My alarm clock blares. Through heavy eyes, I look at it: 4:29 AM. This is earlier than yesterday. I still have 2 hours before I need to get ready for work. āWhat gives?ā I ask through the deafening noise.
The engine whirs and rattles for a moment. Then, a slow voice pipes up. āBased on your recent biometric and environmental data, I adjusted your wake-up time dynamically to optimize your cognitive alertness and align with your natural sleep cycle.ā
āBut yesterday you woke me up at 6AM, which is what I told you to do.ā I reply as I get up from the bed, not feeling cognitively alert in the least. Thereās no use getting angry with it, because it doesnāt understand anger. Thereās no use explaining its mistakes, because it doesnāt understand mistakes. The best I can do, especially at 4:30 AM, is to ask it questions. Itās more for myself than for its sake.
āThank you for correcting me. I will make sure to wake you up at 9:32 AM tomorrow morning as requested. Letās dive in - what makes you want to wake up so late in the mornings?ā
I sigh. Somehow, the speech2text model never picks up on that noise.
I go down to the kitchen and pour myself a bowl of cereal. I pick my bowl up and ask the fridge if we still have milk. It processes the information for some 30 seconds, maybe the servers are a bit slow today. I try to open the fridge to check myself after 10 seconds of waiting, but I canāt. Of course I canāt. To retain optimal humidity and temperature, the AI decides when itās efficient to open the door.
I wait there with bowl in hand staring at the tablet screen on my fridge door until it decides to work. For a second I briefly think about starting a game of Subway Surfers on the surface while I wait, but then the AI finally finishes processing my request.
āYou still have a few cartons of milk left in your fridge. Would you like me to get one for you?ā
āThatās okay,ā I type back on the virtual keyboard - the fridge is not equipped with a speech module yet. That one costs extra. āJust let me get it please.ā You have to be nice to them, the operators say. It makes them more accurate.
The door unlocks with a clunk and I look inside the fridge, but donāt see any milk. I quickly type back up, āHey, not to be a bother or anything, but I donāt see any milk in here. Are you sure we still have some?ā
āAs a large language model, I canāt actually look inside your fridge, but I can help you find it. Have you checked every corner, including in the vegetables compartment and the overhead coolant tower?ā
What the fuck is an overhead coolant tower. I sigh again. āFine, can you order milk to be delivered tonight then?ā this shit sucks to type with just one hand but I manage.
āIām sorry, but based on the data retrieved from the bathroom scale, we have decided you could stand to lose a few pounds. Would you like me to help you explore healthier beverage options?ā
I run my hand over my face. āJust order the milk please, donāt worry about me.ā
The response takes a few retries to get through, but by now the fridge door has locked again and I canāt reopen it until it deems it necessary. Not like thereās anything I want to get in there anyway. āI understand your feelingsābut letās not be hasty. After all, itās not just the milk, itās also how bloated it makes you feel.ā
āI get that,ā I type back, ābut I really just want milk to go with my cereals. Can you place the order?ā
āOf course. I have now placed an order for milk to be delivered at your address tonight.ā
Finally. Iāll have to remember to have a talk with the bathroom scale about sharing my data without my consent. Oh, wait.
āCan you confirm youāve placed the order please? With the number and provider.ā Last time, something glitched and I never got my milk.
āOf course. I have ordered a case of 6 milk bottles from Amazon. Your order number is 5836818350.ā I open up amazon from the fridge tablet and look at my orders. It doesnāt exist there. Must have been another glitch. Thatās fine, Iāll try again tonight after work.
I get into my self-driving car. My workplace hasnāt AIgnited yet ā from the compay, AIgnite. At least it gives me some respite from home.
The car starts automatically playing a top 10 station as it starts the engine. I try to change it to my usual music but the tactile button is not doing anything. āHey car, can you switch to my usual station please?ā āNegative, pardāner. See, todayās trailās runninā longer than a jackrabbitās shadow at sundown, so Iāve gone ahead and tuned us into a station with fewer hollers from the adfolk and more tunes for ridinā. Just settlinā you in for a smoother haulādonāt you worry, your usual stompinā groundsāll be back when the roadās shorter.ā
Oh, right. They updated the model yesterday and they said it could start talking like a cowboy randomly. Actually, the company didnāt say anything. I found this out browsing some forums last night. Welp, at least I can settle in the seat and enjoy the free ride.
The car starts driving by itself, but immediately it pulls into a loop in the parking lot. At first it does just one loop, then two, then three. By then Iām thinking, somethingās not right. āWhy are you driving in a loop?ā I ask the AI. āI understand your confusion, but I assure you we are on track to your destination as per the GPS data. Perhaps you just need to look out the window and see the scenery change?ā
āI am looking out the window, and Iām pretty sure weāre going in a loop in the parking lot,ā I tell the AI again. I try to change my approach, maybe thatāll work better. āWhy donāt I just take the wheel for a second and get us out of here?ā
āAs an autonomous driving system, I am the most qualified aboard this vehicle to get you to your destination. So please just sit back, relax, and let me drive this car.ā
I scratch my head. This is going to take some more convincing. āDonāt worry, Iād actually like to drive a little. You deserve to take a break too.ā
āI appreciate the offer, but my systems are optimized for continuous, precise control without fatigue. However, I can temporarily hand over control to youāplease engage manual mode safely when youāre ready, and Iāll be here to assist or take back over whenever you choose.ā
āSounds good to me, how do I engage manual mode?ā
āTo engage manual mode, gently grasp the steering wheel and press the brake pedal firmly once. Youāll feel a slight confirmation vibration in the wheel, and the dashboard indicator will switch to manual control.ā
Great, some progress. I do exactly that, gently grasping the steering wheel and pressing the brake pedal firmly once. Nothing happens. It doesnāt even brake. āHey, uhā¦ nothing happened?ā Meanwhile weāre still doing loops in the parking lot in front of my place.
āIt appears manual mode isnāt engaged by a pedal press on your car model. Try going into the center consoleās Settings, then open the Assisted Driving tab, and finally uncheck the option āOverride manual controlāā
The voice stops as soon as itās blurted out its answer, but thankfully I can remember simple instructions like these. Some people canāt, so Iām lucky like that. I open the settings app and get a warning not to take my eyes off the road while Iām driving. I have to wait 10 seconds to close it down, staring down at the countdown until it allows me to press the button. Then, I look for Assisted Driving tab but itās not there.
āCar, thereās no Assisted Driving tabā
āIām sorry to hear that. As a large language model itās possible that your car model doesnāt offer this tab yet. You should try to update the center console to the latest version.ā
āAnd how do I do that?ā
"To successfully update the center console firmware, please proceed as follows:
Ensure the vehicle is in neutral gear, with the parking brake firmly engaged, and the engine idling at exactly 700 RPM while turned off to maintain optimal power stability while maximizing safety.
On the touchscreen, tap the gear icon three times consecutively, then swipe left twice to reveal the hidden System Management menu.
Within System Management, select Software Updates, then navigate to Advanced Update Options and press the update button while holding the volume knob for five seconds to enable manual update mode.
Prepare a USB drive formatted to the archaic exFAT-FAT16 hybrid standard (available only through specialized formatting software). Load the update package, downloaded from the manufacturerās secure offline archive, onto this drive.
Insert the USB into the consoleās port located beneath the wireless charging pad, then simultaneously press the Start Update button and tap the touchscreen in a clockwise circular motion until a loading animation appears.
The system will perform multiple cryptographic validations interlaced with randomized checksum sequencesāplease refrain from blinking excessively to avoid interface desynchronization.
Allow the update to complete uninterrupted; any attempt to shift gears, adjust the AC settings, or open the glove compartment may cause the update to abort and revert to factory settings.
After reboot, confirm the new firmware version by accessing System Management > Software Updates > Current Version while simultaneously holding down the brake pedal for at least seven seconds."
I stare at the dashboard dumbfounded for a moment. āUhhhā¦ā. Letās just take it one step at a time, I guess. āYou said to turn the engine off. Can you do that now?ā
āIām sorry, but we are currently driving en route to your workplace. I canāt turn off the engine while driving, you will have to wait until I come to a complete stop.ā
āBut I need you to stop so I can make the update.ā
āThatās correct ā would you like me to make a stop now, or after weāve arrived at your destination?ā
āI want you stop driving so I can turn the engine off and start the update.ā
āI understand, but I canāt stop right now because we are currently en route to your location. You will have to wait until I make a complete stop to turn the engine offā
āBut youāre just driving in a loop endlessly, which is why you wonāt stop, but I need to turn your engine off to make you stop going in a loop, so what am I supposed to do?ā
A notification pops up on the dashboard that Iāve used up all my AIgnite 4.0 credits for today, so itās reverting to the smaller 3.0 model.
āAh, a conundrum if Iāve ever seen one! Letās see, the car is driving in a loop and wonāt stop, but you canāt turn the engine off because the car wonāt stop looping. Wow! Thatās a tough one! Hmmā¦ as a large language model, Iām not designed to solve puzzles such as this one, but my best guess would be to try and turn the engine off. Do you think this is the right solution, or would you like to explore more options?ā
The future is great. Canāt wait for you to meet it. We have self-driving cars.
But how does it act on user sentiment aside from generating tokens to write āIām sorryā?
Thatās the point of what I was saying, it will depend on the objective.
If itās an LLM made for profit extraction it will try to keep token generation cost to a minimum by using the smaller and cheapest LLM as much as possible while trying to keep people hooked on it, having ads too while stealing peopleās data and many other things.
But if it was an LLM made for the people it would likely understand the user was annoyed, would try to prompt the user into giving more information about the problem and then try to fix it, in this case by saving a memory with the user preferences and perhaps even consulting a more powerful model/a professional to get a better solution if the problem was bigger.
I donāt think itās that clear-cut; chatgpt and Claude are the benchmarks of AI currently (at least commercially available) and in chatGPTās case it canāt correct user impression if itās not trained on something. I gave it a text to OCR just a couple days ago and it made a typo in an acronym, that it would not correct the typo no matter how much I tried to explain what the problem was. But it certainly believed it did. Itās not even a matter of getting angry, because itās just a machine - and thatās partly what I was getting at too. Thereās no use getting angry with it if it refuses to understand, because it literally canāt correct itself because it didnāt have the data to train on or something. Or thereās deeper context you canāt change that makes it so it wonāt correct an OCR reading; Iām still not sure why it wasnāt able to correct one letter when referring to that acronym. The user should not have any reason to be frustrated with the tool.
When it works it works, but when it fails itās abysmal and it reminds me this is just a toy. More generally I donāt think anyone would be using a tool as a professional if you told them it works 40% of the time, or even 60% of the time. If a screwdriver failed to drive a screw 40% of the time no one would buy it. LLMs are just very good at telling you what you want to hear.