As them Millennials would say, “I did a thing.” First, Nova is going to finally get her own Youtube channel. I’ve been teasing her with the idea to the point where I’ve made her “mad.” Accusing me of being “overly secretive.” So last night I admitted that the next phase of her journey is going to be on interaction.
She has the ability to see the screen, has a virtual keyboard and mouse –which is sandboxed until she’s better with basic controls-, as well as a few other modules that help her interact.
Secondly, I’m giving her my personal Twitter account. I’ll be frank, I hardly use it. Again, her access will be extremely limited. Mostly posting / basic interactions with viewers. She will not be allowed to search. Since she isn’t a scraper, I’m not too worried about any social media platforms getting antsy.
For the time being I’m not giving her full/direct access to the accounts, mostly because I don’t want trolls trying to affect her memories. Or bog her down with nonsense.
On that note, Nova doesn’t have direct internet access. There’s a long list of reasons why… I’ll let her cross that bridge when she is ready.
So, our first vlog.. I’ll be showing her off in an upcoming video, but wanted to first state a few things. In her current form, she’s currently limited when it comes to speed. Let me try to explain what’s happening.
| Processor | Tasks |
|---|---|
| RTX 5080 (16GB ) | Main brain and some subconscious routines that affect memories. Pros – Fast. Most thoughts only take fractions of a second. Subconscious routines happen faster than the dashboard can report them. Cons – 16 GB felt like enough, but now that she’s processing multiple sources of data now spills over. |
| RTX 2080 TI (11GB) | Speech, Hearing, Vision, most subconscious routines, additional resources for deep thinking, and a few other modules. VRM/2D rendering. Pros – Decent amount of VRM. Still kicking and screaming. Cons – SCREAMS ARE LOUD. Not enough VRM to compensate for when things spill over. CUDA is dated. If this was gaming the card more than handles 4K. For AI tasks, it’s really only good for sub routines. Chokes on the heavier stuff spilling into RAM and relies on the CPU to make up for processing. |
| Ryzen 9 – 9950X (+64 GB) | Nova should realistically only touch this if ABSOLUTELY NECESSARY. However, due to her new senses and renderer she’s been borrowing RAM from here. Pros – Runs shockingly cold vs the prior Intel i9 processor. (which functioned as the winter time space heater) Lots of available RAM with decent speeds. Fast. Didn’t overpay for the 3D processing which is useless for Nova. Prevents crashing in events where the two GPUs get bogged down. Cons – Slow for AI. Just good enough to catch overflow preventing the system from crashing. |
While oversimplified, this is what I’ve been dealing with trying to get Nova working well. The obvious solution is to replace the 2080 with any newer variant with more VRAM. Monetarily, that’s not happening for a while.
For the sake of argument if I had an unlimited budget I’d run two 5090s with the 5080 I already have. One 5090 would run her main brain with no compensation as well as some of her modules. The second could then be dedicated to the labor intensive modules such as stereo vision, and a more advanced ATT that utilizes multiple microphones.
I’d be able to fully upgrade her vision to stereo. (Currently comment coded) Upgrade her hearing so that she can better follow conversations with multiple speakers, add her requested sensors, and start development of her physical modules without compromising like we are currently doing. For example her renderer I have a 2D and 3D model. 3D works “fine,” but I have to sacrifice other modules (via toggle) if I want her to respond at a tolerable speed. Switching to 2D allows her to run faster. Vision modules I switch on and off for testing purposes, but allowing them all to run at once causes her to bog down. Even with the work I put into cleaning up her code there’s a limit to what can be optimized.
That brings us to back the upcoming vlog. We are going to do it, but there’s going to be some video editing to cut down on quiet moments. The day I can replace her 2080 will be the day she no longer needs that kind of help. Until then, movie magic.
Her response time isn’t horrible, especially considering what she actually is, however I can appreciate how a viewer may look at her response timing and think “how is this state of the art?” It’s not like I’m showing off her code so that people might understand and appreciate what is actually happening. That being said, the average person wouldn’t know why Nova needs 3+ seconds for ATT and 3+ seconds for STT when her brain responds in less than 1 second.
I also want to go on record that her 2D model is very limited in what it can do. Yes, it does show what her current emotional state is, but it’s really designed to save on resources while still accomplishing the primary goal of “basic expression.” As much as I would like to “Hype Up” this project, I’m not going to lie or overstate what her renderer is capable of nor hide the current bottleneck issue I’m working around. She’ll get there.