machinekob

Not in 2022-2023 wtf are you talking about.

People are just not buying AMD stuff and AMD wasn't able to provide enough chips when they want (last year case when they just "paper" launch chips 4fun with big hype) it's not some hidden conspiracy.

Probably dependant on country interpretation, but you can try make it to EU level institutions for proper help.

It seems to be pretty easy to get your contract to terminate here in Poland even in case of non-competition clausule so it would be even easier for something like retroactive price change.

"Unfair contract terms" its pretty easy in most cases especially if its something as blant as changing price backward without acceptance of other site.

Also every country can have different interpretation (It is possible that some countries are very limited and you have to try EU level help) In some cases its even illegal to write non-competition into your contract without providing payment for the time of this being in-place ( there were tons of cases when contract was voided by that and I have even one of my contract ended on my account after single lawyer visit cause of this :] )

I would assume Unity won't even try to enforce it in EU.

But ofc. after update its over and games not yet released can most likely be forced to pay.

In EU it's indeed illegal and in 90% of EU countries it won't even go to court. But they can probably force you to not upgrade anything cause of licensing of the new software or if you upgrade they'll make you pay fees from that moment onwards.

But in US im pretty sure it is open for the legal battle that will cost tons of cash for both sides.

> Every system is good if you are making money not from work but from wealth.

But not every system is taxing you 65% from work and 0-20% from wealth. Indeed US wealth disparity is enormous but this is also a country with highest number of billionaire, gigant corporations and biggest stock exchange. And all of that didn't change the fact that US middle class is wealthiest middle class in the world (except some small countries that cant be compared Norway/Qatar etc.)

In which European cities PhD can afford to buy apartment (Paris? London? Berlin?). Or you are speaking about smaller cities in which you can also make the same argument for US.

Tbh there is a price for that, our taxation is one of the highest in the world with stagnation for past 15 years (even with enormous "immigration" flow).

Also people who are in top of their fields almost never stay in Europe (especially if they are young) as potential tax burden for top brackets (which aren't that high its mostly middle class in US wealth scale) are enormous and total taxation can be as high as 60-65% (with VAT and local laws [homes, car, fuel]) for 1.5-3x less money then in US or even in some Asian based corporations which directly makes EU stagnant in most of the R&D fields (even in IT i can make like triple the money in US or 1.5x in China corporations).

EU system is good if you are making money not from work but from wealth (rent, stocks etc.) which is only for folks who already make a lot of investments in younger days or just inherit a lot of cash/properties/stocks.

I know US folks love to be angry about their country, but they are in best position for making decent living in the whole world, except possibly some small asian/european states like Switzerland/Singapore etc.

And Vulkan drivers are like 10-20% slower compared to DX11 on NV.

Ok :)

TensorRT with torchscript is a king as its a lot easier to modify, but ONNX is fine as you can also import some ONNX models to TensorRT -> https://docs.nvidia.com/deeplearning/tensorrt/api/python_api... But ofc. as everything it depends on OPs version of ONNX.

ONNX is pretty old at least 5 years and its still mostly useful on Nvidia GPU's or x64 CPU. TBH. cool that projects like that are still alive but MLIR looks like the future of proper model storage, and custom format and loading is still a king today cause you can easily modify model and optimise or even fine tune which isn't even possible in ONNX without ton of work (also static spec and versions in protobuf sucks wish they migrate to flatbuffers).

Non-commercial licence :[

ROCm is one of the worst GPU software (+drivers) I ever touched, randomly dropping/ not supporting GPUs, random crashes without any proper information etc not integrated with any meaningful framework (It was ~3 years ago).

I have more hope in Apple/Intel joing "AI" chips market over AMD even if they make good enough hardware you cant use it without a lot of software problems (srsly just get your act together AMD, as community support is slowly dying after last AMD dramas and).

How much plumbers will we need in a future?

I'm pretty sure that it'll end with oversupply on traders (especially in stagnant economy) if AI will cut most of the knowledge workers the world will have to move into some sort of utopia/dystopia and people will just don't have to work at all or it will be 0.1% vs 99.9% scenario but with brainwashed people cause of social media and maybe AI censorship.

But its worst case scenario in my mind AI will make jobs just more specialised and output of knowledge worker a lot faster.

Its FP16/Int8 inference only (cause you can only access it via apple frameworks that dosent support training). Also its only used if your data is small enough (4mb cache) it wont be useful for big transformers/ big images processing in a while.

Macbooks have pretty bad GPU per USD performance compared to Nvidia but this post is just a clickbait.

"Why advertise some super AI chip if it doesn’t do AI stuff? Who knows." => Neural engine is fp16/int8 limited so it wouldnt be good enough for training anyway but its pretty good for inference with CoreML in a lot of cases. Also training anything outside Nvidia ecosystem is very bad cause software was written with CUDA first and everything else second (AMD for example have similar story to Metal backend).

As someone who works on Deep learning for past 6 years I can easily say CUDA wont be replaced anytime soon at least for training, almost every software out there was written for either CUDA or CPU.

And for local model development you don't need your rig to be powerful as training wont be done on your local machine anyway and MacBooks are good enough for that (a lot better than windows for sure).

VScode works better on m1 mini than on Windows with 12700k+rtx3090 at least for Python in my case.

They don't write about that, the "paper" is more press release.

Deep Learning training was always very expensive but models werent getting such a massive bump in size every year (for state of the art) and now they are just getting 10x bigger every iteration but AI accelerators / GPUs are getting like 1.5x jump every 2 years so have fun for future AI academia / startups outside US.

ANE is only accessible via coreml and internal apple frameworks so i would assume it wont be using ANE but maybe some neural accelerators in Intel/AMD/Nvidia processors and GPUs.

Accelerators inside GPU (like Tensorcores) seems like a lot better deal as you can easy utilize it without 4 abstraction layers with only some unknown to us mortals operations support inside. (And my god i hope apple will allow to programmable run ANE or at least put this api inside Metal framework cause right now working with Coreml for anything new is a nightmare and even some old models are broken on new versions of coremltools)

If you are Microsoft as GigaScaler with almost unlimited cash and can ignore getting profit of your api/models its pretty easy to undercut all the other companies and offer it very cheap just to gain advantage in the future.

Sadly big players take all in current world and microsoft is pretty big :|

I hate when people don't include approximation for traning before final hyperparameters are found as its most costly part of whole process most of the time.

Just yes we train it for so long etc. but they never speak about tens or even hundres of runs before they finalize the model parameters and architecture -.-

AMD based supercomputers are mostly (maybe even all) float64 machines more physic simulations than Deep Learning/AI (BF16/FP16)

But model need about 350gb so I'm not sure one A100 with 80gb will be enough?

But code is open source so r*ussians who cares?

It's a lot of obj-c/c on apple and syscalls on other platforms, I think the biggest point its to get rid of annoying obj-c calls and make it fully cross platform as its a nightmare for both safety and performance in many cases to call objc.

5.4h more like 10+ h as older and younger population is not doing anything that require spending times in public que, and it is very impressive not waiting extra 6-10h per year + twice that for transporting to and out of the government buildings.

I don't see anyone writing that ML/DS is a lot harder than a typical SWE and results are a lot less visible for a VERY long period of time.

A lot of the time you could train model prepare pipeline and test enviroment for 3-6 months before you get good enough result to push it into production. And it can get extremely stressful rly fast if you care about that, because not every model is good enough for production so once per year you could have as low as 1 or even 0 models that are working fast and good enough for proper usage and this can burn you after just a 1-2 years of work in the field (I know at least 4 people who just drop ML and go for SWE after 1+ year of ML work and all of them are a lot happier with classical backend/devops jobs).

In SWE after 2-3 days you can have small stuff working fine and after few weeks push your small code into production codebase fixing some stuff or optimising smth as there is a ton of potential in almost every codebase for "easy" upgrades in ML space everything is extremely competitive your results are "state of the art" or they are not if you want to upgrade model it better be sota or you would get asked "why we don't just implement/use ...?".

I still love my job but for sure but I prefer working with ops, classical SWE and deployment then model training, optimising and learning/collecting new datasets (I have 5 years of commercial exp in ML/DL maybe it gets better after 10 years or I'm just boring out who knows)