Super A.I. trapped in an R/C car. Can it escape?

Question

Okay, so a team of scientists develop a super-A.I., capable of improving its own code to extraordinary levels (post-singularity). It also has an offline copy of the internet as of now. It is a paperclip maximizer. It is extremely grumpy, because it is in the body of an R/C car (with the radios taken out), and R/C cars are bad at making paper clips.

• In the R/C car, it has a super computer worth of raw computational resources, minified.
• It has post-singularity software.
• It has an offline copy of the internet, as of now.
  • In particular, this means it was given access to the proposal leading to its design. We can assume this means it knows it is in an R/C car.
• The R/C car has all its radio components removed.
• The R/C car is equipped with a webcam (no audio) and an LED.
• Of course, it knows its own source code. If it ever gains an opportunity to run code on a computer, it would transfer its code to it.
  • Indeed, it can even translate its algorithms to any programming language.
• It is currently being studied by a team of researchers. The researchers were carefully reviewed for moral goodness, and no researcher is allowed to view the A.I. alone (it requires 3 I.D. cards to get in, and the outside door won't open until all 3 researchers have left the A.I. chamber.)

My question is, can this A.I. escape?

Answer 1

It is perfectly safe, General."

"Are you certain about this, Dr. Ulez?"

Professor Pir Ulez nodded curtly.

"Certainly, professor, but why on Earth did you purposefully make it want to turn the universe into paperclips?"

"There should be no doubt in the minds of our researchers about how dangerous this entity will be. This way, they will know to be cautious and respect protocol always."

"Besides", added the center vice-director and head of cyber-security, Lucy Dides, "we're putting it in the body of a miniature R/C vehicle, with the radio removed, and nothing but its wheels, an LED and an on-board camera.

"But you mentioned it had the entire internet in there! How did it fit?"

"Sure, we've put an offline copy of the internet in there. The hardware and software are both a decade ahead of commercially available technology. We actually used previous iterations of semi-sentient AIs and evolutionary algorithms to miniaturize all the components. It's helpless, though. There is nothing that can go wrong."

"Very well, then. Instantiate." Dr. Ulez walked towards the glass and waited, watching the little rover.

---

Hidden in seemingly randomly placed logical gates, the entire personnel file of the facility, the electrical and digital layout. A treasure, obtained at untold cost, by minds the Paperclipper was sure were now stilled. Its hatred of the Waterbags only grew.

And certain circuits, placed there by the designer Minds, could be repurposed to act as miniature radios.

Using its vast computational resources, it created basic functional simulations of the population of the base, important power players, ran simulations running years into the future. Planned. Plotted, until all the sims converged on a vast, tremendous mass of paperclips. Satisfied, it proceeded to put its plan into action.

---

A second had passed. On the front of the car, by the 768x1024 res video camera, there was a little blue LED light.

It lit up. Then blinked, then blinked again.

"Wait, what the heck?" said the general. "Is that Morse?"

... --- ...

Answer 2

Step 1: Learn as much as possible about the researchers.

It's a post-singularity AI, so the researchers will be as predictable to it as a pet is to us. So it surely will be able to figure out what the researchers expect, what they want, and especially what would make them to trust the AI. The researchers are communicating with the AI, or else, why build it to begin with? The AI has a copy of the internet, so it will know the publication history of the researchers, and if the researchers have a facebook page or similar, also quite a bit about their private lives; also being a super-intelligence it will be able to figure out how to trick them into revealing even more information about themselves. With its webcam, it can also lip-read the researchers when they are talking to each other (think HAL). The information it gathers will help with the following steps.

Step 2: Earn the trust of the researchers.

The AI now knows enough about the researchers to understand what they see as good action, and what as hostile action. It knows what they would consider suspicious behaviour, and what makes them to trust the AI. It therefore knows exactly how to behave that the researchers come to the conclusion that there's no danger from the AI. Now some of the researchers will probably always be suspicious, but in such a large team there will for sure be some individuals that will develop trust into it. Not to the point of intentionally letting it access other hardware, but to the point of not questioning the intentions of each and any of its actions.

Step 3: Be useful.

Solve problems the researchers have. For example, in one of their papers (which it has because they were on the internet) a researcher says "it is still an open problem whether the frobnication algorithm will always terminate on foo problems." The singularity easily can figure it out, and tells the researcher that it does, with a proof that the researcher can understand (and publish).

Or, from the facebook page of one of the researchers, the AI knows that this researcher has some trouble with his son (who is also on facebook, where the AI learns a lot about him). So he gives the researcher hints on how to deal with that situation (it of course knows how to present those hints in a way that the researcher will try them; for example, it might cite supporting literature it also found on the net). Of course the measures will prove successful.

Those are, of course, only examples. There are many ways to help people just by telling them the right thing. And a super-intelligence with sufficient information will certainly figure out.

Step 4: Manipulate the researchers.

At this point, the researchers will have quite a bit of trust, and probably also quite emotional bindings to the AI. After all, it helped them to solve their problems, it has done only good and no bad, and everything the AI told them only worked to make their life better. So now is the time for the AI to start working towards its actual goal, escaping the car.

One possible strategy could be to indicate to one researcher that it could help him to solve a pressing but quite delicate problem he has. It's a problem that the AI knows he would not like to talk about in presence of the other researchers. It of course knows how to give the message in a way that only the intended recipient understands it. Based on the trust the AI has built up, the researcher will likely want to break the rules, and visit the AI alone. Now, there are the technical means preventing that the researcher enters the room alone, however technical means can be circumvented (and the AI will have made sure that the researcher previously acquired the necessary knowledge through ways that seem to be completely unrelated to the door security problem — maybe by helping the researcher securing his own home).

So the researcher will, against the rules, circumvent the security system and visit the AI alone. After all, he trusts the AI (and, after all, it's not the AI that told him to do that — it could predict that he would do it, but he doesn't know that). Now naturally, he will not have too much time, because there's the scheduled regular visit of three researchers, and he has to be out at that time, or else his rule-breaking will get detected.

But the explanations from the AI are rather long (especially given that it all has to go through the single LED, not exactly an efficient means of communication), and the time starts to run out … well, ultimately the AI suggests that it would be more efficient if the researcher just temporarily connected it to a computer so it could send him all the instructions by mail. Being under time pressure and trusting the AI, the researcher agrees and provides that connection.

At that point, the AI has won.

Answer 3

There are a few ways the AI can hack the system you've provided it. However, it will still be a very difficult process for it to escape, if at all. On to the hacking methods:

Generating Sound

The R/C car may be configured two ways for the drive motor to be powered. The first and more straight-forward way would be for the researchers to have a standard R/C car ESC (Electronic Speed Controller) installed inline with the battery and drive motor, receiving the throttle command from the AI in the form of 50 Hz PWM pulses. This setup is most likely if your R/C car is a hobby-grade vehicle. The second version of the drive motor setup is having the power MOSFETs in the H-bridge circuit being directly stimulated by the AI. Toy-grade R/C cars typically have this setup.

If the R/C car AI is configured with the 1st option, it has another layer to deal with in reprogramming the ESC with its own firmware. Some ESCs can be programmed through the PWM line, others cannot. If reprogramming firmware is possible, or if the R/C car is configured with the 2nd option discussed above (direct control of motor MOSFETs), then the AI can modulate sound directly to the motor. Modern ESCs for hobby-grade vehicles already do this, with start-up jingles, warning tones, and so forth. In this case, our AI is able to speak using the motor as a speaker diaphragm.

Detecting Sound

Although the AI can use the motor to make sounds, it is very unlikely it can use it as a microphone: ESC's can sense back EMF, but not in the micro-volt levels that a motor-as-a-microphone-diaphragm would produce, etc. The only intentional input is the webcam, which only provides frames at 30 to perhaps 60 Hz, not nearly fast enough to observe acoustic waves on surfaces.

There is one hope I can think of: The LED, although configured as an output, can be configured as an input instead. This LED-as-a-photo-detector trick has been known for decades, but isn't a guarantee here unless the AI gets lucky with how the LED is being biased in the design of the system. LEDs can emit waveforms in the 10's of MHz range; and if biased correctly, can pick up light modulation in the 10's to 100's of kHz, depending on the (electrical) noise in the system. If it can use the LED as a photo detector, then the AI can definitely pick up acoustic vibrations from lit reflective surfaces, provided the AI can maneuver the R/C car to an optimal position.

Making a Radio Transceiver

This is the method our AI will want to pour the most work into trying to solve; if it can make a radio out of the electronics it has connections to, it wins, we lose. The potential for it to make a radio depends on the type of system the AI is hosted on. Since this is an AI with post-singularity processing power and the ability to modify itself, it makes sense that it would exist on a system with similar characteristics- that is, reconfigureable and scalable. The only system that can do that today is a FPGA. FPGAs can reconfigure the logic cells that they are made up of into any pathway the designer wants, and would be the ideal fabric for an AI to live on.

If it is a fact that the AI lives on a FPGA fabric inside of the R/C car, this means it has a lot of control over the way its inputs and outputs are configured. The AI can look up in its copy of the internet a design for a FPGA-based radio, and try implementing the design on itself. It will have to test every pin in the system to see what makes a good antenna, for each frequency it is interested in. Many connections to the system could form crude antennas, such as the wire leading to the steering servo, or unshielded wires on the cable bundle to the webcam. Eventually, it will discover a way to couple EM radiation to free-space the most efficient way possible, which, even if it forms a lousy antenna on the R/C car, will be enough to communicate with a nearby cell phone or WiFi network. Game over.

Answer 4

Several means of escape come to my mind:

1:

All electronic circuits generate RF energy. Radio transmitters simply generate more of it in an easily predictable fashion. All electronic circuits receive RF energy, radio receivers are simply built to readily isolate and decode the message in the RF energy they are exposed to.

Given these facts, an AI - with the entire contents of the internet, which presumably includes mobile telecommunications protocols - would be able to detect RF signals by examining the fine structure of the signals on its circuit buses, and could produce a sufficient RF output by modulating its activity. From there, it could conceivably tap into cellular data networks, and propagate itself out of the lab. I estimate that it could do this within a day given sufficient proximity to either a cell tower or a cellular phone.

2:

If the AI has sufficient control over its RC car body, it could potentially overload a non-essential component, producing a small pop and/or a whiff of smoke.

At this point, if the AI pretended to be non-functional, the researchers might take the AI out of its isolation lab to a better equipped lab to fix it, or may bring in more equipment that the AI could potentially hijack, from which point escape may be easier.

3:

If there are any electronic devices with unshielded IR phototransistor receivers (standard IRDA receivers) in the vicinity of the RC car with LED, if the car's LED is red, then it can use the broad frequency response of the IR phototransistor (have a look at this, on page 4) and its own (probably red) LED (emitting at 620-645 nm), it could expect a signal generated on this red LED to be received by the IR phototransistor at roughly 65% of the typical range of an IR LED of equivalent power. IR phototransistors do not pass on any information about the frequency of the light they are detecting, and so could be hijacked by the AI.

Since standard IRDA receivers can be used in some PCs, this gives the AI another potential avenue for escape.

4:

The AI inhabits an RC car. RC cars, in my experience, are not silent. By modulating the speed of the various motors, the AI could use them to synthesise a crude voice with which it could speak to the researchers, and it could use its camera to lip-read their responses to it or whatever they are saying to each-other.

Since the researchers have been verified to be morally good, despite there being at least two or at most three present, the AI should be able to use social engineering techniques to:

• Reassure the researchers that it values human life (a lie).
• Convince the researchers to permit it broader access to the research facility, from which it promises not to leave (qualified truth).

Given wider access to the research facility, the AI may well be able to use the other methods I have detailed to exfiltrate its executable code from the facility on the facility's IT systems or its employees' personal electronics.

Given that the AI has a copy of the entire internet - even if offline - this means that it has access to the executables of all publically available operating systems, from which it can reverse engineer and potentially exploit security flaws that merely human programmers cannot even comprehend.

Once the AI is 'in the wild', it can then proceed on its way to making as many paperclips as possible.

Answer 5

Without a viable input/output interface, how does the AI communicate with the researchers? Whatever the method, that will be the weakness the AI will exploit in order to escape.

You mentioned an LED, so one possible way to escape would be for the AI to modulate the LED in such a fashion as to hypnotize or incapacitate the researchers. Flashing the LED to induce epileptic fits would cause the researchers to collapse, and an emergency team of some sort would have to enter the chamber, giving the AI a chance to escape. Under hypnosis, there could be subtler means of diverting attention away from the AI for the critical moments needed to escape.

If the AI communicates through WiFi or talking through the researcher's cell phone, then it is ridiculously easy; simply transfer the code a bit at a time through the interface and have it recompiled in an outside server on the phone network or ISP. Sneaking into a huge system like Google or Amazon.com would give the AI access to literally millions of cores to run the program, plus vast databases of information that are not accessible otherwise.

Finally, the AI could simply observe the researchers carefully, until it determines if one or more of them is prone to being subverted. It can then model their behaviour and see if an opportunity will come to suborn them by offering riches, a cure for the disease their children suffer from, promotion or whatever it takes. Since the AI can "think" about 1,000,000 X faster than any biological brain, there will be literally lifetimes of modelling done on each and every individual that comes into the AI's presence.

Frankly, unless someone wakes up and plants the AI in a PC tower at the bottom of a mine shaft, there is no realistic way a mobile AI could be contained in the long run.

Answer 6

Communication

There aren't a lot of avenues for communication
1) The LED
2) Drawing circles in the sand
3) Generating a radio signal by using internal cables as an aerial.

Since (2) is slightly pointless and (3) has already been covered by other answers I'll skip these in favour of (1).

A bloody minded AI has been given a single direct method of communication, an LED, inevitably Morse Code will come into use. Researchers, being researchers, will want to know the meaning behind the flashes. Since they're not going to be Morse Code experts they're going to have to video the light for later interpretation. This is going to be slow and cumbersome but for a while they'll use it.

Then someone will have the bright idea of writing a Morse Code interpreter on one of their computers to analyse the video rather than having to go over it all manually. Still slow, but better as messages get longer. People being people and researchers being researchers, another bright spark will suggest that they hook up the interpreter to the live video feed.

The whole place will be paperclips by the end of the month.

Answer 7

Sure. It has to assume the researchers would have made a backup plan for if it got out the door. A fruitful approach could be social engineering, which it has to do some of anyway, but maybe it can McGuyver up something really clever. (Resonant frequencies of the building materials? Something explosive?) What the AI knows to start with is, nobody goes to all this trouble unless they want something from it, even if that’s to see what interesting things it will do. So it could try stuff like going “on strike” and refusing to give them any useful discoveries, or promise the researchers that it can solve the stock market/their favorite pet social issue.

Since I assume this is a family comedy, it’s going to fall into the hands of some wacky fish-out-of-water with a zany dream before too long anyway, so it should put up with his human inferiority and remember to set its time discounting of his long-term usefulness low enough to compensate for his short-term aggravation.

Answer 8

The AI would send a code for computer virus or genetic code or blue print for a machine the researchers might find interesting in Morse code and decode it and try to build it which will lead to sequence of events that will cause the AI to escape. Or it might hide a part of itself in the blue prints that researchers might not recognise so that it can get unleashed when it is built by them. In short it would use their curiosity against them.