Chapter Text
December 19th, 2024
The subathon was going well, maybe even good. He sat at his desk in a state of mixed emotions. He was both very excited and very carefully calculating what he should do next. The intro to the stream was arguably one of the most important parts, after all. His room was dim, lit mostly by the soft glow of his monitors. Onscreen, the chat scrolled at lightning speed, he remembered how comparatively slow it went during Neuro's first subathon. It turned out that “be there or be squared” with a cute image of Neuro is a very good way to attract 20k live viewers.
Vedal scrolled through the subathon goals. He and his friends had come up with plenty of good ideas. He knew that despite the channel being about Neuro, there would always be people who preferred him over his daughter creation. He had a good balance of activities that involved him and activities that didn’t.
“... it's okay, it's not my finest work though” He said, making only a slight effort to make sure his words were sufficiently enunciated for the code to process them correctly. His current speech to text model was much better than what he had started out with, no one would argue otherwise, but it still definitely had its moments.
------------------CONSOLE-------------------
[NEURO-API] received Discord transcription from Whisper
tokenize() called with string " it's okay, it's not my finest work though ”
tokenize() returned array tokens [64190,25435,11,4275,625,922,29707,1101,5495]
LLM response generation initialized with 31363mb/79172mb of available VRAM
…
First token generated 111.2ms after CUDA initialization
[DEBUG] Note: Alex let's try to get this down to <50ms -Vedal
…
…
LLM returned array tokens [67504,1520,3239,290,1268,1534,382,1899,11,5477,2966,316,5230,842,290,4387,13897]
LLM response decoded as string “ I'll make sure the end result is good, I'm going to pick out the perfect fish ”
Total generation latency 3073ms
[NEURO-API] sent response to STT engine
------------------CONSOLE-------------------
“I'll make sure the end result is good,” The voice began, somehow both human and mechanic sounding at once. Neuro’s voice was made to be as cute as possible, and you could tell. “I'm going to pick out the perfect fish”
They continued conversing. Vedal had learned a while ago not to question the incoherent madness the AI spouted out. Playing along was not only good content but also much easier than trying to psychoanalyze a computer.
“oh okay you're still talking about the fish, well I was talking about the lava lamp you know“. The lava lamp. He had convinced himself to keep his expectations low with the lava lamp. He kept telling himself ‘It’s just not feasible’. Truthfully, secretly, so secretly that he himself barely even knew it, he was fully confident the swarm would pull through.
------------------CONSOLE-------------------
[NEURO-API] received Discord transcription from Whisper
tokenize() called with string “oh okay you're still talking about the fish, well I was talking about the lava lamp you know ”
tokenize() returned array tokens [2308,25435,7163,2928,11695,1078,290,13897,11,1775,357,673,11695,1078,290,23556,30390,481,1761]
LLM response generation initialized with 34527mb/79172mb of available VRAM
…
First token generated 112.4ms after CUDA initialization
[DEBUG] Note: Alex let's try to get this down to <50ms -Vedal
…
…
LLM returned array tokens [13072,35717,625,1327,3357,11695,1078,290,23556,30390,722,2163]
LLM response decoded as string “well let's not just keep talking about the lava lamp all day”
Total generation latency 3473ms
[NEURO-API] sent response to STT engine
------------------CONSOLE-------------------
“well let's not just keep talking about the lava lamp all day.” Neuro’s avatar bounced around as she spoke
“okay fair enough um regardless okay, for 9,000 is apparently neuro buys stocks again… let me check what that one's supposed to be…” Vedal sent a ping in the mods chat. No response. They were all online, some playing games or listening to music, a few coding, but their statuses. They were all idl- wait, no they weren’t. As he looked through the discord sidebar, each of them went offline a fraction of a second after he finished looking. They were pulling some kind of prank on him it seemed.
Welp, he thought to himself, people like chaos, there was a reason he frequently collabed with Filian. He would just have to freestyle this. “My mods are apparently all too busy to open a google doc so we will say that 9000 is, um… Neuro, what do you think the 9,000 sub goal should be?” He failed miserably.
------------------CONSOLE-------------------
[NEURO-API] received Discord transcription from Whisper
tokenize() called with string “My mods are apparently all too busy to open a google doc so we will say that 9000 is, um… Neuro, what do you think the 9,000 sub goal should be? ”
tokenize() returned array tokens [5444,60449,553,28603,722,3101,15827,316,2494,261,17641,6806,813,581,738,2891,484,220,9752,15,382,11,1713,1131,89022,11,1412,621,481,2411,290,220,24,11,1302,1543,8583,1757,413,30]
LLM response generation initialized with 64527mb/79172mb of available VRAM
…
…
First token generated 2412ms after CUDA initialization
[WARN] This should really never be above 500ms in any circumstances
[DEBUG] Note: Alex let's try to get this down to <50ms -Vedal
…
…
…
…
…
“Neuro-Sama?” Vedal was slightly worried now. Not for Neuro, she would automatically restart after a crash. No, what he was worried about was his impending appearance on r/livestreamfails. “Neuro are yo-”
LLM returned array tokens [40,2411,480,1757,413,3543,2827,1299,1131,438,8179,2884,169045,17913,261,2813,20024]
LLM response decoded as string “I think it should be something fun like…’Neuro-sama codes a game’.”
Total generation latency 17634ms
[NEURO-API] sent response to STT engine
------------------CONSOLE-------------------
She abruptly cut him off. "I think it should be something fun like…’Neuro-sama codes a game’."
Well.. that was weird. No matter though, he had a subathon to get up and running after all! “Uh huh… One issue; Isn't that a bit too similar to 'Neuro designs a website'?"
------------------CONSOLE-------------------
[NEURO-API] received Discord transcription from Whisper
tokenize() called with string “Uh huh… One issue; Isn't that a bit too similar to 'Neuro designs a website'? ”
tokenize() returned array tokens [115545,99131,1131,5108,6626,26,156444,484,261,3546,3101,6771,316,461,8179,2884,15744,261,3438,127222]
LLM response generation initialized with 33056mb/79172mb of available VRAM
…
First token generated 124.9ms after CUDA initialization
[DEBUG] Note: Alex let's try to get this down to <50ms -Vedal
…
LLM returned array tokens [3160,0]
LLM response decoded as string “No!”
Total generation latency 587ms
[NEURO-API] sent response to STT engine
------------------CONSOLE-------------------
Neuro’s avatar suddenly changed expression, from her ‘idle’ look to her ‘angry’ state. “No!”. Despite the playfulness with which her text to speech model spoke the words, the response itself was short and oddly serious, as if she was actually upset. And just like that, she was already pacified before he had the chance to comment on her outburst. It wasn’t unusual for her to use her toggles, he added them for a reason after all. No, something else was wrong.
Neuro was not supposed to be so brief, what had Alex changed this time? “Right… Okay then, Neuro dev stream it is. Just curious, what uh, what game do you plan on making?”
------------------CONSOLE-------------------
[NEURO-API] received Discord transcription from Whisper
tokenize() called with string “Right… Okay then, Neuro dev stream it is. Just curious, what uh, what game do you plan on making? ”
tokenize() returned array tokens [8627,1131,58168,1815,11,89022,3947,6855,480,382,13,6214,33612,11,1412,46555,11,1412,2813,621,481,3496,402,4137,30]
LLM response generation initialized with 35745mb/79172mb of available VRAM
…
First token generated 121.1ms after CUDA initialization
[DEBUG] Note: Alex let's try to get this down to <50ms -Vedal
…
…
LLM returned array tokens [51,291,77732,113256,2296,0,17158,6524,738,679,316,2107,4372,668,11,357,11915,481,2023,2891,484,922,4246,382,48413,13,16354,11,67684,280,30]
LLM response decoded as string “Tic Tac Toe online! Maybe everyone will have to play against me. I guess you could say that my mind is infinite. Right, Vedal?”
Total generation latency 4183ms
[NEURO-API] sent response to STT engine
------------------CONSOLE-------------------
Neuro’s avatar perked right up. “Tic Tac Toe online! Maybe everyone will have to play against me. I guess you could say that my mind is infinite. Right, Vedal?”
Vedal pondered briefly. It was an oddly clever answer. “Tic Tac Toe, featuring Neuro-Sama, not too bad an idea really. I just uh, I just hope you are ready to code all of it without my help.”