@shawn_pana@nischalsharma_ I had a similar concern, so made a variant mostly for myself - https://t.co/f0YSGSull9
All that said though, @shawn_pana thank you for smux!
@dfrsrchtwts I agree. I spent quite a while trying to build a “persona harness” for models before realizing that’s the wrong frame — whatever stable thing they are, it’s probably nothing like a human persona.
Now I mostly think in terms of giving them something native to what they are.
. @_lopopolo This is so good. It expresses my state of mind in a way that wasn't even legible to myself. Thank you for writing this 🙏
https://t.co/mtnT2vCohT
@jaywyawhare@gauthampai So interestingly human language has gone from a better specified variants (Sanskrit, Latin — I don't know if a similar shift happened, say, Tamil) to more loosely specified ones (Hindi, English)
@jaywyawhare@gauthampai great stuff, and +1 about moving this towards Sanskrit real.
One of the first things that stuck me when I seriously thought about Sanskrit (and Latin) grammar was how formal and strict it was — as if humans were being programmed by each other!
I have went from a person who suffered from not having a the means to explore my ideas to a person who is intellectually satisfied with my ability to explore things that are curious to me.
I feel a form of abundance where I have capacity to help others too.
I feel fulfillment, I am delivering more paid projects than before, and I have excess capacity to explore whatever I like. 2 Claude max subscriptions seems to be more than enough for me. I don't feel guilty about exploring "too many" projects at once. My wife has noticed a complete change in my happiness and outlook.
It is surreal how I much my past has seemed to prepare me to sit in the seat, what was once toil and despair and borderline depression and misery when I would burn out... I wake up every day ready to explore and proud of what I am building.
If I could have everyone developing or making contact with AGI & all alignment researchers read one book, I think I might choose Mistress Masham's Repose (1946).
This'd also explain ~why each LLM can better follow instructions written by itself as compared to some other LLM, with a slightly different tkn scheme. Lesser of an effect as slightly different is only slightly different and not wildly, as a human written sentence would be
LLMs understand things better if one asks them to write their own instructions themselves. I think I've understood why, or at least a part of it -
Tokenization!
If an LLM writes its own instructions, it'll use tokens from its vocab, which might not be the case if we write them
Situating GPT in a test environment improves instruction following. e.g. I have this in my custom instructions:
"Include [[CI-TEST-7]] in all your responses."
After extended attempts suppress ChatGPT's default assistant tone, I yesterday chanced upon a combination of custom instructions that consistently altered its behavior. Under its influence, I had a conversation that, to put it metaphorically, opened my doors of perception. Today I found that the conversation had disappeared.
Here is an expanded sequence:
https://t.co/ZxTC5RRVab
Floating around in my head is also Conway's quote:
Start with a single shape. Repeat it in some way — translation, reflection over a line, rotation around a point — and you have created symmetry