What is missing from the picture with all these articles is the numbers. LLMs already have a few solid use cases as translators, general document processors, coding helpers ...etc. So the first question is, to what extent does this demand support the investment? Would it be enough if basically every SP500 corp provided paid LLM access to their employees? Or is the investment so big, that people are betting on less solid applications, like Agentic AI, with some non-trivial automation?
If AI replaces workers, we wouldn't have an economy. It would probably be the end of capitalism. Or at least the end of the consumerism driven capitalism that we have known since the end of WWII. I don't know what would follow but it probably wouldn't be pretty. Honestly at that point, I could see the end of humanity. If truly we get to the point that machine intelligence is more capable and people are entirely marginalized then it's game over. At best a few human specimen end up on display in zoos, but maybe machines might not even have any use for zoos, since they can just share "experience" digitally.
Car ownership is pretty expensive. But holistically speaking it's not more expensive than the Deutschland Ticket, because it gives you access to cheaper housing options that you wouldn't be able to live in if you depended solely on public transport.
Can confirm for the US too. I live in a rural county with zero public transport, but when I tell city friends what the lot cost and the property tax on it, they have to hold back tears.
There's a way to combine those, and some countries have it done well (I believe this includes Germany) - you have a rail network that is good enough that you have small stops on feeder lines in rural areas; if you want to be rural you can be a mile from a train that gets you to a moderate-sized city where you can get the fast train.
I recently tried to port my model to JAX. Got it all working the "JAX WAY", and I believe I did everything correct, with one neat top level .jit() applied to the training step. Unfortunately I could not replicate the performance boost of torch.compile(). I have not yet delved under the hood to find the culprit, but my model is fairly simple so I was sort of expecting JAX JIT to perform just as well if not better than torch.compile().
JAX code usually ends up being way faster than equivalent torch code for me, even with torch.compile. There are common performance killers, though. Notably, using Python control flow (if statements, loops) instead of jax.lax primitives (where, cond, scan, etc).
Interesting. Thanks for you input. I already tried to adhere to the JAX paradigm as laid out in the documentation so I already have a fully static graph.
I would test how much of the total flop capability of the hardware you are using. Take the first order terms of your model and estimate how many flops you need per data point (a good guide is 6*param for training if you mostly have large multiplies and nonlinearity/norm layers) and then calculate the real time performance for a given data size input vs the actual expected theoretical max perfomance for the given GPU (eg 1e15 FLOPs/s for bfloat16 per H100 or H200 GPU). If you are already over 50% it is unlikely you can have big gains without very considerable effort, and most likely simple jax or pytorch are not sufficient at that point. If you are at the 2–20% range there are probably some low hanging fruit left and the closer you are to using only 1% the easier it is to see dramatic gains.
Hypothetically speaking, if the average Chinese person was as wealthy as the average American, how would that affect the world economy and geopolitics? What is America so afraid of?
China is clearly aiming to take over Taiwan, and if they can't do it peacefully they are building up a military that can do it. America should be afraid of that.
China clearly support Russia in their war on Ukraine despite pretending neutrality. Europe should be afraid.
China is doing lots of other geopolitical things that you should be afraid draws you in.
Everything. The whole empire exists on the basis of exploitation, and existence of enemies. The US military industrial complex has immense power, and it needs a reason to exist, and keep existing.
The problem with that index is that each culture reacts differently to corruption. In some cultures, if a public servant buys a coffee using the company card, that's a scandal, and some of those cultures have a reputation for being corrupt.
In Germany, everyone downplays corruption for some reason. But I see it everywhere, especially in everything that had to do with public funds. But it's never called corruption, so corruption does not exist because it is never acknowledged.
I think he makes a good point about immigration. Immigration is not going to save Germany, because Germany is not an attractive destination. The only people who move to Germany are the ones who failed to get a VISA from an English speaking country, and even those will leave as soon as they gain enough experience to reapply for a VISA. The only way forward is for Germany to really invest in its own population and to retain them by giving them attractive enough opportunities.
But how do you accomplish this? Where do you get the money from? Nobody wants to invest in Germany, not even the Germans themselves. So the government tries to jump start investment, but the government is both incompetent and corrupt, so for example the German government bet on Quantum Computing instead of AI. Now the government in investing in AI, but instead of investing in LLMs they are investing in "AI for science" projects, 99% of which are complete deadends. Actually this is a sort of theme. The government funded startups are always doing some sort of "science" based product. For some reason this appeals to the founding agencies because science sounds like a solid investment. But most of these startups are either attacking a problem that is way too improbably of yielding any results, or they are projects that sound good to the uninitiated but are actually fundamentally flawed when you actually dig deep into them (like a lot of AI for science crap).
I think, if there is a way forward, it's for the government to stop trying to be a startup accelerator. They are too incompetent and corrupt for that. Instead, it should be a mediator between foreign investors and local talent. Make it attractive to build in Germany and use local talent to do that. And make sure the local talent gets competitively compensated so they do not emigrate.
I disagree. The main downside of moving to Germany - or any other EU country except Ireland - is the need to learn German (or French etc) if you want to fully participate in daily life. Other than that, the quality of life is better compared to USA. Example: EU Bluecard better than H1B and Greencard, health care, public safety, public transportation, walkable cities,... even now, I would recommend Berlin over San Fransciso - especially if you plan to start a family.
If there would be a magic pill that enabled anyone to learn a new language instantly, that would be the end of the US and UK as major immigration destinations.
I know where you are coming from, but I want to point out that this is not the typical skilled immigrant experience.
I have met a couple of Americans with 10+YOE who have been imported to Germany with competitive salaries, like 150k€/year. They live in Berlin or Munich, are already partnered up, have a nice nest egg from their years working in the US, and just enjoy being expats on a little adventure.
This is not the typical skilled immigrant experience. If you start with 0YOE in Germany, it's much harder to go anywhere in life. And when you don't already have a well populated nest egg, you start getting anxious about your inability to accumulate savings. And a lot of jobs are unfortunately not in Berlin or Munich. A lot of these skilled immigrants end up in some 20,000 pop town, 1hr away from the nearest mid sized city. And then, if they come from non-western countries and they do not look European enough, that also adds something to the experience.
Honest question: What major German company is in a 20,000 pop town in the middle of nowhere?
And if so, would that mean life in such a town is worse than in a comparable small town somewhere in the middle of the US?
Indeed, salaries are better in the US, especially AI jobs salaries.
But you also need less savings in the EU due to a much better social security system, e. g. free healthcare in case you get unemployed, free university (that alone saves you > 100K US$/kid),...
Lots of bigger companies like Zeiss or Bosch have many locations and depending on your team you might or might not end up in a really terrible location.
> would that mean life in such a town is worse than in a comparable small town somewhere in the middle of the US?
You lose a lot of the advantages that you listed such as public transportation, walkable cities and free healthcare (availability is terrible in certain locations). But honestly, it's a terrible location to live in, if you want to have any kind of social life, and consider that as a foreigner you probably don't have loads of contacts in the country.
Even a failed project generates high profile jobs for the duration. Which still could be years. As I understand it, the government isn't lacking profit, just opportunities for people.
The problem is there are no good long term prospects.
For example, you might get a job at a research institute or at government funded startup and gain experience working with cutting edge technologies. The pay is mid, but you don't care because you are gaining experience. Then your contract ends. The private sector cannot make use of your experience. So you have to pivot to programming CRUD webapps or you emigrate, or you get another public sector contract with the corresponding public sector salary.
reply