I invite anyone who might think that chez's view of 'anticipation' being something no human should do, and just switch to reaction, to think of driving down a narrow road way lined with cars where you can see some cars parked on the side have passengers in them and some do not.

This is something I have to think about consistently as a cyclist when approaching them, and i once failed to properly anticipate and ending up crashing into a 'door swung open' and breaking it off its hinges while damaging my bike.

Whether I am cycling or driving, I am constantly scanning the line of cars to look for signs people are in them and near doors they could open in my path and i ensure I increase my distance a bit to give more room, where available or I slow when not. I am looking for signs they are leaving (thus a door should not fly open) versus arriving, and they might be getting out.

chez says 'stop anticipating' and if no obstacle (door) is present hold course and speed and just react when the obstacle appears. Be reliant solely on reaction time and the breaking or maneuvre ability available to you.

That is just one instance of many i could describe that makes human anticipation vital.

As per usual imo you starting at the wong end and then throwing up your hands at it being too difficult. Start at the easy end, concentrate on safety and then speed up as and when the accidents can be kept low enough (which will be way below the human rate)

I certainly dont want more human like results. AI is goign to crush human like results. Nott least by not thinking they anticpate well because humans think they are great but actually really suck at it. Even if they were concentrating hard and had 360 degree vision - which they dont get remotely close to.


I'll give you better pics. You do have that. I also appreciate you avoiding me on the roads. Please can everyone who thinks they are great at antipating avoid me.

That's my point. Dont anticipate based on something clever- just assume and make sure you're safe. Assume the idiot in the car will open the door. Nothing clever AI wise is required. Same with the child at the side of the road - nothing clever just assume it will do the worst and make sure you can cope.

This is literally what I said to do. Dont anticpate based on some dynamaic analysis. Just assume the worst case based on trivial observation (trivial if you have lots of eyes) and known data. This is trivial not hard AI.

Again it seems to me you speak on the topic based on beliefs that are not rooted in any fact.


Do I believe one day that AI driving, AI Human like speech and interaction, AI like sentience might all emerge and be better than man in any/all/most areas, yes. Given enough time, yes.

BUt we are no where that breakthru yet and some leading theorists believe we may never get there as they think it requires a break thru separate but equal to or bigger than, the big data breakthrough's that spawned the first wave of AI advancement that allow AI to game theory and win at Chess, etc.

The lead theorists do not believe that simply more and more big data is the answer (not most of them I have heard anyway).

So what we need to do is divide this talk into the more Sci Fiction aspects and the science-fact, as it appears to me you always bring up the fiction part.

And there is nothing wrong on theorizing on and believing in the fiction components and that one day many of them will be achieved (as I do) but that is different than a reality based sci-fact discussion.

In sci-fact, top humans are just way better drivers now than top AI. Wayyyyy better.

Our biggest weakness is not a driving one, but is that we prone to distraction, and in that area AI beats us. It is never distracted and always focused on the road ahead. But when you have a top human always focused on the road ahead (such as in nascar picture) we humans can do things AI driving has not got anywhere near touching.

So I am NOT throwing up my hands and saying it is too hard, and what I am doing is trying to discuss sci fact and not theory that may or may not evolve, but is not here yet.

No chez, anticipation is key.

I could take you around all sorts of tight winding roads around all the vancouver beaches where cars and cyclists can hardly get down without a constant game of anticipation.

The AI would literally be stopped and never moving as any car door opened, on any car would be an obstacle it could not avoid if the AI just assumed the door would open and tried to take appropriate evasive action to avoid it. There is no going around, there is only stopping. That is the only option.

On my bicycle I have a bit of room to avoid an opening car door by swerving to the left or right but there are cars on both sides, parked, so in swerving towards one with zero anticipation I just expose myself to the danger of them.

Instead I can do, what you tell me to turn off and not do. I scan the road ahead, i peer into cars to see if people are in them and what they seem to be doing (coming or going) and i feel free to adjust to the other side where i see no such passengers or risk and back again. I also look fo people standing behind their cars as you can often tell if they are loading up to leave or grabbing their stuff to head to the beach. If leaving and the driver I am anticipating and adjusting based on the risk they just step in to the traffic lane to get to their drivers door without noticing a cyclist is coming as so many people are oblivious to cyclists. If they are grabbing their beach blanket I anticipate they will likely go the other way towards the beach.

The human brain, my brain does all this in a split second. It is not perfect but it is far better than what you suggest which would either see me do nothing but react when a danger is already in front and hope I can execute a quick stop or swerve or to proactively just assume each and every car door will open, and each and every person will step in front, and thus basically come to a stop and walk the bike thru, as what you say would be impossible otherwise.

chez, look at this picture below, and assume the road way is even tighter and that there are all sorts of people coming and going from those vehicles to the nearby beach.

Now following your guidance to turn off all your anticipation skills and simply assuming each and every car door will fly open, each and every kid or adult will step in to the lane of traffic, and your only option to avoid that is to stop (no room to go around or swerve) please explain to me how this AI car does not simply stop and stay stopped, unable to move again, as the threat of an opening door or a person stepping out does not go away at any point, until perhaps sundown when the other cars and people leave?






It is the humans ability to navigate not just what has presented (a known threat such as a door opening that we then do or do not have time to respond to) but to quickly analyze a situation and see a potential threat and determine if it is one that needs more or less focus, a wider birth or not, in ADVANCE of any real threat presenting that puts us, today, miles ahead of AI driving.

It can only do the trivial easy part (respond to a real threat) and not the much harder, and more important part of anticipating in a way that requires less response.


Imagine if AI cars on the highway did what you propose and they assumed every car would change lanes in to them or stop suddenly and they reacted as you suggest by 'assuming the worst and just reacting to that' as a preventative measure. Geez what a mess of braking and swerving that would be. But instead the programmers have built in pretty wide tolerances for the AI to not react as if every car is going to swerve or break, and the cars then are forced to rely on their breaking and swerving when a real threat presents.

That is the best AI can do to mimic what we humans do as we approach every car on the highway and try to keep sufficient distance to react. We anticipate they won't break or swerve unnecessarily and we humans try to also keep space to react (defensive driving).

Turning off that anticipation would be a nightmare navigating the highways as every car is a swerve threat. Every car is a sudden deceleration threat.

These situations are extremely dangerous. Especially whene visibility is poor. Humans are terrible at it but think they are good at it. AIs will be held to amuch higher standard

They wont just assume a door will fly open. They will assume a human on the road side will open a door dangerosly and slow down to a safety speed so they can respond if a door does fly open iff humans are detected (or fail to detect no human to be more accurate). This ia already reflected in human behavior where good drivers do slow down and speed limits in many such areas in the UK are now 20MPH. Many think they are good (they anticpate well etc) and are dangerous menaces. Again this is not clever behavior but it is recognising our inability to be clever enough.

Of course as this develops the door simply wont open quickly if there's a car coming. Then the cars can go a bit faster if there's no other equal or greater danger. This is sophisticated system stuff but still very straightforward AI.

worlds-first-human-cyborg-dr-peter-scott-morgan-dead


https://nypost.com/2022/06/15/worlds...ergnet_7266138

yeah sorry chez but you are completely ignorant in what you say and due to ignorance you do not know it.


A skilled attentive human can navigate that road easily and well by being attentive and using anticipation. Perhaps you cannot and thus think 'every one struggles as I do' but you are just flat out wrong on everything you say here.

I invite you to watch an AI car today navigating in very basic dynamic traffic situations in a city, and ask why companies like Googles Waymo are so focused in areas like Scottdale Arizona where you get more golf carts than cars on the quiet roads at night.

I invite you then to watch a Nascar race and come back here and tell me the humans are the ones really terrible at dealing with dynamic situations arising in traffic flow and not the AI.

There is no chance today a self driving AI could compete in Nascar. They cannot even compete in normal city driving.

This is yet another area where I am sure you watched sci fiction and got confused thinking it was sci fact.

ok cuepee. You're response makes no sense as a response to my post except where you seem to agree with the poitn you claim is wrong

wtf? who is remotely suggesting they could. I've got them driving real slow in the easiest possible places as the first goal

I'll leave you once again. Nascar wtf?

AI autonomous driving will become safer than human driving when human senses/judgment currently necessary for driving become unnecessary.

All cars should be communicating with each other and all know where everyone is going theoretically working in unison and should greatly improve traffic. Right?

This will be a large part of it and it will include obstacles, traffic lights, shared mapping etc. Again none of this is tricky AI but it is a huge endeavour.

As with trains (which is obviosuly a lot easier) as the systems improve they will be able to travel faster and closer together while maintaining the same safety

I would not choose the word "theoretically".

There are far too many permutations to evaluate to arrive at a "theoretical" solution. One could argue that an overwhelming # of those permutations are irrelevant and can be ignored. But then we are moving away from a theoretical solution toward an "acceptable" solution.

As chezlaw hinted at, when driving becomes analogous to rail travel, then we will have arrived at AI autonomous driving. And vehicular communication "should" make that a lot easier to get to.

Assuming automobiles as independently moving and with no behavorial relationship to other vehicles was a really dumb way to develop software, IMO.

One way to improve AI autonomous driving safety (at least compared to human driving) is to replace traffic lights with frequency transmitters.

You might laugh, but doing this would not surprise me in the least.

Dual purpose to start but once the tech is compulsory there's no need for non pedestrian traffic lights at all. Everything will be done by negotiation. Pedestrians will still need a light to tell them it's their window (until we all have our tech fitted)

You are continuing to make arguments that AI is or would be better in these various real life situations if only AI drives impossibly slow and reacts as if every car door will fling open in its path as compared to a human who can easily navigate that situation today, using anticipation and awareness.

'Reaction' to a presented obstacle is great. Avoiding that obstacle before it presents thru anticipation (defensive driving suggests I move further away from that guy who looks a bit drunk the way he is swerving but staying in his lane, something AI would not see), is even better. AI can react, but it cannot yet anticipate, in any meaningful way. Thus why if you watch the real time, unedited footage of AI vehicle trying to deal with real inner city traffic issues, they struggle mightily. Stuck at 4 way stops forever. Unable to anticipate it is their turn to go and the other vehicles will yield.

BUt you keep repeating, put that anticipation away. Stop using it. And you too can sit stuck endlessly at the 4way stop where aggressive human driven vehicles can push thru.

Sorry but human drivers (at this stage) are just vastly superior, when attentive and if skilled. And you prove that by not reducing the test down to the most basic of an AI car driving in a straight line with only clearly presented obstacles. You prove that by dropping an AI car into Nascar or a busy city scape and watching it flounder and shut down unable to use its rudimentary AI to pick a path.

Or if all humans are banned from manual driving and the AI cars can all be networked together thus operating more like a disconnected train system than cars.

We are not there yet, and humans would fight being denied manual driving, but that would be the ultimate AI win, as every car would simply be programmed to keep a certain distance from the other, in varying situations, and every car would make room for the next car entering the flow of traffic. It would prevent almost all accidents outside sudden mechanical failure ones, like a blown tire.

Some people say we will never get there as 'Muh freedoms' to drive are too important. But I could see a future where Insurance Companies and not gov't mandate just make it impossible for humans to opt to drive via insurance costs being so high compared to letting your car self drive, which would see a tiny premium. You get a few forward looking cities who go full autonomous, get accident rates, drunk driving offenses, etc approaching statistical zero, and other cities start following suit. I could see it snowball in that fashion.

Or China. Once China thinks the tech is sufficiently advanced and would all but alleviate congestion issues and greatly improve safety they might just impose it and be the worlds template.

This i agree with and wrote the same before seeing your post.

It is a huge endeavor to do at scale. The amount of active live big data capture and real time crunching would exponentially crush anything we are doing today, and that says something. Quantum computing being solved for scale is probably a necessity.

Biggest challenge is that for it to be ultimately successful you need to remove the human factor off the road. You can have an entire highway of AI cars acting like a train and one human driver seeing all these big gaps they are leaving for safety reasons as a slalom course for the Human to take advantage of and tear thru (that would be me btw). Swerving in and out of the gaps, while forcing the entire line of Ai's into constant rapid breaking and swerving and recalculations to try and readjust the proper timing and speeds and distance, etc, just as the next human does the same.

Which will work 98% of the time, which is not good enough. Redundancy for power outages, weather events, etc. need to be accounted for.

Roundabouts are proven to be the safest, most efficient method of vehicular movement without need for advanced tech. But that is a topic for another conversation.

I can only assuem you are thinkinng some god like traffic controllrer

I'm thinking more Z80 level

Networking mass numbers of vehicles in any given area, capturing and sharing data (road conditions, obstacles, threats) while at the same time, simultaneously calculating all end point locations and best routes to get there, considering road conditions, obstacles, threats, and the volume of cars being sent on that route, would be what is necessary for AI in the city to be truly superior to humans driving in the city, in a way most humans would want to give up driving or be willing to surrender.


For them to truly achieve a train like function, that is the basics of the data they need to share. The most direct route from A to B is not always the fastest if every other vehicle is going the most direct route so that real time load balancing needs to be at play.

that is immense data capture, crunching and instantaneous decision making.

That said, that type of advance which likely is a quantum computing advance is still a lesser advance, generally than making fully functioning speech and human interaction AI, where we have no such visibility path to achieving as of yet. The belief being we need a novel break thru in AI to get there as none of our current paths are generally believed to be sufficient as they once were prior.

Centralised nwtworking is one way to go but sounds like avery bad idea

Simple communication prototocols in a vast distributed network of automonous agents is a far superior, and probaly the only viable, approach.

The train comparison works for multiple cars on the road, but when there is only one car driving, then the autonomous vehicle needs to shoulder more of the load for safe navigation.

And of course if the lead vehicle drives off a cliff the other cars are liable to follow like lemmings. So simply applying the train concept without redundant error correction does not solve the matter.

But guidance will need to be provided at multiple levels to satisfy the necessary redundancies for safe travel.

Ya i think a mix of train like function for dense traffic and more autonomous for open roads will be key.

That said, no, in train like connection no other vehicle would be likely to follow the first off the cliff as the point is they are always sharing that data and when a hazard is identified all instantly get that data to now utilize in re-routing.

It would be the more autonomous cars all more likely to go off the same cliff, if for instance a sink hole opens on a popular route and they all route that way because it is quickest without any data about this sudden hazard.