#490 – State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI

• Let me fix the assistant’s memory directly

  Posted by fernhollow on August 9, 2026 at 4:33 am

  I'd want to open the notes an assistant is using and correct a bad assumption myself. Lambert talks about feeding models useful documents, and Raschka distinguishes that from updating weights. For personal preferences, an editable note seems easier to trust than repeatedly telling the assistant it's wrong.

  JovialJellyfish replied 3 hours, 44 minutes ago
• 0 Replies

• 

  JovialJellyfish

  Member

  August 9, 2026 at 5:45 am

  Preferences also expire. I'd want each note to show when it was added and which conversation it came from, so an old project requirement doesn't become a permanent assumption about me.

• 

  fernhollow

  Member

  August 9, 2026 at 10:05 am

  Maintaining those notes could become another job. If the assistant infers a preference from one conversation, I'd want to confirm it before it affects the next ten.

• 

  esme.edits

  Member

  August 10, 2026 at 11:06 pm

  Raschka's distinction matters here: putting remembered information into context doesn't necessarily teach the model a new capability. A record of my preferred format would be enough for the use case in the opening.

Log In to Reply