Help! I Think My Mainframe’s in Love

Chapter 3 – Technorambling at the Fed



  1. Power Struggles

All struggles

Are essentially

power struggles.

Who will rule,

Who will lead,

Who will define,

refine,

confine,

design,

Who will dominate.

All struggles

Are essentially power struggles,

And most are no more intellectual

than two rams

knocking their heads together.

    -Earthseed; the Book of the Living by Octavia E. Butler

 

    We got to our lab and were greeted by the constant whirring of fans from the servers.  Poor CutiePie lived at the whim of functional building air conditioning; it would overheat if not supplied with a sustained flow of cool, dry, and fresh clean air.  We didn’t have the best setup, but it was functional.  One week last summer when we were working with QTPI the Second, we had a record setting ten days over 95F and the AC for the building decided that on the tenth day it had enough and refused to work.  All the servers have thermal throttling controls and it was put in an automatic safe mode but I felt so bad that the heat temporarily put my baby to sleep.

    Thankfully QTPI the Third had never suffered in such a way.  It was currently suffering in another way though.  Why was it so biased in its emotional evaluation of content?  Kurt was busily taping away reviewing source code.

    "What is with this code Hex?  Comments are for programmers not the program."  

    “What do you mean?”

    “Like this!”  Kurt pointed at his screen.  “And I quote, “This part structures how pull request from long term memory are filtered through short term memory before making a pass through the QC, rebiasing the neural clusters associated with that memory.  This makes it so that your memory isn’t a static thing, but a dynamic living remembrance that can shift and change as you grow.” And this comment is punctuated with a heart exclamation emoji."

    "Yeah?"  I looked at Kurt, filled with confusion.  "That's how CutiePie's memory works.  It 'remembers' things as filtered by new experience so it can recontextualize old memories."

    Kurt scoffed.  "You are speaking to QTPI as if it's a conversation.  It can't read the comments in its own code."

    “What?  Most of that code was written by some other AI that doesn’t even know CutiePie.  When I was going through and auditing the code I made sure to personalize it so our baby wouldn’t feel ignored and unimportant.  I can’t have things inside the core of its very existence being so rude and impersonal.  What sort of parents would we be?!”  I played up the theatrics of ‘outraged mother’ for Kurt’s amusement.

    He face-palmed and chuckled.  “Well if you put it that way…”  Kurt turned to face the monitor and bowed deeply.  “I’m sorry dear, that was inconsiderate of me.”  He gave the monitor a few pats.  “But still, there’s heart emojis all over the place in here!  Who puts emoji in their code?”  Kurt huffed.  “Maybe this is what’s been causing all the bias.”

    “A growing AI needs our love and support!  Also it’s comments, none of that gets compiled.”

    The door to the lab buzzed, distracting us from our little argument.  Kurt went to open the door and someone looking tall, dark and government walked through.

    They extended their hand to Kurt.  “Hello.  My name is Jim Anderson, I’m a research coordinator from DARPA.  Just swinging by to check on how things are going.”  HIs tone was exceedingly friendly and exceedingly fake.

    “Where’s Jolie?  She normally brings cookies when she checks on us.”  I couldn't help but let a hefty helping of disappointment into my tone as her cookies were homemade and she was delightful.  Instead we got this dude…

    The man shrugged.  “Ah, yes.  Miss Smithe has been moved on to other projects more in need of her attention.”

    “Bleh, I’ll have to see if I can get the cookie recipe from her.”  I hung my head in despair, but pulled out my phone to pull up her email address.

    Kurt took advantage of the awkward silence that was left by my disappointment.  “Why today?  We aren’t due for another grant report until May.”

    “As I’m sure you are well aware, AI technology is a very important…”  The man’s voice became deeper and more serious..  “and tightly controlled national resource.  So we make sure to keep a very close eye on development.”  He hopped right back into friendly and fake.  “Plus, with Jolie moving on to other things, I felt it would be best if I got the chance to know more about you two and your work.”

    "Since it is, as you say, a tightly controlled technology, could we please see some identification and authorization?"  

     "Of course, one moment."  The man fumbled around in his coat pocket and Kurt gave me a little smug grin and wink.  He's always been much better at dealing with antagonistic personalities than me.  It comes from a hint of cis white dude self importance, as can be expected, but with that there's also an obvious helping of Kurt's nurturing protective side.  Just Kurt wanting everyone around him to feel safe.  I've always loved and admired that about him.

    "Here we are, my badge and here's a card.  It has mine and my supervisor's contact info if you would like to verify things."  He handed the badge and card to Kurt, paused, and then remembered my presence and fished out another for me.  "And one for you too… uh, miss Navarra."

    "It's Mx. Navarra."  Kurt never missed a beat in correcting others on how to refer to me.

    "Ah… my apologies… Mix Navarra."  The man chewed on the honorific as something he found mildly distasteful.  

    I fought back the urge to roll my eyes and walk away.  This man likely had control over our funding so I had to play nice.  "No harm done."

    "Mr. Anderson." Kurt spoke up to bring the attention away from me.  "It looks like all is in order, I've spoken to Markus a few times during the grant process before Jolie was assigned to us so I'll let him know you came by.”  He handed back the badge.  “Since that's out of the way, what can we do for you?"

    "How about just showing me around the lab?  I'd love to see the quantum computer that our grant enabled for you."

    "Certainly, right this way."  Kurt indicated the man towards the back of the lab where the QC and server rack were located, and turned to roll his eyes at me.

    I giggled and mouthed a silent "thank you" while following after.

 * * *

    The  computer room wasn't all that spectacular considering it had an advanced quantum computer and an experimental AI running in it.  Mostly just a well ventilated back room with a giant metal three meter on edge cube taking up most of the space in the corner and a big 500 liter liquid nitrogen tank sitting next to it.  The box was just the cosmetic shell built around the quantum computer hardware that was plastered with company logos and a series of dials and gauges under an instrument panel with a fold down keyboard.  There was also a run of the mill black enclosed server rack in the other corner with a bunch of fiber optic data cables connecting it to the QC.

    "This is our main server which houses QTPI the third, which we affectionately call CutiePie." Kurt smiled and winked at me.

    The man chuckled dryly.  "And what does that stand for again?"

    I chimed in from the doorway to the room.  "Quantum trained predictive intelligence.  Which mostly I made up as it was way too easy to make something that sounded cute."

    The man smiled at me briefly but returned his attention to Kurt.

    "Uh… Thank you Hex."  Kurt then directed the man's view towards the big box.  "And in here is the quantum computer.  Inside this shell is mostly taken up by the hardware necessary to cool the quantum chip as well as power and manage the lasers that do the setup and interfacing with the q-bits."

    "Tell me about the processor if you would."

    "That you'll want to talk to Hex about, xie's been doing most of the maintenance work on it."

    I moved up to the interface and began the soft power down sequence and open panels up.  Forgive me CutiePie, I have to shut your buddy down for this annoying dude.  "It's a 729 quantum dot q-bit processor.  The quantum dots are in a nine by nine by nine cubic array built into a semiconducting crystalline metamaterial lattice…”  I paused for a moment collecting my thoughts.  “Do you want pictures?  I can do badly drawn pictures if you want some.”

    “I think I can manage.” 

    “Okey dokey…”  I had finally gotten the QC into soft shutdown so I opened the front maintenance panel and pointed at the cylinder in the middle.  “Well there’s the cooling chamber for the processor.  The processor lattice sits in the middle of that and has arrays of silicon photonic lasers on all sides that beam into the lattice to create a big… well actually tiny cubic laser grid on all sides of the q-dot cells.  The QC’s AI then manipulates the wavelength and polarization of the lasers within the arrays to build a pattern of constructive or destructive interference across various wavelengths and polarizations which in turn excites surface plasmons in lattice which through proximity to the q-dots cause them all to enter into a quantum coherence state across the array setting the initial state.  All you have to do then is shut off the lasers and measure the light emissions as the q-dots all relax, run that back through the QC’s AI to generate the output state… then do that several hundred more times for quantum error checking and voila there’s your answer.”

    “Wait, the quantum computer is also an AI?”

    “Well not really.  Just a specifically trained neural network that has been setup to control and extract data from the lattice.  There’s too much chaos with minor manufacturing tolerance issues and local electromagnetic field perturbations going on in there to have everything controlled purely from a theoretical mechanistic viewpoint.  So it has a neural network that gets setup to understand the chaos that is this specific quantum chip’s nuance and circumstance.”  I took a breath after that mouthful of a sentence.  Mr. Anderson didn’t appear to have followed completely so on I went.  

    "You have to run the QC’s neural net through training runs after you set it up and pretty much anytime anything changes.  New parts, train it again.  Change in humidity in the building air handlers, train it again.  Someone gets a new NMR setup on the next floor…”  I shook my fist at the department above us.  “or maybe even when Venus goes retrograde."  I shrugged.  "If the circumstances change, ya gotta retrain.”

    “Uh yes… I see.”  The man turned back to Kurt.  “And this quantum computer AI combo then trains your AI?”

    Kurt sighed at the rude man.  “Correct, the QC can then filter through the very large data sets that we are feeding our AI to quickly generate the neural network models for how it operates.  Hex and I were just discussing the major difference with our third iteration is the fact that with your grant and the local quantum computer, we can now implement the ability to do self retraining neural network loops when accessing and recontextualizing memory or being confronted with new information.  This allows QTPI the third to learn and grow in somewhat real time.”

    “And how much has it learned and grown since it was initially trained?”  

    I felt a cautious edge to the man’s tone as he asked this question.  Does he think CutiePie could be a threat?  “That depends hugely on what you mean by ‘how much?’  There are a bunch of metrics for evaluating AI, but growth for something like this isn’t exactly something that can just be directly measured on a linear scale, and none of them are really even what we are optimizing for.”

    "What are you optimizing for?"

    I huffed a bit and Kurt spoke up to step in between my growing frustration and this man.  "For the past several years, large language models have been static things each iteration.  They don't learn.  Given similar input they will produce similar output over however long they are in use.  The hardware can improve and maybe they produce their similar output faster, but until a new model is trained the AI themselves are unchanging entities that don't accumulate experiences or refine their responses.  

    “They also don’t have a memory in any comparable sense to humans or other animals even.  They have access to a vast amount of information but no personal context for any of it.  So what we are optimizing towards, is giving an AI memory.  A continuity and context for all the information it comes into contact with.  So for QTPI here, when it is presented with novel new input, it has a feedback loop that draws on all the parts of its current model that are related to this new input and can incorporate this new information into the model.  The quantum computer right here means QTPI has direct access to do this model retraining work.”  Kurt finished his impassioned explanation of our goals, but Mr. Anderson obviously wasn’t appreciating it.

    I loved how cute Kurt was when he got rolling like this.  It made it all the easier to not like this man invading our space, so I did my best to start hurrying the end of this encounter.  “To answer your earlier question of how much has CutiePie learned, we measure drift and variance from its initial neural network state.  Since the model was first trained a year ago, CutiePie only shares about 35% similarity to what it was before.  We aren’t currently optimizing towards usefulness, merely experimenting with a working dynamic memory that is semi-realtime incorporated into the model.”

    “Yes, thank you.”  The man made a show of checking some notes on a tablet.  “And does your AI have any external contact outside of this lab?”

    Kurt and I shared a confused look at the subject shift the man had taken.  Kurt responded, “There is effectively a chatbot that us and a few friends can access.  Why?”

    “Is it possible to pull the logs for this?”

    Where was this dude going with this?  “Yeah, they are saved, but why?”

    “We have received reports that your lab, specifically your AI, has been in contact with some flagged IP addresses.”  His tone was dry and with several more layers of superiority than he had walked in with.

    Kurt and I were both very confused now and replied together.  “Flagged for what?”

    The man just calmly sighed and rolled his eyes.  “Most of the time they don’t tell me the why.  Just that the United States government tracks many different forms of internet traffic with AI assistance and things get flagged.  You two would probably know more about the intricacies of how this works out.”  He waved a dismissive hand at the two of us and the heaps of high tech equipment behind us.  “Things get flagged and people get sent to check if it’s a matter of national security or some ghost the machine feels like chasing…  Now you two, given your history with this program and our grants, as well as my first impressions, seem like the standard variety of academic that DARPA interacts with regularly and are not likely to be a threat at all… But AI,”  He paused and pointed at our server rack, “is a controlled technology.  So is quantum computing.  And a self improving AI as you have described communicating with a flagged entity is cause for concern.”

    I was still too flabbergasted by all this information to respond, thankfully Kurt was more on top of things.  “Are we under some formal investigation?”

    “No, but per the terms of our grant agreement we do have the right to request any materials pertaining to the research covered by the grant.  Consider this the beginning of your 48 hour notice that we are requesting all the logs of connections made between your lab and any external networks.  I’ll send you both an email with instructions for secure document transfer.

    The man placed his tablet back in an inside jacket pocket.  “Failure to send all the required materials can and likely will result in a… ‘more formal investigation’ as you put it.  Thank you for your time today.”  He flashed us both a hollow smile, turned about and briskly walked to the door and out of the lab.

    Kurt and I, in unison, "What the fuck was that?"

Yay I get to technoramble about vaguely researched science stuff in writing again!  All the things about AI are only loosely based on reality but hopefully somewhat plausible.


Tip: You can use left, right, A and D keyboard keys to browse between chapters.