There are endless posts on HN about how Macs are good for LLMs. Somoene is running some model on their Mac for local inference.
What the posts dont mention is how unusable that experience is with dogshit slow tokens/second. To make a local model usefull you need to run the highest parameter models at 100+ tok/sec, otherwise you are just better off paying for cloud inference. So either all those people are dumb as hell, or Apple is doing clever advertising. And I personally have more faith in the tech sector.
Back in the 2010s, the best, most efficient software engineers were characterized by 3 things
* end to end rough map of the entire space of compute in their head
* ability to search the internet for the right things
* ability to quickly experiment and try things to figure out how to do things
AI hasn't change that, it just made 2 and 3 into a very efficient thing.
The characterization of psychosis is best described by believing AI can do the first thing. No modern LLM can "reason" - otherwise you could give it a task like "make me money", it would ask you all the questions it needs about information that it doesn't know about and needs to know to make you money, then it would set whatever it needs to set up to make you money.
As such, you still need to know the domain entirely to be effective. When you do that, AI is fantastic at getting you to the right solution. Furthermore, its still in large part actually cheaper to higher a developer who then can use AI to build you the product that you need long term.
> No modern LLM can "reason" - otherwise you could give it a task like "make me money", it would ask you all the questions it needs about information that it doesn't know about and needs to know to make you money, then it would set whatever it needs to set up to make you money.
If that's your bar for reasoning, then most people can't reason either.
The thing about these math proofs is that they really add nothing of value. Its a cool tech demonstration on how LLMS can search the trained space, but fundamentally, LLMS haven't "discovered" anything groundbreaking.
For NS equations, they are trying to model something that is discrete (i.e molecules colliding) in a continuous manner. You can easily think of a condition where they fail - imagine a vaccum where there is sufficient space between air molecules, so that collisions aren't always possible. NS won't be able to predict the state of the fluid in every single point in space.
In practice, when you do CFD, no package uses direct differential simulation of NS equations, you usually have simpler approximations that are good enough for the space you are working for. And if you want accuracy, you usually do something like LBM which simulates particle collisions using probability distributions.
I can't wait to tell my pure maths professors that their most of their research adds nothing of value. I mean I am sure most of them would agree to some extent, but like, dude, have some more faith in the utility of pure maths, esp. centuries down the line
There is plenty of work that goes on in the math field that is super important. But any field has research that is irrelevant. So far, all the proofs done by ai have been largely irrelevant.
sure I get that, but like, my field has plenty of counterexample as proofs. we have had non-constructive proofs like probabilistic arguments. i don't think we can play the game of "oh this proof is useful that proof is not useful" well
Generally, the proofs that are in the form of "here is a single contradiction to an established statement that proves that its not always true", are generally useless. We can prove that newtons laws don't apply when you start considering relativity, but because they still apply for a large domain, they are still used.
Same with NS equations. Who cares if you can find a singularity.
And if you want an example of something novel that is worth pursuing - Its highly likely that the modern transformer architecture is sub optimal, you probably don't need to do full matrix multiplies in the transformers. There potentially could be a higher level mathematical formulation of minimal math operations that are needed without having to do trial and error - especially because all of the math involves linear combination passed through smooth activation functions.
But coincidentally, there hasn't been any research in terms of point LLMS to self optimize in this way, because there isn't enough human math literature on the LLMs to train on.
i really think we are opening a can of worms with these “who cares if you find a single counter example as disproof” arguments. i think the better version is “ok any lemmas or techniques we can generalize from this” or “what did we learn about maths through this” and use this as a basis to say LLM proofs are not useful
like say if god lets me find a single counter example to P=NP and thus disproving it — I think we can learn tons about complexity theory from this counter example by studying it. we should not have the hubris of assuming “oh a single counterexample is generally useless” — why, how. this is the same hubris imo that produced like “number theory is useless” until it is not
P vs NP is a more fundamental problem that if proven, will have insane consequences, perhaps more than anything else out there. For starters, you would be insantly able to design an an actuall all knowing AGI.
The NS equations are far,far,far less meaningful. Like I mentioned earlier, if you actually want accurate CFD, you dont even use them.
i am not a complexity theorist but I am a CS academic by training (I never was a good one, but welp), and during my PhD it is often said that maybe P vs. NP an initial proof/disproof to the statement is not that practically important, e.g., if P=NP, maybe the NP -> P reduced algorithm is very very cosmic. P=NP by itself hardly proves that one would instantly design an AGI whatsoever. Often the downstream potential theoretical/practical insights/results seems more exciting;
> The NS equations are far,far,far less meaningful. Like I mentioned earlier, if you actually want accurate CFD, you dont even use them.
Sure. Consider this: in algorithm research often the most optimal algorithm in big-O is not the one used IRL; examples are numerous: matrix multiplication, LCA data structures, many variants of shortest paths.
An academic can work two years on faster-in-theory matrix multiplication that no one expects to be used in practice (in our currently imaginable univese). Do you consider that less meaningful than working on faster matmul kernels?
Essentially, every task that AGI can do is the same as the traveling salesman problem - you have some system with dynamics where you can make decisions and to get to the right outcome, you are trying to find an optimal path. The best it can do at this moment is to create parallel simulations (i.e AlphaZero/MuZero MCTS - the tree represents the dynamics and evolution of the game board and it effectively does a simulation of the game prior to making a move). To do this with something as complex as the world, including all the humans within it, most likely means that you have to have more compute than reality allows - otherwise you are taking a hit in accuracy, which given chaotic effects creates problems where you can be way off on the predictions.
On the flip side if P=NP, that means that instead of dedicating compute to running branching simulations, An AGI can dedicate compute to just solving directly the actions it needs to do for any given outcome. This is a shortcut to reality, which means that reality in itself is compressible.
There still has to be some logic behind what is the meaning of a proof. Generally, its pointless to explore random equations of numbers and try to prove that this equation holds for every value. That equation needs to have some use, whether its cryptography, or description of a physical process.
For example, look at Poincare conjecture proof. As cool as it is, can you name one area where the derivation of that proof or the proof itself has been used (without asking an LLM)?. Note that the core concept, Ricci flow, is used in lots of places, but the application of the proof is largely irrelevant - the homeomorphism of any 3d shape (say like a surface in Blender) to a sphere can be determined in other ways, more efficiently than what Pointcare conjecture states (i.e that every loop can be tightened to a point).
The problem is, people who are the most spoken about personal freedom are the ones also voting conservative all the time. The libertarian square is still on the right.
By the time the government overreach starts being used on them, its gonna be too late.
Electric planes can in theory get longer flights, provided they change the flight profile - you spend a lot of energy going up to higher altitude, and the essentially glide down.
The energy used to achieve altitude is then saved in the descent stage. So that’s not an issue for aircraft, regardless of how they’re powered. This is why every flight you’ve ever been on climbs to cruise altitude as fast as possible.
High altitude where the air is less dense is better for speed due to less drag. This is also power source agnostic.
yes, but current battery energy density is nowhere near fuel ATM, and evs operate differently. Gas engines use fuel/hour as a function of RPMS, independant on the load on the engine. EVs use battery juice as a function of the load.
When you size a gas engine for a plane, you design around the power needed to takeoff on a given runway, which is fine because in cruise, you will throttle down and use less fuel.
When you design an electric vehicle, if you design it the same way, you end up carrying effectively dead battery weight, which also affects power consumption during cruise (you either go faster, or use use more wing, which creates more induced drag).
IF you design it for cruise, you end up being horribly inefficient at takeoff.
The only 2 solutions are
1. 10 mile long runway where you can take your time accelerating, to reduce the torque requirement on the motor. You would basically limit the current draw and the plane would accelerate slowly, and eventually you get to cruise speed
2. Changing your flight profile, where you get to high altitude ASAP where you can be more efficient,
Batteries aren't rapidly improving. Every single battery that has proposed higher energy density also has downsides in max discharge current. We may get at best like 10% improvement over what it is today.
If you are consuming X watts per cell, then each cell will drain by the same amount of current regardless of whether the cells are in series or parallel (or more likely a combination of the two).
This is supposed to be able to handle "up to" 2-hour flights, and (for jets at least) takeoff power is about 3x cruise power, so about 1.5 watts per watt-hour of battery, or a 1.5C peak discharge rate. What chemistries are you thinking of that cannot handle this?
The full power system for a prop looks like this: you have a battery of a specific voltage, which then runs a motor, which then runs a gearbox, which then turns a prop.
The motor and the gearbox can be considered as one unit - an electric motor has two factors, KV(RPM/volt) and KT(torque/amp). The higher the KV, the lower the KT. A high KV motor spins fast, but draws a lot of current for the same torque - putting it through a reduction gearbox turns it into a low KV, high KT motor. Naturally, low KV motors or (low KV setups) are more efficient because they draw less current for a given torque, and the heating power loss varies with current^2.
The prop needs to spin at certain RPM for max aerodynamic efficiency. Given the slider for motor/gearbox selection between high KV/low KT and the opposites, you generally want to have as high voltage as possible, so that you can run a low KV/high KT setup, which means that for the given torque, the current is minimal. I.e you have a motor spinning really fast, through a large reduction gear, driving a prop at the necessary speed and torque without much load on the motor.
So lets say you determine that you want a certain voltage, which requires a stack of cells in series. The only way to get more capacity is to duplicate that stack and put them in parallel. So your weight becomes quantized by the number of stacks you have in parallel. And the more stacks you have in series, the higher the weight jumps between parallel stacks counts.
Subtracting cells from stacks doesn't work well. Lets say you have a single stack of 10 cells 10s1p. If you do something like 8s1p, you lower the output voltage, which means you need to have slightly higher gear ratio to spin the prop at the same efficient rpm, which means you draw more current, which means the extra capacity in the cells doesn't really matter if you are drawing more current.
If your motors draw W Watts at peak, and you have N cells at V volts then the peak per-cell current will be (approximately -- there are internal losses) W/(NV) regardless of the geometry of your stacks.
If N is small, your designs may be restricted, but I'll hazard a guess that N will be large for a 30 passenger plane capable of 2 hour flights (e.g. a Tesla model 3 is almost 3000 cells; 31s96p).
A plane needs to be able to takeoff in a given runway length, with a given payload, and climb at that power to a safe altitude for an option of an aborted landing. And just Tesla set to a high power mode, if when starting to drive it you do max acceleration for 1-2 minute you drain a significant portion of your battery very quickly.
The overall point that Im trying to make is that slight battery specific energy density improvements don't matter when compared against the power losses which are proportional to square root of the current.
>Yes the problem is very hard. Mainly because high DOF generalization is very difficult.
>We have self driving cars because what are the control inputs? Pedal, brake, steering wheel. This already took many many years.
Its actually amazing to me that this hasn't been solved yet. Its really not that hard of a problem.
Modern robotics, including self driving, are famously all about end-to-end training. We are trying to replicate what humans do through muscle memory. But muscle memory is not what makes us good at operating in the physical world. The thing that matters the most is our ability to simulate the world around us in a compressed form into the future, which lets us predict how our inputs will affect the world.
A similar system in a self driving car should be able to drive perfectly without self inflicted accidents 100% of the time, especially with basic lidar to serve as an error correction mechanism to the camera 3d scene reconstruction.
Gaming as a whole is a shadow of what it used to be, specifically because it has been tailored to the masses who have no real opinion on any of this other then they want to play what their friends play.
My guy, there’s more weird and niche indie games than there’s ever been. Not too long ago, even AAA studios like EA were trying super weird stuff, like It Takes Two.
Nostalgia glasses are real, and yeah we don’t have a Good and Evil game anymore, and we’re not in an RTS renaissance; but there are a LOT of VERY good games out there, even for your particular niche - and if there’s not, then Godot and Unity make it VERY easy for you to add to the pile, like hundreds of thousands of teams before you in the last decade.
My takeaway from this is apparently you can build a datacenter in a place where the weather reaches 110 degrees in the summer. AC has come a long way. I wonder what cost feasability of building essentially a biodome that is actively climate controlled that spans an entire town or city.
More niche use case is typing. Im on zfold 7 because I use Termux quite a bit.
reply