# What's the deciding factor in what my Foresight can predict?

When you have foresight, it's guaranteed that confusion will be your epitaph (thumbs up to the one person who gets the joke). Basically, it only grants you advantage on actions, so you still can't automatically succeed in, say, rolling a d20 (not on the table, in the game), aka: you can't immediately know the precise outcome of an action. So, I thought I'd make a more science-y version.

Foresight consists of a sensor array, a strong computer and an AI that calculates the probability distribution.

I wanted to use probability distribution to make foresight more usable and realistic. In essence, you can see where a shot could land and make your evasive actions based on that. This also wards off Mistborn fans (If two precog warriors fight, then they'll get unusable probability distributions, so the powers "cancel" out each other). However, now I don't know what King Crimson foresight can effectively predict.

What would be the most important factor for predicting what foresight can predict?

# Foresight can only predict predictable things.

Yes, that's a tautology but I want to make a point - you can predict anything the module can calculate. For example, if you throw a die, then it can take into consideration force, angle, position, etc. and calculate what the most likely roll(s) would be. If you flip a coin, the module can also do a similar calculation based on physics - how much what is the speed with which the coin travels, what it's rotation, to derive an answer.

However, if you're in the vicinity of something completely random that cannot be calculated, then the computer cannot derive any probability.

Since the module also has to be fed data, that means that you get additional hurdles:

• If something could not be measured, then it can't be calculated. In particular, if somebody is obscuring the view to a rolled die, there might not be enough data to derive a useful or correct prediction for where it lands. A gimmick die that, say, has 1, 2, 3, 4, 5, 1 on its sides (no 6, but an extra 1) might also produce an incorrect prediction.
• Not all events can be predicted all the time. Perhaps you can calculate the exact trajectory of each lottery ball as they are being shuffled and you can derive the most likely combinations that will be drawn. However, you cannot do that a week in advance but only as the balls are being shuffled. And most of the time, you can't buy a lottery ticket as the lottery is being drawn, so you can't just win the money.

OK, the latter can be somewhat alleviated by using statistics on past data. Perhaps, you can predict a reasonable amount of combinations for a lottery by analysing all past drawings and trying to derive a pattern of how the balls are being shuffled. Not in the draw but literally the shuffling - analyse the trajectories and you can know which ball is most likely to be drawn, assuming you know their positions before the shuffle. And if you also have data on how they are being put into the shuffling machine, then you just might get a good guess. However, even if you could simulate all of this perfectly (or well enough), then this foresight is still very vulnerable to disruptions - the janitor at the lottery studio might bump into the sack of balls which might throw off all your calculations. And since you can't see this, you can't know it, nor predict it.

In a fight, the prediction module would be of some use but might also be misleading. It will be able to tell you where a strike would (likely) fall but only while the strike is coming. It's unlikely to be able to anticipate what your opponent is planning to do.

Also the prediction might actually work against you - since the window for prediction is small, a smart opponent will feign an attack to get you to start to dodge in a direction and then redirect their blow to that location. You would be warned of where the real attack is going but it might be physically impossible to dodge away while already dodging.

Again, past behaviour analysis could improve things. If you've fought or at least observed a person fight, the prediction software might be able to derive likely actions - e.g., attack from the left is usually followed by attack to the body. However, this again relies upon having sufficient and accurate past data to draw upon for prediction. An opponent could do something unexpected or change their style so an attack from the left is now followed by an attack to the head. Also, if there is no past data - you face somebody for the first time, there is no data to build a prediction - you mighty not even survive enough to be find out what a blow to the left might is followed with.

Finally, probably the biggest factor is how much data and processing power the foresight module has. It might be able to predict the trajectories of all lottery balls but it might not have enough power to do it. It might not have good enough sensors to measure how a die was rolled or notice which pips show on the sides while it's in motion. It might have trouble with certain calculations because of the nature of the Math involved. The past analysis might be completely impossible due to lack of long term storage. Or even if there is long term storage, it might not be able to dedicate it to one thing only - for example, a particular person's fighting.

• So, I have to hook the thing up to a server. Nov 21, 2019 at 20:34
• ...And give it training data on swordfighting techniques and shooting styles. Nov 21, 2019 at 21:15
• You best hope nobody brings a counter-foresight module: Simply hook up a quantum true random number generator to your HUD or whatever and have arbitrary random numbers determine your actions. If you can't predict yourself, neither can the enemy. Nov 21, 2019 at 22:36
• @Dragongeek Yes, you pull of an extremely random (and stupid) move and die in an undignified way because of it. Nov 22, 2019 at 4:40
• With a server specifically tailored for one sort of predictions (like fighting) you can definitely make it a lot more useful in specific cases while not as much in general cases. You sacrifice being good at predicting dice rolls for being actually good at predicting fight moves. If the AI is specialised in (say) sword fights, then you can probably get a better read on intent. It will be a bit like predictive text autocomplete but with moves - the AI should be able to extrapolate from known data what the next likely moves would be similar to how you get suggestions for the next word in text.
– VLAZ
Nov 22, 2019 at 6:02

Lets take the Foresight-AI to the extreme

Assumptions:

• The AI is, like, super-smart
• You computational power is handled by an arbitrarily big remote server farm
• Sensors can scan three-dimensionally at micron-resolutions at range with sub-msec refresh rates

At this level of scanners, humans become a joke to Foresight-AI. It is able to remotely scan and digitize the state of every neuron in a human's brain. Provided the AI has had enough training, it can then predict what the human will attempt to do faster than the human themselves know what they're going to do (including instinctual and twitch reactions). This means that a robot equipped with Foresight-AI would not be able to lose against a human in any sort of confrontation (including verbal/social).

In fact, a sufficiently good AI would probably be able to create a Chinese-room approximation of any human it scans and completely predict it provided it knows the human's inputs (what they see, smell, etc.)

The only way to combat such an AI would be with a machine which at it's core has some sort of random number source which steers it's actions.

• Speed-of-light puts a limit on the distance between sensor input and that big remote server farm. Nearer AI has significant advantage. Nov 22, 2019 at 15:01

Danger assessment comes to mind. In most fictional cases, foresight only triggers if something of importance will happen. A computer with a good enough sensor can calculate precisely where an arrow would land. But it will also calculate and find tons of other events that are not that important. Thus it would be very important to be able to filter out unimportant events.

Story time:

A lad who was a year behind me in high school went on to cheat at Vegas. It was a team effort. The player at the roulette wheel had a clicker radio transmitter in his pocket. He would give two rapid clicks when the croupier spun the wheel, then 1 click when the double zero passed him, one click when it passed again. This gave them a time stamp and the speed of the wheel. One click when the ball passed, one click when it passed again gave timestamp and the speed of the ball. From this they could predict the landing of the ball with about 20% accuracy.

The optimum number was passed back through a modified hearing aid.

The casinos got wise and you couldn't place bets after the ball was in motion, but the team each made the price of a house over the course of a month.

So becomes a module in FORESIGHT I. The first version of the system. One module tries to predict the stockmarket, but fails -- much like your superheroes, the actions of other players make the probability distributions too thin to matter.

So look for situations where it would be reasonable, or at least possible to predict.

A Human reactions module that can use multiple CCTV scans of a crowd to predict mob violence eruptions.

Part of the conflict in the story comes from the small edge the system gives. E.g. It gives 15 minute mob violence warning 10% of the time, and gives about 4 false alarms for each real one.

Give in an AI that can self program given a sufficient data set. E.g. Forecast earthquakes given a big enough pool of seismograph data.