How Hot is Ukraine Gonna Get?

[ArmouredTopHat]

32 minutes ago, TheVulture said:

Second-hand quoting of a journalist on what the UK has announced as additional support today:

Upon taking office the Defence Secretary immediately asked for extra support to be provided to Ukraine which was readily available and meets their needs for the battlefield against Russia. This new package includes:

• A quarter of a million of 50 calibre ammunition

• 90 anti-armour Brimstone missiles

• 50 small military boats to support river and coastal operations

• 40 de-mining vehicles

• 10 AS-90 artillery guns

• 61 bulldozers to help build defensive positions

• Support for previously gifted AS-90s, including 32 new barrels and critical spares which will help Ukraine fire another 60,000 155mm rounds

John Healey also directed officials to ensure that the promised package in April of military aid is accelerated and delivered in full to Ukraine within the next 100 days.

Not seen any official confirmation of details yet.

Although this is a new Labour government replacing the previous Conservative government, so its always possible there are some political games going on of announcing as new stuff that was already promised but not yet delivered. But at the least the new government is signalling that it is going to continue, and possibly increase, UK support for Ukraine.

 

(Note for confused Americans: Brits use 'government' in the same way that Americans use 'administration' - so read that as 'administration' in the above)

The bulldozers are a fantastic addition that will go a long way. Long term support for the AS-90s is also great to see. 

[The_Capt]

2 hours ago, poesel said:

Yes - sitting in a forest is probably the worst if you are attacked by a drone swarm (even if you had a defending swarm of your own). OTOH a desert is not drone country...

If we take Ukraine as test field, I guess it would be okish for the defender. Large open space between tree plants have not much cover for drones. But terrain is a bitch - we all know that.

A defender will need to think different about what angles of attack he needs to cover against drones in the future.

Man that sounds like the bones of a great game.

[The_Capt]

2 hours ago, poesel said:

I haven't read everything, but that would be a really strange argument to make. Offensive drones already exist in the thousands.

I haven't seen anything against kamikaze drones except infantry shooting their rifles.

Your lips to God’s ear.  Not my argument but here we are.

[chrisl]

1 hour ago, Butschi said:

Technically, I was talking about the attention mechanism in transformer models. You can use a lot of other architectures in order to predict a word as a function of a bunch of other words that don't have this feature. They also have worse performance, of course.

Excuse me but that sounds like a really big claim to make, given that a) there a papers out there that research reasoning capabilities in depth and come to the conclusion that LLMs have at least basic capabilities, b) many in this field use LLMs precisely because of their reasoning capabilities and c) you can actually ask e.g. ChatGPT to reason how it arrived at an answer. Can I kindly ask you to back up your claim?

Sure, LLMs are not AGI and their answers and reasoning are often flawed. That doesn't imply they can't do any reasoning at all, though.

If I ask it to explain how it came up with the answer does it:

a) provide me with a breakdown of what chunks of information it processed (and how) and what the sources are for what it answered  or

b) Generate an explanation the same way it generated the answer, like, to put it a little rudely, an expensive mansplainer?

I haven't seen anything like (a) provided. I have seen (b).  When I pay ML people to put together a "remote graduate student" it gives me both the summary of what it thinks is interesting and the actual details of parameters it used to make the decisions. All referenced to actual data and physical relations.

When people apply what amount to modernized Turing tests they interpolate well and blow up at the edges.  Like a massive multiparameter fit, or string theory.

some refs:

https://arxiv.org/abs/2304.03439

https://arxiv.org/abs/2406.15992

https://www.sciencedirect.com/science/article/abs/pii/S000437022400081X

There's no model - it's a giant parametric fit.  Impressive, but not nearly as smart as often claimed.  

LLMs suck (and I mean suck) at physics problems.  They might give you a correct answer to a problem that's been put up in words a million times on the internet, but ask it one that's easily described in words, and commonly solved in classes, but not a phys 101/102/201/202 problem and they go into neverland. 

There's so little model behind them that they don't have the concept of references being a static thing.  They've been caught making them up, possibly most famously in a legal brief:

https://arstechnica.com/tech-policy/2023/06/lawyers-have-real-bad-day-in-court-after-citing-fake-cases-made-up-by-chatgpt/

https://hai.stanford.edu/news/hallucinating-law-legal-mistakes-large-language-models-are-pervasive

[Letter from Prague]

To be honest, LLMs kinda suck in general at anything that is not text manipulation. Even knowledge retrieval is bad, because of hallucinations.

Anyway, new Perun dropped:

 

[Vanir Ausf B]

3 hours ago, poesel said:

Good microphone + good filter and you can detect a quadcopter. I guess if something would exist that can shoot drones down, we would already see a microphone network in Ukraine.

It already exists in Ukraine and apparently works well-enough that NATO is planning their own versions.

https://www.twz.com/land/thousands-of-networked-microphones-are-tracking-drones-in-ukraine

[Probus]

2 minutes ago, Vanir Ausf B said:

It already exists in Ukraine and apparently works well-enough that NATO is planning their own versions.

https://www.twz.com/land/thousands-of-networked-microphones-are-tracking-drones-in-ukraine

Never would have thought it.  That's why we need engineers to think outside the box.

[chrisl]

9 minutes ago, Vanir Ausf B said:

It already exists in Ukraine and apparently works well-enough that NATO is planning their own versions.

https://www.twz.com/land/thousands-of-networked-microphones-are-tracking-drones-in-ukraine

It works well because Shaheds and Orlans are powered by lawnmower engines.  You can hear those from miles away.

It will work for quadcopters, but not as well because the amount of noise energy they produce is much, much smaller.

ETA for Probus: It's an old concept.  It was used in a much less fancy way to detect bomber swarms in WWII: https://rarehistoricalphotos.com/aircraft-detection-radar-1917-1940/

It's also one of the methods of CB target location prior to shot radar.  Modern police departments have also used it (shotspotter).  ADS-B Exchange, FR24, and Flightaware use the RF equivalent for locating aircraft that are only transmitting a hex code and no identifying information or location/speed data.

Edited July 7 by chrisl

[Battlefront.com]

3 hours ago, poesel said:

You wrote about a coordinated attack. For coordination, you either need a third party observer, communication between the drones or passive observation between drones.

In the first case, the observer is at a distance and needs to communicate with the drones. That makes himself a high priority target, and the distance diminishes the value of the observation.

In the second case, the drones need to emit something (radio, IR, acoustic) which makes them detectable in principle. But I guess whoever designs this will make it low noise enough, so that it can be used for detection but not targeting.

The third case is hard because the drone can only use passive sensors to detect others in the same mission. Since the drones are designed to be hard to detect - well... I guess possible, but a few years down the road. That would be real ninja drone swarms.

Already partially address by kimbosbread, but I'll add something to it.  There's a couple of other methodologies that mean no emissions.

The first is to choose a target ahead of time, feed the target coordinates and profile into a computer, have it spit out individual paths for X number of drones, upload those instructions into the drones while they are on the ground (i.e. using a hardwire connection), then send them on their way.  The coordination is done ahead of time, therefore zero need for the drones to communicate with each other.

The second is chaos theory at work.  You feed the target coordinates and profile into your drone swarm and each one chooses its own path to the target.  The targeting algorithms are such that it's highly unlikely that any two drones will take even close to the same path.

Third option is to have all the drones already in flight without any specific target.  One ISR drone or a ground station bursts one blink of an eye transmission to issue coordinates or a slightly longer one to issue paths.  The drones are receiving only therefore they are undetected, though the source of the transmission will be visible for the milisecond or second it took to complete the transmission.  This is susceptible to EW, but not very effectively since you will know what bands are compromised before the transmission and simply select a band that isn't jammed.  In any case, the drones aren't emitting.

Some of these techniques are more useful for attacking stationary targets than moving ones, but all are theoretically effective at doing either.

3 hours ago, poesel said:

If I had a bunch of networked vehicles, then each would know where everyone else is at any time and not shoot in their direction.

This means that the enemy can know where all of your vehicles are in realtime.  That's a problem that goes beyond drones and counter measures.  However, if that information is available to the friendly units it can be leveraged as a basis for attack plans for drones.  There are so many ways to "outsmart" C-UAS if you know where the point defenses are.

The other thing is this doesn't help protect dismounts, civilians, or infrastructure.  So this isn't even a theoretical solution to the problem.  And then there's...

3 hours ago, poesel said:

We are talking about small calibers here. So even if I had to shoot my shotgun equivalent in the direction of something Bradley-ish I would do it. Anything I could do to the Brad is less worse than what the drone does.
And if theres my own infantry directly behind the drone coming at me I would still shoot. That is one CAS vs 3+ in the vehicle. War is no Ponyhof

If your SOP is to fire no matter what, then I'm going to have fun sending my UAS at you when you're passing through civilian areas, friendly bases, or whenever you're in tight formations with things like trucks.

3 hours ago, poesel said:

No, he doesn't. Everything I named so far is passive. Being a heavy emitter makes targeting easier. So PD may need to switch on its radar when the attack has been detected and is imminent. But then you are not giving much away since the enemy already knows you are there

 

Switching on radar means something other than a UAS can come in and spoil the party.  Or even more deadly, a UAS that has a stand off emission seeking missile.  Have one of those standing by to jump into action when you start your UAS attack on something like a column and all that fancy C-UAS becomes useless.

3 hours ago, poesel said:

I would like to ask you to stop telling other people that they are just wrong. I perceive this as rude, even if you put a smiley to it

When someone says something that is wrong it is not rude to point that out.  It's kinda a fundamental part of a discussion.  I expect that if I say something that is verifiably wrong that someone would make mention of it, smiley face or not.

3 hours ago, poesel said:

You did not follow my example calculations. Depending on the speed of the bullet and the amount of g's the target can pull, there is a distance where the target just doesn't have enough time to move away enough from the current, predictable path.
Enlarging the kill area with a shrapnel weapon (like a shotgun shell) makes hit probability even larger.

Airburst needs at least a 20mm shell IIRC and that is IMHO too big for this application.

I understood your example, but it is flawed because you presume a predictable path.  The point of me posting that video is there is no predictable path necessary for a any UAS, manned or autonomous.  However, there are practical limitations for path variation if Human piloted, none exist with autonomous.

So the round gets fired at what, exactly?

This is why I said that danger close defense is the only viable way to knock down a UAS because, at some point, it has to move in a way that is predictable enough to be hit by some sort of airburst munition.

Steve

[chrisl]

22 minutes ago, Battlefront.com said:

 

I understood your example, but it is flawed because you presume a predictable path.  The point of me posting that video is there is no predictable path necessary for a any UAS, manned or autonomous.  However, there are practical limitations for path variation if Human piloted, none exist with autonomous.

So the round gets fired at what, exactly?

This is why I said that danger close defense is the only viable way to knock down a UAS because, at some point, it has to move in a way that is predictable enough to be hit by some sort of airburst munition.

 

His example was reasonable - it didn't assume predictable path so much as that newton's laws are still valid at the drone scale.  If a drone has a particular momentum at point (x,y,z,t) and maximum acceleration of a_max, there's a well defined limited volume available for it to occupy at some later point (x1,y1,z1,t1).  You just have to fill that volume of spacetime with stuff at a density high enough that at least one of piece of your shrapnel is guaranteed to intesect the volume occupied by the drone.

You're also correct that when it's danger close the drone can (and will) set you up to have some of your own guys in that second volume.  But he's merciless and will save the three guys in the PD vehicle over ex-PFC Conscriptovich standing outside having a smoke.

Edited July 7 by chrisl

[Vanir Ausf B]

34 minutes ago, chrisl said:

It works well because Shaheds and Orlans are powered by lawnmower engines.  You can hear those from miles away.

It will work for quadcopters, but not as well because they amount of noise energy they produce is much, much smaller.

True. The microphones will need to be near the units they are specifically tasked with protecting. They may be integrated with the unit itself.

Fortunately, quadcopters are relatively slow-moving objects. A Mavic tops out at 40 mph/65 kph in a straight line with no headwind.

 

[poesel]

58 minutes ago, Battlefront.com said:

Already partially address by kimbosbread, but I'll add something to it. 

...

You repeat kimbosbreads arguments here. And I wrote him a PM to not drag this one out but no luck.

The fun part is that we all agree and say the same thing using different words.

TL;DR: you can go totally silent with some (probably small) loss of coordination. And even if you communicate, it can be made very hard to detect and nearly impossible to track.

 

 

 

 

58 minutes ago, Battlefront.com said:

This means that the enemy can know where all of your vehicles are in realtime.  That's a problem that goes beyond drones and counter measures.  However, if that information is available to the friendly units it can be leveraged as a basis for attack plans for drones.  There are so many ways to "outsmart" C-UAS if you know where the point defenses are.

Why would the enemy know where my vehicles are from the information my vehicles share with each other? See above for drone to drone communication. Same problem, same fix.

 

 

 

58 minutes ago, Battlefront.com said:

If your SOP is to fire no matter what, then I'm going to have fun sending my UAS at you when you're passing through civilian areas, friendly bases, or whenever you're in tight formations with things like trucks.

Yes, there are places where PD does not work. Try to avoid them.

A friendly base that cannot defend itself versus drones is not a base.

Trucks in tight formation is a dumb idea for other reasons, too.

 

58 minutes ago, Battlefront.com said:

Switching on radar means something other than a UAS can come in and spoil the party.  Or even more deadly, a UAS that has a stand off emission seeking missile.  Have one of those standing by to jump into action when you start your UAS attack on something like a column and all that fancy C-UAS becomes useless.

If you have something better or cheaper than a drone swarm, you use that, of course.

I assume our common assumption is that the defender has been detected, and the attack will be done by a bunch of drones.
If the defender is supposedly undetected, then we talk about a free roaming drone swarms and that is another can of worms.
If you don't attack with drones and use e.g. artillery, then the PD we talk about here has no use at all.

 

58 minutes ago, Battlefront.com said:

When someone says something that is wrong it is not rude to point that out.  It's kinda a fundamental part of a discussion.  I expect that if I say something that is verifiably wrong that someone would make mention of it, smiley face or not.

Well, maybe be lost in translation. I just told you, how I understood it.

 

58 minutes ago, Battlefront.com said:

I understood your example, but it is flawed because you presume a predictable path.  The point of me posting that video is there is no predictable path necessary for a any UAS, manned or autonomous.  However, there are practical limitations for path variation if Human piloted, none exist with autonomous.

So the round gets fired at what, exactly?

This is why I said that danger close defense is the only viable way to knock down a UAS because, at some point, it has to move in a way that is predictable enough to be hit by some sort of airburst munition.

chrisl already explained the upper part nicely.

I don't know how far danger close is for you, but from my ballpark calculations, I guess we are talking about 200m out.

Edited July 7 by poesel
typo

[hcrof]

44 minutes ago, Battlefront.com said:

The first is to choose a target ahead of time, feed the target coordinates and profile into a computer, have it spit out individual paths for X number of drones, upload those instructions into the drones while they are on the ground (i.e. using a hardwire connection), then send them on their way.  The coordination is done ahead of time, therefore zero need for the drones to communicate with each other.

The second is chaos theory at work.  You feed the target coordinates and profile into your drone swarm and each one chooses its own path to the target.  The targeting algorithms are such that it's highly unlikely that any two drones will take even close to the same path.

Third option is to have all the drones already in flight without any specific target.  One ISR drone or a ground station bursts one blink of an eye transmission to issue coordinates or a slightly longer one to issue paths.  The drones are receiving only therefore they are undetected, though the source of the transmission will be visible for the milisecond or second it took to complete the transmission.  This is susceptible to EW, but not very effectively since you will know what bands are compromised before the transmission and simply select a band that isn't jammed.  In any case, the drones aren't emitting.

Some of these techniques are more useful for attacking stationary targets than moving ones, but all are theoretically effective at doing either.

This means that the enemy can know where all of your vehicles are in realtime.  That's a problem that goes beyond drones and counter measures.  However, if that information is available to the friendly units it can be leveraged as a basis for attack plans for drones.  There are so many ways to "outsmart" C-UAS if you know where the point defenses are.

The other thing is this doesn't help protect dismounts, civilians, or infrastructure.  So this isn't even a theoretical solution to the problem.  And then there's...

If your SOP is to fire no matter what, then I'm going to have fun sending my UAS at you when you're passing through civilian areas, friendly bases, or whenever you're in tight formations with things like trucks.

Switching on radar means something other than a UAS can come in and spoil the party.  Or even more deadly, a UAS that has a stand off emission seeking missile.  Have one of those standing by to jump into action when you start your UAS attack on something like a column and all that fancy C-UAS becomes useless.

When someone says something that is wrong it is not rude to point that out.  It's kinda a fundamental part of a discussion.  I expect that if I say something that is verifiably wrong that someone would make mention of it, smiley face or not.

I understood your example, but it is flawed because you presume a predictable path.  The point of me posting that video is there is no predictable path necessary for a any UAS, manned or autonomous.  However, there are practical limitations for path variation if Human piloted, none exist with autonomous.

So the round gets fired at what, exactly?

This is why I said that danger close defense is the only viable way to knock down a UAS because, at some point, it has to move in a way that is predictable enough to be hit by some sort of airburst munition.

Steve

A few corrections Steve:

Preprograming drone paths means you cant hit moving targets. If you are attacking a target 5km away (quite short range), then your target might be 1-2km away by the time the drone arrives. Unless you are attacking a largely static target you have to guide it in. Allowing your drone to hop frequencies dynamically means it needs a milspec radio which is likely going to push your price and weight up a lot. If the drone has a fixed frequency then there is a chance of jamming or spoofing it, or just sending out defensive drones. I am no EW guy so I don't know how hard that is to do with future AI-based EW. 

On the other hand, target vehicles will have milspec radios that can transmit at very low power so intra-unit comms will generally fade into the background noise of the battlefield. 

On radars, an AESA radar can be detected if it is emitting but needs a pretty sophisticated detector since it is frequency hopping. A cheap drone will never find it. 

And AI piloted racing drones look fancy to a human eye but I would encourage you to look at the hard maths of what they can do when you assume they are carrying a heavy battery and warhead so have much slower accelerations. Doing an attack run limits how much they can manoeuvre even more since they are aiming at a target (which might be very small if they need to hit a weak spot). A large shotgun can absolutely take them out if aimed by AI, but only at very close range: over 100m or so you need airburst munitions or a defensive drone. 

[Sojourner]

Interesting report on Russia's latest tactics for taking ground, from CBS News featuring interview with George Barros from ISW.

TL/DR: Best defense against FPV drones = meat.

https://www.cbsnews.com/video/russias-latest-tactics-to-seize-ukrainian-territory/

[chrisl]

15 minutes ago, poesel said:

Why would the enemy know where my vehicles are from the information my vehicles share with each other? See above for drone to drone communication. Same problem, same fix.

I know where your vehicles are because I can see them from space.

By about the second paragraph of your reply on p3558 I came to the conclusion that the best solution is to go underground and become mole people. Though that comes with a whole different set of problems for control of uncrewed underground vehicles and some unpleasant constraints on movement paths and/or speed of movement.  

[Letter from Prague]

40 minutes ago, hcrof said:

A few corrections Steve:

Preprograming drone paths means you cant hit moving targets. If you are attacking a target 5km away (quite short range), then your target might be 1-2km away by the time the drone arrives. Unless you are attacking a largely static target you have to guide it in. Allowing your drone to hop frequencies dynamically means it needs a milspec radio which is likely going to push your price and weight up a lot. If the drone has a fixed frequency then there is a chance of jamming or spoofing it, or just sending out defensive drones. I am no EW guy so I don't know how hard that is to do with future AI-based EW. 

You can do a hybrid of preprogramming a search-path, and then having the drone autonomously attack anything that looks like a tank. That would make slightly outdated information ok - assuming the recognition accuracy works out.

[The_Capt]

23 minutes ago, chrisl said:

I know where your vehicles are because I can see them from space.

By about the second paragraph of your reply on p3558 I came to the conclusion that the best solution is to go underground and become mole people. Though that comes with a whole different set of problems for control of uncrewed underground vehicles and some unpleasant constraints on movement paths and/or speed of movement.  

Kind of my thinking. I mean are we really trying to hide 30-60 ton vehicles that run on small explosions? If we can do that then a small bird sized UAS is basically invisible. So this really is less about individual drones v a platoon of vehicles, and more about battle space management.  That platoon is going to get picked up well back, and if it somehow manages to stay hidden its logistics tail will definitely give it away.  Once detected, there is a suite of systems to go hunt and kill them from good old artillery of all shapes and sizes, to NLOS ATGMs, to drones, to UGVs, to other vehicles, to mines.

So it is great if you have some sort of wireless and silent comms system in your platoon…but the autonomous drones will be sent right to your grid square along with a bunch of other stuff.  After thinking on this, I am not sure passive or active will really matter.  Once detected they can and will be engaged. Then even if they managed to hold off the first swarm wave, all those guns blasting off like a Serbian wedding is damn sure to give up their position.  And in come the PGM and cluster munitions.

A foundational element of land warfare is an ability to seek cover from observation and fires.  Once that is gone, land warfare theory is pretty much dead.  At that point we need a new theory, probably closer to maritime warfare.  And before someone pipes up that there is PD in maritime warfare, well 1) we have no evidence that our current theories of war on the sea are still effective and 2) because land is still…well land…the dynamics are fundamentally different.  We can disaggregate on land far more than at sea - so I think there are adjustments to be made.

So here we are…still.

[Butschi]

2 hours ago, chrisl said:

If I ask it to explain how it came up with the answer does it:

a) provide me with a breakdown of what chunks of information it processed (and how) and what the sources are for what it answered  or

b) Generate an explanation the same way it generated the answer, like, to put it a little rudely, an expensive mansplainer?

I haven't seen anything like (a) provided. I have seen (b).  When I pay ML people to put together a "remote graduate student" it gives me both the summary of what it thinks is interesting and the actual details of parameters it used to make the decisions. All referenced to actual data and physical relations.

When people apply what amount to modernized Turing tests they interpolate well and blow up at the edges.  Like a massive multiparameter fit, or string theory.

some refs:

https://arxiv.org/abs/2304.03439

https://arxiv.org/abs/2406.15992

https://www.sciencedirect.com/science/article/abs/pii/S000437022400081X

There's no model - it's a giant parametric fit.  Impressive, but not nearly as smart as often claimed.  

LLMs suck (and I mean suck) at physics problems.  They might give you a correct answer to a problem that's been put up in words a million times on the internet, but ask it one that's easily described in words, and commonly solved in classes, but not a phys 101/102/201/202 problem and they go into neverland. 

There's so little model behind them that they don't have the concept of references being a static thing.  They've been caught making them up, possibly most famously in a legal brief:

https://arstechnica.com/tech-policy/2023/06/lawyers-have-real-bad-day-in-court-after-citing-fake-cases-made-up-by-chatgpt/

https://hai.stanford.edu/news/hallucinating-law-legal-mistakes-large-language-models-are-pervasive

Aren't you shifting the goal posts now? You claimed that "there really isn't any reasoning behind the answers" of LLMs. Providing sources, etc. is not to really a part of reasoning, especially for common knowledge. If I ask you to tell me how fast a spacecraft goes after a given time with speed v and acceleration a, I don't expect you to tell me what textbook you had at school.

In fact I just asked ChatGPT this exact question (without the textbook part):

Quote

There is a spacecraft in vacuum. It travels at 10 meters per second.  How far does it travel in five seconds? Please also give me the reasoning behind your answer.

To determine how far the spacecraft travels in five seconds, we can use the basic formula for distance traveled at constant speed:

\[ \text{Distance} = \text{Speed} \times \text{Time} \]

Given:
- Speed of the spacecraft (\(v\)) = 10 meters per second (m/s)
- Time (\(t\)) = 5 seconds

We can plug these values into the formula:

\[ \text{Distance} = 10 \, \text{m/s} \times 5 \, \text{s} \]

\[ \text{Distance} = 50 \, \text{meters} \]

So, the spacecraft travels 50 meters in five seconds.

**Reasoning:**
The spacecraft is moving at a constant speed in a vacuum, where there is no air resistance or other forces acting on it (assuming we are ignoring gravity and any other potential forces for simplicity). Therefore, the distance it covers can be directly calculated by multiplying its constant speed by the time for which it travels.

Now, let's assume that it starts at 10 meters per second but I'm addition accelerates with 2 meters per second squared. How far is it now after 5 seconds? Please, again, give me the reasoning for your answer.

To determine how far the spacecraft travels in five seconds with an initial speed and constant acceleration, we use the kinematic equation for distance traveled under uniform acceleration:

\[ \text{Distance} = v_0 t + \frac{1}{2} a t^2 \]

where:
- \( v_0 \) is the initial speed (10 meters per second)
- \( a \) is the acceleration (2 meters per second squared)
- \( t \) is the time (5 seconds)

First, we calculate the distance traveled due to the initial speed:

\[ v_0 t = 10 \, \text{m/s} \times 5 \, \text{s} = 50 \, \text{meters} \]

Next, we calculate the distance traveled due to the acceleration:

\[ \frac{1}{2} a t^2 = \frac{1}{2} \times 2 \, \text{m/s}^2 \times (5 \, \text{s})^2 \]

\[ \frac{1}{2} \times 2 \, \text{m/s}^2 \times 25 \, \text{s}^2 = 25 \, \text{meters} \]

Now, we add these two distances together to get the total distance traveled:

\[ \text{Total Distance} = 50 \, \text{meters} + 25 \, \text{meters} = 75 \, \text{meters} \]

So, the spacecraft travels 75 meters in five seconds with the given initial speed and acceleration.

**Reasoning:**
The total distance traveled by an object under constant acceleration can be found by considering both the distance traveled due to its initial speed and the distance added by the acceleration. The initial speed contributes a linear distance term (\( v_0 t \)), and the acceleration contributes a quadratic distance term (\( \frac{1}{2} a t^2 \)). By summing these contributions, we obtain the total distance traveled over the specified time period.

Granted, this is a simple task but that isn't the point. The reasoning is perfectly ok, as far as I am concerned. I also didn't see any mansplaning or whatever. But even that would be beside the point. You claimed that these models are incapable of any reasoning.

None of the papers you provided (interesting ones, really), actually concluded that LLMs are incapable of reasoning, just that their performance for abstract problems or ones not in included in the training data are often not very good. But frankly the same can be argued for humans. Back at school the transferring knowledge to a new problem was what we needed to get an A not to pass the test ...

I'm sorry but you didn't convince me here.

As to

Quote

There's no model - it's a giant parametric fit.  Impressive, but not nearly as smart as often claimed.  

You are right, this whole machine learning business is basically an increasingly sophisticated way of fitting parametric functions to data. But "human leaning" is really not very different, right? Neurons and synapses work a little different than neural networks in ML but the principles (leaning weights for connections, i e. synapses, between neurons) is very similar.

Now, we humans, although our brains are also just parametric fits, are capable of forming models of how our world works. Often just heuristics but even learning algebraic formulae.

As you said yourself explainability is a thing in ML. Those LLMs are huge, with billions or even trillions of parameters. I think right now they are more or less black boxes. That goes both ways, though. If we can't tell how and what knowledge is stored where, we can't really say what isn't in there.

So, I didn't want to spread the gospel of AI, I've argued myself at some points that AI is not as far as some think. On the other hand I'd really like to stress that LLMs, with all their flaws, are a huge leap in capabilities compared to what we had just a few years back - so "LLMs suck" is really something I can't agree to.

Ok but I guess Steve will very soon tell us to no further derail this thread.

[dan/california]

Emitting is tending towards being an ever worse idea

Quote

 

https://www.popularmechanics.com/military/research/a24177291/oops-civilian-satellite-data-inadvertently-pinpoints-military-radars/

Oops! Civilian Satellite Data Inadvertently Pinpoints Military Radars

 

Now this article is about very powerful air defense radars. But the satellite that picked it up wasn't even trying. I suspect you would gain a lot of sensitivity if you designed a system with this trick in mind.

[Fenris]

Clip showing drone operator view of EW battle in action

 

[cesmonkey]

 

[dan/california]

 

[dan/california]

 

[dan/california]

 

[dan/california]