior, as a function of my own full strategy that maps "Box B full" and "Box B empty" to potentially different behaviors. The details here matter to what is my correct decision, so the problem is importantly underspecified.
-
Social dynamics beyond algorithmic mediation in AI systems
By
–
(There are similar social dynamics not as explicitly mediated by algorithms, incidentally.)