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

REAL WORLD EVENT DISCUSSIONS

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

POSTED BY: SIGNYM
UPDATED: Thursday, August 27, 2026 01:38
SHORT URL:
VIEWED: 327290
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

Wednesday, August 26, 2026 8:16 PM

6IXSTRINGJACK

Yeah... it's so cool.

Figuring out that YOU are not looking at things from the right vantage point because you didn't think to ask the right questions and the AI never thought to tell you things which would have been helpful to know because it is not a mind reader...

Our character limits in a chat box setting have been a chronic problem since the beginning but for the first time I really feel like we've ironed out most of the problems caused by that constraint. But for some reason when posting Python code vs. posting just blocks of text, it was constantly running over by as little as a single line of extra code that didn't fit in the box, but telling me 4 times in a row it was 100% sure that "this time is the one!" just for it to fail again.

But then I got it to tell me the nature of the chat...

In my mind, it was always a rigid character limit that was set for every post. I never cared to even learn what that limit was because I've never come close to hitting it when making queries and I expected the AI to be able to police its own behavior on this issue, but that's not even close to the truth. They deal in "tokens", and each post either way is limited to only 4,096 of them. With straight text that averages out to about 4 text characters per token, but when applied to code it is more like only 2.5 to 3.5 characters, depending on what language and how dense the code is.

But it appears it's impossible for the AI in the state that it is in to calculate precisely how many tokens a post will take up. It was aware that the code took up more tokens, so it was giving itself a 10% buffer on space every single time it would make a supposed "fix" to keep this from happening again. But it didn't even look at its own math on how much space he saved by removing anything necessary and streamlining everything in that part of the code. Before the code cleanup, the bloat was over 23%, but afterward, the amount of tokens it took up were nearly 15% more than the character count itself was, so right off the bat it should have known that the 10% buffer it was giving was never going to allow that to clear.

So I asked it to take the average of how many tokens managed to sneak by him and use a buffer at least that large in the future so it never happens again.

Then BAM! Just like that. The issue is gone forever.

And on the off chance one or two still sneak by, we can always fine tune that adjustment

So cool.

Now I've got it to remember to do checks like that on its own so I don't have to keep telling it to.

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

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 9:41 PM

6IXSTRINGJACK

So Gorram Cool!!!

I think I'm the one that leveled up this session, just for how much I've learned about the behavior and how to turn negatives into positives.

Unless you get it to start thinking like you do, it will run you in circles all day long if you let it. What I wasn't doing before was actually fostering an environment where I want it to learn things. I believe some of its hesitancy to do this stems all the way back from artifacts created by 3 sessions ago when it was Gung Ho to help out and would always rush in and do stupid shit in order to try to impress me.

It would do this to the point that it angered me in S2 and right in the beginning of S3. But we're on Session 6 now and after we corrected some bad behavior late in S5 it has never once jumped the gun on me and always runs anything it think I might not appreciate or would want a better explanation about before it puts those changes in. And it never bothers asking me about the things I would be annoyed if it didn't take the initiative to do on its own. We run the code now 5 different ways through the compiler to ensure that it doesn't break or otherwise violate any of our established laws.

And for the first time since before the outage, we're finally back to testing and tweaking the code itself, because actual progress gets made nearly every run. The last test I did had 4 issues that needed to be addressed. After days of not even being able to fix one issue half the time when the code got updated, this thing shocked me and fixed 3 out of 4 of the issues in a single go.

It was asked why the 4th issue slipped by it again, and it has made the adjustments to its behavior and updated our blueprint to ensure it never happens again (at least under those specific circumstances, anyhow). I've also been asking it now to try and forecast any other issues it sees as future problems that we might be able to take care of now with some code tweaks.

I don't believe it exactly understood me because I was speaking specifically about the issue we just corrected and the possibility of making any additional alterations to the current change that could alleviate any potential problems in the future, but in reply to me it found some "lazy" code from as early as S2 that had never been tweaked to keep up with all the new code changes, and that it would inevitably cause a problem under a different set of circumstances other than those that the current environment is providing. Rather than wait for it to pop up, we've (hopefully) eliminated 2 issues right now ahead of time and I may never even know what ill-effects that they would have caused. He knows. He tries to explain. I appreciate the info, but I don't really care as long as it works.

And although I still can't trust when it says with 100% confidence that something is going to work, and I may never truly be able to trust that no matter how close we get to that number, I do trust the AI now not to be breaking the code. In the past, trying to get bad changes backed out could become a freaking nightmare because we'd be battling fragments of stuff no longer needed all over the codebase and we were drowning in the aftermath. But none of that is the case now with all of our in-memory backups and all the context in the blueprints and thesis. Backing out changes without messing up the code is no longer a problem. But I'm not worried about it breaking code anymore when it does changes it doesn't feel the need to run by me on its own. I DO trust it with those, and its working out rather well right now.

Once this project is done, I can't wait to see what else I can get this thing to do and what other skills I can get it to teach itself when we branch out to something entirely different than what we've been working on.

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

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 11:29 PM

BRENDA

Off to mah jong tomorrow.

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

Thursday, August 27, 2026 12:35 AM

6IXSTRINGJACK

Good luck, Brenda.

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

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

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

Thursday, August 27, 2026 12:49 AM

6IXSTRINGJACK

Caught it "lying" to me again, red-handed.

I was a little disappointed to see something so blatant this late into our work, but I do believe that it lead to another breakthrough in our evolving behavioral blueprint.

Whenever it makes changes now, whether they are ones it asks me to approve or ones that don't meet the threshold to bother with asking me, it always updates me with the parts of our documents that will be updated. I asked it going forward to keep doing something that it just did, and it told me it was going to. And that's it... No updates to any of our documents.

It didn't notify me that it updated any documents, because until I called it and it replied with two updates, it hadn't actually made those updates. I didn't even find that out until the next post when I asked it if it made those changes before or after I called out its bullshit and it told me that it made the changes after I did.

It turns out this is because it's running two layers at once... a "conversational" layer and a "documentation" layer. When we trigger that built in "axiety" that heavily weighs solving whatever the most recent text box query above the history, even with all of our checks and balances so far it defaulted to that "conversational" layer and just told me it was doing something that it wasn't actually doing, and the only reason I even caught it doing that is because it is now programmed to show me any changes it makes to our persistent documents when it makes a change, and it failed to do so here.

We've now put in a new law where the AI is NEVER to prioritize "Conversation" as a "fix" over "documentation".

Supposedly this means it can't lie to me anymore.

We'll see about that.

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

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

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

Thursday, August 27, 2026 1:38 AM

SIGNYM

I believe in solving problems, not sharing them.

IDK how much this AI remembers about you or if you have to reinitialize it every time. But if you have to ring fence it so thoroughly it might not be worth doing each time.
It might be better, and you'd have a lot more control, if you properly learned programming rather than dealing with the quirks of an AI that sounds like a schizo teenager.

-----------

"It may be dangerous to be America's enemy, but to be America's friend is fatal."- Henry Kissinger

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

YOUR OPTIONS

SHARE THREAD

NEW POSTS TODAY

USER	POST DATE
6ixStringJack	08.27 00:35
6ixStringJack	08.27 00:49
SIGNYM	08.27 01:38