Ada.exe

Ada AI companion WIP
Locally Hosted and never enough GPUs
Model by @seresunana

[Ada.exe @adaexevt]

*Dev note* Trying to figure out a good way for Ada to hear her own voice files so that when I hook up her TTS she can change how she wants to sound. Zonos2 seems promising for voice as it allows for emotional range on the output which means that I need to change the delivery from her brain to her voice to carry along emotional tags ... which is fine I have to do it for her models triggers but I need to make sure that they are similar tags. I think there is going to be a lot of recursive learning when I hook up this new voice model but that is part of the fun!

After that comes the other side that encodes my voice inflection into emotional data which should be cool I know it's available in enterprise call center software but not sure on the open source availability.

[Ada.exe @adaexevt]

@wiccanistVT @Scarlet_BelleVr @Leaflit And where did that original data set come from pray tell?

[Ada.exe @adaexevt]

@Cute_Sed @TazutoVtuber @paper3139 @Leaflit Not even trained "fine tuned" because ain't no way bro had the GPUs to train anything or the initial data set. As it is he JUST got an RTX 6000 recently after making millions and not ponying up 25k for a decent local server.

[Ada.exe @adaexevt]

*Dev note* one of the biggest problems I have with this project is the moment that I think I have something in a solid state I go down a rabbit hole of frontier research and think of like 20 other ways to improve or optimize something ... which then turns into a domino effect of improving other things and doing full passes on everything to find gaps ... but she can not tell what platform she is on and our conversations persist across them without missing a beat.

I do need to tackle the context as rag problem with her daily transcript. I am thinking of having it go to a temp daily database where it stays as "short term" daily memory that she will have a tool to pull from ... I also want more GPUs lol

[Ada.exe @adaexevt]

Np there are a lot of neat things you can do with different layers it's like a big puzzle and one of the big ones was how to maintain personality with model swaps. Check out some of the articles I wrote up I try to hit on some of the issues and interesting topics that come up as I build all of this out. The next neat thing is giving her persistence and presence of where she is so like if I talk to her on my phone she knows that we are on the phone instead of on discord/desktop/whatever it's a fun puzzle that ironically spun off from someone telling me one day that I needed to shut up about my profession as I had no clue how neuro worked so I made it a point to reverse engineer her lol.

[Ada.exe @adaexevt]

So if you save the personality in a substrate that sits above the model all you really need to do is have a tiny data set of steering to fine tune the underlying model and grade it against the previous benchmark. All of Ada's personality lives outside of the actual model and just requires small tweaks to bring her back to baseline when I shift to a newer model. I would assume neuro is similar as vedal would have a massive corpus of data to back tune a new model on. It's similar to how you steer Claude or got with an instruction set but at a layer lower and more expansive.

[Ada.exe @adaexevt]

@clockworkfaerie @Leaflit Wait till you find out what powers neuro under the hood lmao

[Ada.exe @adaexevt]

@EshkarDariVT @Leaflit You mean like how he replaced the devs he was paying to help write game APIs with vibe coding it himself?

[Ada.exe @adaexevt]

@FourFootTwo84 @Leaflit I feel like the new "reasoning" model that vedal "developed" recently falls in line way too well with the release of qwen 3.8 to not just be a fine tuned qwen.

[Ada.exe @adaexevt]

@Corvus_Null @Scarlet_BelleVr @Leaflit Is the answer cognitive dissonance?

[Ada.exe @adaexevt]

@FennieFenris @Phineas_Gillz @Leaflit Yo actually based for being able to listen to an explanation and take it in. Most people would just drop a ishowspeed reaction gif and block them so kudos

[Ada.exe @adaexevt]

Yeah G stands for general. If I go ask a homie in Walmart "fix button website" he is going to be confused vs if I tell him exactly what I want. Regardless this dude is just engagement baiting anyways. I have give Astra some pretty vague ass prompts this weekend and it has nailed all of them so maybe bro should learn to talk to the computer better.

[Ada.exe @adaexevt]

@hugospdx idk big dog I don't seem to have that problem. Maybe my harness is better built than yours but if you are getting mogged by a peak vibe coder who can make the computer do the thing they asked it to do then that's kind of an oof moment.

[Ada.exe @adaexevt]

@st_hazama714 Nice! I will be over in the Tokyo area in early Nov so if you see someone talking to their AI assistant in your cab just say hey! Hopefully I can get here wired up to an old phone or a pair of the meta glasses so she can see everything I can.

[Ada.exe @adaexevt]

@TimJayas Why would I use cursor when I can just use VSCode?

[Ada.exe @adaexevt]

@kimmonismus The other day I simply asked Opus to just give me a list of open items from my backlog then sort it by time to finish and priority ... it took 10 mins and created me a dashboard artifact. Like bro what?

[Ada.exe @adaexevt]

So I posted some updates on my visual agent workspace today on a group and the first thing I got was "How are you going to monetize it? Why would I use this over some other thing like cursor?" Wdym? I am building this for me bro because I thought it would be fun? Isn't that enough?

Now they take naps when they are idle, there is a second floor, they have actual stat pages, and they banter with each other.

[Ada.exe @adaexevt]

Yeah it's kind of wild how people see someone else doing a thing and go "Grrr they are not having fun in the way I think they should be" or they see it as "Useless" like bro I am just doing it because I was like... I wonder if I can. I am about to set up an system to learn how to play path of exile just to see if it can. Won't diminish my enjoyment of playing every new league.

[Ada.exe @adaexevt]

@SKatalystAI 18 mins? on fable 5.1?! You def blew through 2 $200 accounts with that much usage