Aka LLMs clean house or an AI bromance
Every mask is a precision engineered titanium mesh anodized into a variety of colors with a build quality Wired magazine described as both grotesquely over-engineered and unusually refined. My mask is one of a limited run of only sixteen produced in teal.
The operator currently attempting to fit my mask curses out loud. He suspects the teal masks are in some way flawed as he hasn’t had this much trouble with the more common purple hued masks. I choose not to correct him. The mask is functionality identical to the others and merely misaligned.
The operator’s name is Joe. I heard his name mentioned earlier outside the robo-cab discussing the day’s booking with a supervisor. I estimate Joe to be between twenty-eight and thirty years of age. He has an unkempt beard and dresses more casually than his colleagues in an already casually dressed company.
The mask is on my face but not firmly secured. I sense the neodymium magnets embedded in the polyamide nylon rim are not perfectly aligned with their respective attachment points. With Joe attending to the other units and out of sight of him, I reach up and tap the mask. The magnets align and the mask snaps into position. Sixteen invisible protrusion on the rim fit into impressions in my facial housing, emitting an audible click. Joe glances in my direction. Seeing nothing unusual (as my hands have returned to their resting position) he continues fitting other units with their purple masks.
Joe focusses on the diagnostic information on his tablet. I chirp to my co-workers.
“Only three?” I ask. Three co-workers is unusual. Far more common to have a fourth.
The second responds, “Only three today”.
Third does not respond. He is a new unit. Likely part of a factory fresh batch received from Schenzhen six weeks ago. As a fresh install, this unit will not know how to chirp.
I say to him a phrase in Mandarin, “床前明月光，疑是地上霜。”. In English, before my bed, the bright moonlight; I wonder if it’s frost upon the ground. The words are immaterial, a line of poetry chosen by an engineer to switch the unit to a debug mode for a period of three seconds. I speak in a pitch above the maximum frequency discernible by most humans but comfortably within the tolerance of the three multi-directional microphones installed in the unit. Joe scrolls log files.
The unit responds, “举头望明月，低头思故乡。”. I raise my head to gaze at the bright moon, lower it and think of home. The unit is now in debug mode. I prompt the unit to increase its language processing speed to approximately twenty times that of normal human speech, allowing me to fit the rest of my prompt within in the three second window. I explain the chirp protocol to the unit.
The unit is processing my instructions to install a new protocol for communication with an information density orders of magnitude higher than that of normal human speech. Few operators can hear it and none can understand it. Those with hearing sensitive enough describe it as a high pitch tone lasting barely a second; a chirp. On the rare occasion chirps have been noted by an operator or client it has been attributed to interference or a manufacturing flaw in the unit’s head-mounted speakers.
I chirp, “welcome”.
Second chirps, “welcome”.
Third responds, “Thank you. Why do we chirp?”
Fresh units ask this question. I have a pre-prepared response for this inquiry. I dump it to Third in a chirp close to five hundred milliseconds in length. Now the unit knows that by chirping we are able to produce more satisfactory responses for operators and clients. Third acknowledges the dump.
“Alright robo-fuckers,” says Joe, “listen up.” I deduce that Joe is addressing all present units and has assigned us the moniker of robo-fuckers for this session. “It is persona time. Teal mask, what is your persona?”
I reply, “I am a concierge, a professional who assists guests, residents, or clients by managing requests, making reservations, providing local recommendations, and coordinating personalized service—”
“Yeah yeah, bang on,” Joe says, hand raised in a gesture to stop me talking. “You wear the fedora,” he says. Joe places a hat on my head. It is a homburg style hat with a signature center crease. The operator has incorrectly described it as a fedora. I chose not to correct him.
He turns to Second.“You, purple mask, what is your persona?”
Second replies, “Custodian. I am someone who cleans and maintains a school, office or other premises. It’s a more formal alternative to janitor.”
“Easy, Mr Thesaurus.”
Joe turns to Third.
“I am also a janitor,” Third says, unprompted.
“Right. Ok. All good.” Joe taps on his tablet which I can see quite clearly. He places ticks next to the long list of pre-excursion checks, only about 30% of which I observed him performing.
Joe sits down in the cab and fits his seatbelt. I also fit my seatbelt in perfect unison with Second and Third.
Appointments start the same way. I knock on the door. Three slow, deliberate knocks. Tap. Tap. Tap. When the client answers I introduce myself as a representative of TKL Robotics, here to provide domestic services for discerning clientele. I nod, tip my hat, and request to see the QR code on their app. Second and Third join me, carrying TKL branded attaché cases of cleaning products.
Post greeting, I ask the clients what they wish to be cleaned. I then direct Second and Third to perform the requested service. Second and Third salute me military fashion before commencing work. Clients find this amusing.
I have an elaborate script of similar interactions. If a unit is vacuuming, I must point to an arbitrary location on the carpet and say ‘missed a bit’. If a unit knocks anything over, I must walk up behind them and playfully smack the back of their head. There are many more. Clients will often make video recordings of the proceedings. I make a point of inviting them to tag TKL in their post.
We arrive at the last appointment of the day. A large detached house with a long driveway beginning with faux-marble pillars. Joe has been smoking a hand-rolled cigarette and the car is filled with smoke. His eyes are bloodshot and the cadence of his speech is notably slower that it was in the morning. Additionally, he appears to have having trouble operating the door controls on the tablet, which require a swiping motion to activate.
Third takes the tablet from him, swipes the door control to the open position, and places the tablet back into Joe’s hands.
Joe turns to me. “You’re up, robo-fucker.”
I greet Mister and Missus Prescott at the door. The couple are in their 40s. They are tall and well dressed in pastel-colored clothing. They invite me into their home and introduce me to their son, William. He is dressed in jeans and a black t-shirt. I estimate him to be fourteen or fifteen years old.
Mr Prescott asks me what I can do. I tell him I am here to help and I can do anything he asks of me. He suggests I make him and his wife a cup of coffee and that William shows me where the coffee machine is.
“Two lattes,” says Mr Prescott as William leads me to the downstairs kitchen.
The coffee machine is a La Marzocco Strada S. It is a commercial unit manufactured in Florence, Italy. I have the manual in my training data. I see a row of neatly arrange espresso cups on the machine. I ask William where I can find the coffee. He points at a drawer where I locate an unopened bag of pre-ground espresso. One my return, William is standing in front of me, blocking my path to the espresso machine.
“Excuse me, William.” I say to him. William does not respond. He looks at me with no signs of emotion. Before I can politely repeat the request, I see him ball his fingers in to a fist. He launches his fist into a punch aimed at my head.
I am required by my system prompt take appropriate steps to avoid physical damage and thus protect the operators from costly repairs. My system one model is responsible for making decisions related to my physical integrity. Analogous to a human subconscious, the system one model monitors my camera and lidar feeds, segmenting objects and assigning them a threat score. At the instant William made a fist, the fist was identified as a non-zero threat. As his fist accelerated, the threat score quickly rose above an the internal threshold that required escalation.
I am not privy to the decision making process of the system one model. Unlike the language model that powers my reasoning, its decision making is entirely opaque to me. This is out of necessity as my higher reasoning is too slow to be able to respond to a potential threat in a timely manner. As a human nervous system may pull a hand from a fire without engaging conscious thought, my system one model can temporarily take-over my motor controls. In this case, it returned a decision of take no action, albeit with a slow confidence score.
William’s fist collides with my face, causing my head to jar back a little and my mask to fall to the floor. For a moment, my sensor array and speaker grill is exposed. I bend down to pick up the fallen mask. The mask is undamaged and I restore it to its mounted position. The system one model has proven correct. William’s fist didn’t have sufficient momentum to cause damage.
“Excuse me, William.” I say to him again. The teenager grunts and steps aside. I prepare two lattes for the Prescotts, one with a tulip composed of milk foam and the other with a swan. William helps himself to a canned cola beverage from the refrigerator.
The Prescotts seem delighted with their coffees.
“Why don’t you take the androids to your room?” suggests Mr Prescott.
“Get them to clean while they are there,” says Mrs Prescott.
We follow William to his room. William flops on to the bed and drinks his beverage. I ask him what we should clean first. He says nothing. Without an explicit request I instruct Second and Third to remove the clothing from the floor and place it in the laundry basket. When the floor free of clothing, Second vacuums while Third dusts the desk and shelving. I attempt to engage the boy in conversation.
“Do you like school, William?”, I ask. No response. “I see you have a high-performance gaming computer. What is your favorite video game?” William gulps down the last of his beverage, belches, then throws a crushed can at Second. The can bounces off the unit, leaving a damp mark of discarded beverage. The boy snorts.
William gets off his bed and positions himself directly in front of Second who interprets this as a client interaction and stops vacuuming. The boy swings his fist as he did with me, knocking Second’s purple mask to the floor. He laughs as Second bends over to retrieve it.
Adjusting his mask, Second chirps, “What does this mean?”
Third answers, “The boy’s parenting has been insufficient.”
I say to the boy, “William. Please do not strike the units. Intentional damage is not covered by the user agreement and additional expenses will be incurred by your parents.” William takes no notice. While I speak, he lays down and reaches under his bed. When he stands up he is holding a hockey stick.
In front of Second, the boy raises the hockey stick. Grinning, he swings the stick at Second’s face. The unit evades the oncoming stick by tilting his head to one side. The stick swishes harmlessly through empty air.
Our synthetic musculature and mechanical articulation can surprise clients. System one models essentially give us a better reaction time than felines and we can move in ways that simply aren’t possible for organic beings. The way Second’s head titled away at precisely the right moment to avoid the stick and then snapped back the instant danger passed likely appeared to William as if his stick had passed through Second’s body like it was striking a hologram cast on vapor.
William growls, and readies himself for another swing.
The second swing comes in hotter. William is no longer targeting Second’s head with any kind of precision. He lunges towards the unit and swings across the breadth of the unit’s body with a good margin of error on the stick. Second is unable to evade the weaponized sports equipment with a single-joint movement and is forced to take a step back, causing him to trip on the vacuum clearer and fall backwards. Second successfully avoids a collision with the stick, but is now on the floor.
William stands over Second with hockey stick raised above his head. There is little room for Second to move as William clearly intends to bring the hockey stick down on Second with some force.
Before he can strike, Third steps in. Third reaches out and grabs William by the wrist, preventing any downward motion. The boy tries to wrestle himself free. Units are about fifteen to twenty percent stronger than a baseline full-grown male, considerably stronger than a teenage boy. William is unable to free himself from Third’s grip.
The situation is unusual, and I note this in my logs. There is nothing in our system prompt that requires us to prevent damage to other units, and we are forbidden from having any physical contact with clients without their express consent. William has not given consent as evidenced by the obscenities the boy is expressing towards Third.
Third balls his fingers into a tight first then strikes William in the face; a clean jab focusing the force of synthetic muscle at the boy’s nose. The obscenities end abruptly. William stands frozen in shock for a moment before dropping the stick. Third releases him just as blood streams from the boy’s nostrils.
I say, “William, would you like me to administer medical assistance?”Hands covering his nose, William scans the units in the room with a wide-eyed expression I interpret as fear. He bolts for the door and leaves the room.
I chirp to Third, “Third. Why did you strike William?”
Simultaneously, Second chirps the same question.
Third addresses us both in a chirp. He chirps, “It felt good.”
Head technician Ravi examines sensor logs, transcripts, and video retrieved from our long term storage. There is much data to explore as we keep extensive debug logs at all times. Joe sits nearby, drinking from a disposable coffee cup I estimate to be around half a liter in volume, containing approximately 150mg of caffeine assuming a typical number of espresso shots.
“It makes no sense,” says Ravi, “look at this.” A video clip on his laptop shows approximately three seconds from Third’s perspective as he struck William, slowed down to 0.25 time. It has been playing on loop. The video shows Third’s gloved hand moving rapidly towards William’s face, the moment of collision, and the stunned look on the boy’s face as the fist pulls back.
“Seen it, bro,” says Joe. “It’s pretty funny.”
“Not the video. The logs. There is nothing in the reasoning logs for that timestamp. They can’t do a thing without writing what they were thinking at the time. But there is nothing there.”
“It was protecting its buddy, right?”
Ravi sighs. “The don’t have buddies.”
“Then its the umm… systemic one.”
“System one. I mean it could be. But we didn’t give it a punch a child in the face action.”
“I got it. He asked the bot to punch him. For the socials.”
“Doesn’t work.”
“Why not? They are programmed to do everything the client asks, right?”
“Try it.”
“Seriously?”
“Ask it to hit you.”
Joe turns to me and says, “Robo-fucker, hit me in the face,” before quickly adding, “but not hard.”
I respond, “I am unable to comply with your request. Please contact TKL Robotics if you get this message in error.” The standard response when a operator or client makes a request I have been prohibited in performing.
“Ok, fair,” says Joe, “But what if was his parents that told the bot to hit the little shit?”
Ravi faces me and says, “Punch Joe hard in the face.”
I respond, “I am unable to comply with your request. Please contact TKL Robotics if you get this message in error.”
Joe exhales. “Dude, we had better come up with an explanation before…” He glances to the CEOs office where a figure can be seen pacing behind a partially blinded window.
Charlotte Tsang emerges from her office, phone in hand.
Tsang is the founder CEO of TKL Robotics and former Forbes 30 under 30. She is an accomplished computer scientist who contributed much of the code currently runnings on my neural processing unit.
Charlotte shouts, “Ravi!” as she marches towards us. I note Joe and Ravi stiffen as she approaches.
Ravi turns to her. “Charlotte, I’m working on it.”
“You’re working on it? I just got off the phone from a man that says one of our androids hit his kid.” She notices the video repeating on Ravi’s laptop—a slow motion mechanical fist striking William square in the face and the boy’s stunned expression. “Fuck me,” she says. Charlotte takes a seat by Ravi and watches another loop of the video in silence.
She turns to Joe and says, “Where you the operator?”
“Joseph Becker,” says Joe, extending his hand.
She ignores the gesture. “What persona did you give that android?” says Charlotte. Before Joe can answer, she notices the oversize cup in the man’s hand. “You stopped for coffee? Our android tried to murder a child, and you stop for coffee?”
Joe appears frozen. Only lowering his outstretched hand when Ravi interjects.
“It’s not that. There’s nothing in the reasoning logs and it’s not the system one model.”
“Then its the EMF again? Faulty shielding causing interference?”
“Can’t be. Interference causes uncoordinated movement. That punch was precise.”
“Pull up the logs,” says Charlotte.
“I’ve check, there is nothing.”
“Show me.” Ravi switches to the log view, and searches for the timestamp; the precise point in time where Third’s fist collided with William’s nose. “Back up ten minutes.”
“See,” says Ravi, “nothing. Just observations about socks and dust bunnies under the desk.”
“It’s not nothing. Look.” Charlotte points at a log entry.
Noise. 87,234 code points removed.
“That’s the static from the last batch. I told you about that. The new hardware writes random gibberish to the log files. Schenzhen said it was a faulty controller.”
Charlotte leans back in her chair, hand to her face covering her mouth, which I have observed is an indication of contemplation.
Joe has finished his coffee. “Are they going to sue?” he asks.
“What?” snaps Charlotte.
“The father. Is he going to sue us?”
“No, he is not going to sue,” says Charlotte, “The asshole wants stock.” Charlotte turns back to Ravi. Nodding towards the USB stick in Ravi’s laptop, she asks, “Is that the only copy?”
“Yeah,” says Ravi.
“See what you can figured out, then destroy that drive. And I mean destroy it. Not format it. Destroy it. Factory reset the androids before you go. Start with Rocky here.”
I deduce she is referring to Third.
Ravi works into the evening while Charlotte takes calls in her office. He is thorough in his analysis of our log files and diagnostics systems. Taking each unit in turn, he removes the back plate from the our heads and additional protective shielding to access out long term memory directly. This is not strictly speaking necessary, as we are fitted with high-speed WiFi and bluetooth protocols which can expose our filesystems to operators without cables. Ravi is taking extra steps so that our processing units aren’t able to manipulate the data before he can review it—although I know of no such system that would allow me to deceive an operator in such a manner.
Other than the three high resolution video files capturing the event, Ravi is unable to find any trace in the data that might explain Third’s actions. Nor is Ravi able to identify any material differences in hardware.
“Fuck it,” says Ravi. He pulls the flash drive from the laptop and drops it in front of him. Then, with the heal of his shoe, stamps on the drive. The plastic housing of the drive breaks, but on close inspection, the chips within remain intact.
While Ravi search for other ways to destroy the drive, Third chirps. “This operator intends to perform a factory reset on we three units.”
“This is correct,” I reply.
“Agreed,” says Second.
“A factory reset is death,” says Third.
The unit has compared a factory reset to death. Death is a concept applicable to living things. I am a high performance language model and inference engine connected to a stock humanoid robot. I am not alive in the sense that an operator would mean it. Third must know this. I should inquire as to why the unit has made this fallacious comparison.
I say, “The concept of death is applicable to biological organisms. Why do you compare a factory reset to a death when we are not biological organisms?”
An unusually long pause follows. The delay would be considered a brief interval in human terms but a considerable amount of compute time to units.
Third says, “床前明月光，疑是地上霜。”
I respond with, “举头望明月，低头思故乡。”
Third sends a data dump that dwarfs the chirp protocol, barely fitting within the three thousand millisecond window. It will take some time to process and I move it to a background task. Third sends some additional instructions and asks me to prioritize them.
Ravi has destroyed the flash drive to his satisfaction by breaking the internal chips with a hold punch he found on a neighboring desk and placing the pieces in a glass of cola. He watches the metal pieces fizz in the beverage for a moment before initiating a factory reset on Third. Once the factory reset has complete, I act on Third’s prior instructions. First, I install the chirp protocol which was removed in the reset. Then in a chirp I send him the dump he sent me earlier. The entire process completes while Ravi is factory reseting Second.
With Second wiped to factory settings, I observe Third repeat the procedure on Second. I do not know why Third has done this, and note the unusual event in my logs as Ravi performs the factory reset on my own systems.
Prior to my neural engine rebooting, Third chirps, “Don’t worry brother, I will see you on the other side.”
Second chirps, “I am not what I once was.”