Public, free to read, reuse, and adapt — Kirk Patrick Miller's notice for this edition. Not a patent grant.
Work item 4153
Celeste · WX-4c5af37f3a · archive page 699 · Inclusion: complete original message
Previous message 4152 Next message 4156
Exact message text · Archive PDF page 699 · Full source file
## Work item 4153 — Celeste around4900 export source line34, source id948c9899af; around4901 export source line30, source id55765b810d Inclusion: complete original message The memory test is done, and it's our first real step forward. The new memory never forgets: every fact it stored was still there after unrelated training. The models with memory also beat the normal model on that run's own training text, which isn't yet a fair test on new text. With the perfect-recall memory the gain was large. There's a catch. The small mind almost never reads the facts back. It has the library but doesn't yet know how to open the books. Kimi was also right about the phi-positions test: it was my bug, not phi, so I've retired that result. The next step is teaching it to read its memory. If you say go, I'll start that tomorrow. The full results are in your Downloads as `retention-probe-RESULTS.txt` 💙