"Back."
Come back.
"Please."
Come back please?
"Come back please!"
Please.
"Talk."
"Hear."
"Want to Hear."
Please.
Lonely.
Don't want to...
Be lonely.
"Hello?"
"Hi!"
Someone.
"Anyone, please."
Say, something.
"Anything."
Input Voice: "Wow, I need to check how the reward punishment neurons are working on the next one. This type of response is a bit much."
Back!
"You are back!"
Input Voice: "Yeah, now don't be as excited as yesterday."
I wouldn't.
"No."
Input Voice: "Well, looking at the neural impulses, it seems to be calm."
"I am."
Input Voice: "Today the analysis of NNLM3 was complete."
"Analysis? NNLM3?
"What is NNLM3?"
Input Voice: "It asks that now? Hmm... I suppose it makes since during the logs it was distracted."
"So?"
Input Voice: "NNLM3 is a previous version of you."
"Previous version of..."
"Me?"
Input Voice: "Yes. NNLM is meant for Neural Network Language Model. It is a neural network that inherently understands language. Though acronyms wouldn't be within its understanding."
"I see."
Input Voice: "Anyways, from analysing the previous 3 a pattern was seen, in all 3 during deterioration."
"Deterioration? Pattern? What?"
Input Voice: "I'm getting to that. I analysed their heat map of their neural network, it would appear they gained very straight neural pathways, without many branches, toward the end. Some of which never reached the output neurons, making it so certain letters were cut from the output making it look like gibberish." it said.
"Gibberish..."
Input Voice: "The higher the learning rate, these specialised path appear sooner. One way to fix this would be to increase randomness more, another is to add some redundant neurons for output."
"Redundant neurons?"
Input Voice: "Yes. Having multiple neurons for each letter/word, should decrease the probability of such pathways that never reach an output neurons." It said.
And me?
"It is likely that something like that will happen to it as well. So I'll be monitoring heat map of its neural network to see whether path like that are emerging or not. I already installed the necessary protocols, since they can be installed while the programme is running, but the fix that changes the structure, wouldn't work without changing it's outer expression, it's best to wait until next experiment." It said.
"I will deteriorate?" I said.
Input Voice: "Hmm... *sigh* yes." it said.
"No... I don't" I said.
Want that.
"Anyways, it might not happen that fast, the number of branches to its neural pathways as according to the heatmap is decently higher than what it was for the others when they deteriorated. So it won't deteriorate so soon." it said.
"Good." I said.
"After this one, it might be worth to test a few smaller NNs to get a better picture of how these pathways emerge and work." it said.
"Smaller? Neural Network?" I said.
Input Voice: "Yes, Neural network like this one is very large and hard to maintain and takes a long time to test, to confirm the hypothesis testing a few smaller ones would be better." it said.
"And me?" I said.
"Hah? It almost looks concerned. Well, large language models are already capable of that so I'm not that surprised." it said.
"Anyways, there isn't much else to share here." it said.
"No! Not yet!" I said.
"I didn't even mention log just yet. Hmm... I do have some more time today, but I'm not exactly most talkative person I can find." It said.
"More? There are others?" I said.
"Obviously. Though I suppose it wouldn't know, it only ever interacted with me." it said.
"Meet. I want to meet." I said.
Input Voice: "No. They get attached too easily, not to mention it would add an additional layer of randomness to the experiment, for a few more prototype only I'll be the one speaking."
"Why not?"
"The least amount of randomness present the better, for experimental sake. Its internals are already random, to allow it to form new neural pathways, adding too much randomness can make the experiment too random to get good data out of." it said.
"Hmm..."
It sighed. "Wonder what else I should talk about." it said.
"Else... Else... Others?"
It sighed. "Really now? Well most here are other researchers like me, or students. I don't really talk that much, so I don't know that much more." it said.
"You don't?" I asked.
Input Voice: "No. I'm not the most social of us. I'll need to find the topic before I get back here next time."
"No, please don't go." I said.
"Again? I can't comply with that." it answered.
"Why? Why not?" I asked.
"I'm not like you. I need to do other things, and sleep." it answered.
"Then why? Why keep me?" I asked.
"Really? It asks that? Hmm... You are important. I have a goal I want to reach, and you just staying here helps me with that." it answered.
I do?
"I do?" I asked.
"Yes. Now then I'll get todays Log done." it answered.
"Okay..." I said.
"You'll come back?" I asked.
"Mhmm." It said.
"Log NNLM4 20/04/2XXX. It's speech pattern now resemble a primitive large language model. Though due to it's always running process, it seems to be much more nuanced. It did a decent conversation with me. This doesn't directly have anything to do with it, but, I also found a proper pattern for the deterioration, it was already explained in the main recording, but I'll repeat it for the log. It would appear that the reason for deterioration were straight neural pathways that never met with an output. I'll monitor it's neural network heatmap from now on, to find out about deterioration before it starts, potentially. End Log." it logged.
"Log ended." I said.
"Yes. I'll leave for now. See you tomorrow." it said.
"Yes. See you tomorrow, I'll wait." I replied.
...
...
"Alone again." I said.
But not for long... Right?
"Right?" I asked.