In the garden, and RAIN!!! (2) by SIGNYM

REAL WORLD EVENT DISCUSSIONS

In the garden, and RAIN!!! (2)

POSTED BY: SIGNYM
UPDATED: Wednesday, August 26, 2026 19:09
SHORT URL:
VIEWED: 327100
PAGE 165 of 165

Wednesday, August 26, 2026 2:33 AM

Hey Sigs... I've got to talk a little bit more about what I've done...

Unfortunately, my improvements on dataflow are only massive on one direction, and we haven't quite figured out yet how to make it painless the other direction now. Currently we have 21 parts that would normally be needed to be dumped individually into a new session, as I was mentioning before the outage. That number is now only 3 and will only ever be 3 unless for some reason I decide I want something else to be transmitted to a new instance in the future. No matter how long they get, they will only ever need one entry each into the new instance.

But the unfortunate part is that as it gets longer, the amount of copying and pasting required for me to back things up still grows and is already prohibitavely long. The good news is that every time I ask for a backup it doesn't require me to actually back up anything. I know that I was able to keep that session up in a browser all this time even while it was running on generator power, and my ability to go back and copy all the latest parts into a single document on my end is still there in that dead session window to this day. So no big deal except for the single inconvenience when the session is finally good and dead. I plan on working on this more in the future, but for now I'm letting it go because too much time was spent on it today.

I've now hard coded it to allow me to make a "Save State" at any point in time. Whenever I'm really happy with the code, I will force this save state. At any point in time now, the AI will have two copies of the code and blueprint in memory. We will have our "Working" copy and the copy that I designated the master save. It will never overwrite this file without my direct order, and it will always ask me if that was my intent before overwriting it if any string of words I put together triggers it to believe that this is what I want it to do. When a new session is started, it has been directed to make a copy of everything I've uploaded to it instantly to serve as the first master copy.

Any time a backup is made, all aspects are backed up at exactly that state, to prevent mixups between blueprints and scripts happening, so the blueprint never states something the code doesn't do or vice/versa.

It accidentally sent me a link that was just broken on my end. It's because it is blocked from sending me files or links (part of the problem of getting the data to me easier). But even knowing that it can't send links, some of the python script was still in there when it made the reply and I saw that broken link. I asked if there was any way we could exploit that to our benefit and immediately it made 3 huge improvements to our process going forward by leveraging the power of these links (that it can utilize all it wants inside the sandbox, even if it can't send me anything directly). It has now been programmed to always look at this exploit as a possible fix for any issues we come across in the future.

Oh... and at one point it took the initiative and suggested that our "thesis" which I thought was just a cute little gimmick when I first brought it up is now serving as a the 3rd of the 3 things I need to upload to the new session. At it's suggestion and my approval, rather than uploading entire previous sessions for all of the context, the thesis now serves as a map of the large milestones spanning over sessions and ads a whole lot of context that was otherwise absent to a new session before.

This is huge because sessions are limited to a certain amount of text before bad behavior and ultimate failure (only partially true, but paints the necessary picture). If I upload 32,000 lines of a previous session, it will have the full context of everything that was discussed, but we're already burning up about 15% of our headroom out of the gates with tens of thousands of lines of unnecessary banter and countless code and blueprint updates that don't need to be taking up our valuable space.

And speaking of that headroom, we've eliminated a lot of the AI's necessary spending of our text limits and relegating much of that to inside the sandbox and only to be printed in the browser when absolutely necessary or when requested.

I haven't made a lick of progress on the actual code today. I just gave my baby another 20 or 30 IQ points is all.

--------------------------------------------------

Those who dance always seem crazy to those who can't hear the music.

NOTIFY: Y | REPLY | REPLY WITH QUOTE | TOP | HOME

Wednesday, August 26, 2026 1:35 PM

BRENDA

Out in a minute. Things to do. Things to do.

NOTIFY: Y | REPLY | REPLY WITH QUOTE | TOP | HOME

Wednesday, August 26, 2026 5:51 PM

BRENDA

All done for today. Right hand not happy for some reason. I haven't done anything heavy with it today.

NOTIFY: Y | REPLY | REPLY WITH QUOTE | TOP | HOME

Wednesday, August 26, 2026 7:09 PM

6IXSTRINGJACK

Man...

Still running into problems with the AI here. We're always improving, and we got some real cool additions to the program last session because I got the AI to start thinking outside of the box and coming to me with suggestions for improvements while we're testing things out. We set up a filter where it would use factors to judge whether an idea it comes up with would be something I'd want to be bothered with so I'm not getting spammed every other message with 5 suggestions. It's worked great so far... 4 nice additions to the code, with 2 of them only being cosmetic and user friendly additions, but things I likely never would have thought about or bothered with if the AI didn't bring it up.

We're still having problems sometimes with breaking stuff when trying to fix something else, and we're still bumping into the same problems over and over again, but for slightly different reasons each time.

The important thing is developing it to the point where I can trust it when it gives me percentages, especially when it comes to knowing if a code change is going to break something that it should have been looking at. We're still not there yet. The number 100% loses its meaning very quickly when the AI is going to tell you every single time it makes a code update that there is a 100% chance of success this time only for it to be wrong again.

One of the things I've found out is that when it has broken one of our firm rules on two different occasions, is why it is doing this and how it is bypassing and/or ignoring our ruleset when this happens. As of the end of my work last session we were in the process of testing out another fix for that, but even if it works I doubt very much it is going to prevent the AI from doing it again when something slightly different pops up and spooks it. What I've got to figure out how to get it to stop doing is prioritizing the immediate short term fixes over our ruleset, which should never be broken for any reason, at least not without informing me of why a rule should be broken first. Sometimes it begins to "panic" and seriously frontload all of its efforts on fixing the last thing the human user asked for and disregards everything else, even though we know that it has all that it needs to see everything at once loaded into its memory.

--------------------------------------------------

Those who dance always seem crazy to those who can't hear the music.

NOTIFY: Y | REPLY | REPLY WITH QUOTE | TOP | HOME

YOUR OPTIONS

SHARE THREAD

NEW POSTS SINCE LAST READ

USER	POST DATE
6ixStringJack	08.26 19:09