> “Our companies are competing not only with Chinese rivals, but with the Chinese state budget,”

That's rich coming from the car industry in Germany. Volkswagen is partly owned by the government, and the whole industry has received all sorts of incentives over decades. The result is that they botched the electric transition and are now behind technologically.

Yeah - but long before "electric" was a thing, the German car industry was screwing up in multiple dimensions. Just talk to a few mechanics about the traditional reliability and TCO of German cars, vs. Japanese ones.

I don't know, I think intent has also become cheaper. Agents have intent, and can summarize intent from the changes pretty easily (as demonstrated by this skill).

In my experience Claude Code already does a pretty good job of including relevant context in the PR, although it's true that for commits it usually defaults to "this and this changed".

Yeah for sure, the skill leans on the agent being able to identify the intent. I'd say roughly 30% of the time I have to either prompt it to fix the intent, or it will ask me explicitly what the intent is.

Without the skill, the commit messages are often a list of files changed with a list of bullet points at the end saying:

- Added 8 tests to file.xyz

Pretty useless for a reviewer, they can see that from the diff.

Title could be "The Museums of the University". This is not only a US phenomenon, some European universities are essentially becoming museums if you look at their budgets and administrative priorities.

One aspect of the AI bubble that is not talked about very much is how the European market is a key factor in any serious calculation about future revenue. If Europe decides to, or is forced to decouple its digital infrastructure from the US, that essentially slashes the addressable market of a company like chatGPT by a third. And Europe has some of the richest customers too.

In other words, Sam Altman et al. should be hardcore Atlanticists at this point.

You are right, but I have the feeling that the Google, Microsoft, ... and the IA companies think that the EU is a acquired market. It's false, they can shift off the US, they eventually will.

Maybe this will happen eventually but decoupling any time soon is a pipe dream. For the foreseeable future, Europe's BATNA is shit.

Forget Microsoft and Google services, what about the hardware? To support all this new demand for European infrastructure you'll have to buy tons of new gear from mostly American companies: AMD, INTC, NVDA, MU, etc.

Are cutting-edge European competitors going to suddenly spring into existence to satisfy that demand? Is TSMC gonna allocate wafer spins to some scrappy EU startup instead of NVDA, AAPL, AMZN, MSFT, AMD, INTC, AVGO, QCOM?

I dunno if you've been paying attention to the market but demand for all data center components has gone through the roof and supply is already spoken for for years to come. The hardware you'd need to decouple simply isn't available, when it becomes available you'll be competing with nearly $1T in annual hyperscaler CapEx, and Europe has no capability to produce domestic alternatives.

Decoupling would be painful and certainly hardware would be the most difficult challenge. However, consider that for all those companies you mentioned, Chinese alternatives are starting to appear and will be very competitive in the next few years. Europe also has some key strategic cards to play when it comes to hardware, such as access to TSMC.

The most likely scenario in my view is that Europe will be a consumer/bystander in the AI race, not a protagonist. For that, you don't need a lot of hardware. China and the US will compete for the European market, with the US already far ahead and earning billions from European customers. But it can lose that post in the future, either through crazy foreign policy or simply being outcompeted by the Chinese.

Look, I support Europe's push for computational sovereignty. It's frankly long overdue. But Europeans still consistently underestimate the scope and difficulty of that task, and it must be undertaken simultaneously with building military and energy sovereignty while China's eating their economic lunch.

There isn't the money, the attention bandwidth, the industry, the IP, the skilled labor with know-how, the tools, or the raw materials to do all of those.

> Chinese alternatives are starting to appear

Maybe, but they are seriously handicapped and not competitive. And certainly when it comes to inference: They also don't have the hardware capacity to supply their own market and the European market.

> will be very competitive in the next few years

Maybe. More likely they'll only be competitive in 10+ years.

> Europe also has some key strategic cards to play when it comes to hardware, such as access to TSMC.

Then why haven't they played them? Both Japan and America got commitments for significant domestic advanced node fab and packaging facilities from TSMC. Europe only got 40k WSPMs of already-outdated 12nm+ lines.

Further: Which European company is designing or will soon design competitive CPUs/GPUs/switching to hand off to TSMC? Will a European company cut in line for cutting-edge tools, conjure up IP and know-how from nothing, and start selling HBM and DRAM, too? And will all these fantasy companies be decoupling from US EDA tools, as well?

Finally: If you're saying that Europe will be able to outplay the US in a zero-sum game for access to TSMC wafer spins, you might want to consider also the strategic cards that the US can play to prevent that.

> Europe will be a consumer/bystander in the AI race, not a protagonist

That's not just likely, that's absolutely guaranteed. Sorry, Mr. Draghi.

> For that, you don't need a lot of hardware

How do you serve inference at scale without a lot of hardware? If both US and China are supply-constrained for GPUs that they're turning into high-value-added products/services, why would they give Europe any hardware at all?

If I can turn a $100k GPU into $1M of value, there aren't enough GPUs in the world, your companies don't have the fiscal firepower to be price-setters, and your products/services compete with mine on global markets--why the fuck would I sell you any GPUs? Charity?

And if the plan is for European computing to remain totally dependent on US and CN... what are we even talking about here?

> But it can lose that post in the future

Sure, nothing lasts forever. But I thought this was about European computational sovereignty, not dependence on the US specifically. I guess not: Depending on the US is a crisis, but depending on Qatar and China is A-OK. What could go wrong?

It also destroys the winner takes all market. Investors would count on the winner takes all market and give infinite VC money to a start up, so that they would make a product that is slightly better than the competitor and kill the competition early on.

It's true that the cloud act and other data handling issues were already there. There is one thing this US administration did that was unique though, which was to threaten the territorial integrity of an European country.

I remember multiple people at HN saying "show me ONE example where AI was used to produce commercial-grade software" like a month ago. Cloudflare alone has posted a couple of examples recently, and yesterday Ladybird was ported to Rust using AI.

The most interesting aspect I see in all these examples is that extensive test suites make the work very straightforward. Maybe AI will produce a comeback of test-driven development.

How is this production grade? The last few things CF posted with AI were outright lies or omitted large swaths of functionality.

If your stance is assuming we have an existing implementation of something in the training set, and we have a robust test harness already, and we have thousands of dollars to throw at tokens, and we're not at all concerned with "works" THEN this is viable then sure? But that doesn't seem to be what most boosters are saying.

Honestly, might try this in my next project. Just code the tests and have AI implement against it.

There will be a lot of those scams from now on. Take millions in funding for allegedly training a "sovereign AI" and buy a bunch of GPUs, delivering an actual good model is optional.

The EU is about to fall for the same trick, maybe multiple times over, by funding multiple toy model initiatives that are far behind SOTA. I believe they will also impose ideology on every model trained in the EU, similar to what reported in the article.

Right now EU is tricked by their own leading research institutes. Sad to see all these invented metrics and based-on-nothing statistics that lead only to “gib monay” at the end of the presentation.

>>I believe they will also impose ideology on every model trained in the EU

As compared to what. Every AI model has ideology of its creators baked in.

Right. But we're at stage 2 now.

Stage 1, the SOTA LLMs of current top vendors, started with general feel-good alignment and ideology that's compatible with progressive western sensibilities (so quite neutral, even if not completely), and gradually adjusted to address issues that could generate bad press and hurt revenue.

Now that these are established, anything new - such as, but not limited to, those "national" LLM projects popping up around the world - need to differentiate themselves from the incumbents - and that implies more bias and ideology, since you can hardly have less bias and ideology than what Stage 1 LLMs have.

When a government funds an AI model, the issue of its ideological control and manipulation is different than a private AI model, for several reasons:

- It uses tax money from everyone to reinforce views held by a portion of the population

- Governments have an inherent conflict of interest in representing truth (see examples in TFA about Modi in India)

- That AI model might in the future be the "official" source of truth in the country, not just another commercial alternative that has to compete in the market

EU gerontocracy with social democracy characteristics is what the author is implying I think

people seem to prefer private censoring over govt. one. govts wanting to censor certain content online always has pushbacks not applied to social networks, for example

People haven’t had a choice, so this is not a revealed preference. From what indications I am seeing, people were hesitant to empower government, but are increasingly amenable to the idea.

For countries other than America, it’s not just private enterprise, it’s Foreign Private enterprise.

For both, self serving and genuine reasons, Tech is highly resistant to calls for openness, data sharing, and cooperation with civil society.

A perennial questions from civil society is “how do we engage with tech?”, “what did you do after we sent you data and evidence?”

Ignorance is the only product of this, and that is driving suspicion, fear and anxiety in voters.

Without this swell of resentment, there would be no support for social media bans.

I was surprised to see this "impose ideology" line show up in the Heritage Foundation's project 2025 discourse, citing the example of generative AI making images of black of female Vikings. An obvious dog whistle to white purity. The argument seeming to be that early attempts at guard rails to prevent model bias of generating mostly white men when prompted for images of things like managers etc, were imposing woke ideology, but removing these guardrails all together, and letting models act in alignment with existing social representation biases, there is apparently no ideology being imposed at all. Because their way is just the normal way, and everything else is ideology. I remember first seeing this Viking thing show up in the Discord channel discussing the Dalle Beta, then the next time it was in a white-house statement.

>>Because their way is just the normal way, and everything else is ideology.

I think you hit the nail on the head there.

>The EU is about to fall for the same trick, maybe multiple times over, by funding multiple toy model initiatives that are far behind SOTA. I believe they will also impose ideology on every model trained in the EU, similar to what reported in the article

Most important think for EU is an option of EU hosted inference , you can't fake that with a "clever" prompt. Then you also need support for all eU languages and you also can't fake that with a system prompt.

Anw what ideology is EU pushing that scares USAians and Ruzzians? that humans have value and should be equal in front of the laws ?

Btw do you guys remember when X had a prompt to make their AI stop saying bad things about Trump and Elon? It was found out, X blame it on some intern and the story dissapeared from the public debate, and IMo is a very on topic story, where tech is used to bias the AI to suck up on a politician and a billionaire.

How many of you think that was actually an intern mistake and that now the Elon AI is 100% not RLed to be biased to be pro Elon and MAGA ?

The ideology is that if your speech goes against the establishment they will (mis)use hate speech laws to censor you i.e Merz and the pensioner story, but this has become a common tool in many EU countries. Also, see Chat Law where they are trying to "save the kds and from terrorism" to massively spy on all private chats. The law keeps being struck down but they keep presenting it again, and again and someday will pass because they just need to win once. While other nations have their issues, one cannot remain blind to ours because "look at muh ruzzians and americans"

I never said it was Merz himself who prosecuted, but if you want examples it's ok. - Stefan Niehoff, abother pensioner who insulted Habeck himself, was acquitted but after a lot of drama ensued. Habeck has filed 805 suits on speech so yeah - Nancy Fraser, and the journalist who was sentenced to 7 months in prison - The Pimmelgate

I am sorry but the average German politician has very authoritarian tendencies and Germany itself has many laws that constrict freedom of speech. For now that is ok that they keep building these laws (its for democracy guys and against hate speech) and when AfD eventually gets in the government, you will get a Coyote's law, where you build tools the opposition will eventually use against citizens as well

The chat law is some proposal paid by USAian people, I bet in USA there are also lots of idiotic proposals.

The hate speech thing is not a new thing in Europe, defamation, supporting genocide always was illegal what is new is Ruzzian bots pushing fake shit.

I am not afraid to post with my real name that I think politicians are stupid, corrupt and whatever , what I can't do in Romania is glorifying the fascist and their genocides and crimes.

No it actually isnt a proposal by "USAian", it is a homegrown issue. I dont care about American idiocy because I dont live there. On the other hand I live in the EU and thus I care about our idiocy. Should I be ok with authoritarianism because "China is also authoritarian".

Please tell me, what Russian propaganda pushed Germany to use its anti-semitism laws to deport foreign students who were protesting against the actions happening in Palestine. We have to stop with the "but look aborad its worse" mentality

Are Trump and Elon still friends? I'm not on top of these things, but I read about their disagreements. One is of course about the "liquid gold" vs electrification.

Looks neat and original, congrats!

I don't quite grasp how to interpret the training data attribution process. For example, it seems to say that for a given sentence like "They argued that humans tend to weigh losses more heavily than gains, leading to risk aversion", 24% is attributed to Wikipedia and 23% to Arxiv.

Does that mean that the concepts used in this sentence are also found in those datasets, and that's what's getting compared here? Or does it mean that you can track down which parts of the training data were interpolated to create that sentence?

Great questions. We weren't quite explicit about the training data attribution process. We'll discuss this in more detail in future work. We can track down which parts of the training data were interpolated to create that sentence. For those training data sentences, we then compare the concepts between generated and training.

We can attribute to exact sentences and chunks in the training data. For the first release, we are sharing only concept similarities. Over the coming weeks, we'll share and discuss how you can actually map to the exact training sentence and chunk with the model.

For a technical overview of how some of these models work, check this link out: https://www.guidelabs.ai/post/prism/

That would be great because "I got it from Wikipedia and Arxiv" isn't exactly useful.

From reading your second link (and please tell me if I got it wrong) it sounds like it isn't actually tracking to training data but to prototypes which are then linked a posteriori to likely sections of the training data. The attribution isn't exact, right? It's more like "these are the likely texts that contributed to one of those prototypes that produced the final answer." Specifically the bit in PRISM titled "Nearest neighbour Search" sounds like you could have a prototype that takes from 1000 sources but 3 of them more than the others, so the model identify those 3, but the other ones might matter just as much in aggregate?

It says that the decomposition is linear. Can you remove a given prototype and infer again without it? That would be really cool.

This part of the claim is involved, so we have future posts to clarify this. And yes, you can remove a prototype and generate again. We show examples in that prism post.

In prism, for any token the model generates, you can say, it generated this token based on these sources. During training, the model is 'forced' to match all the prototypes to specific tokens (or group of tokens) in the data. The prototype itself can actually be exactly match to a training data point. Think of it like clustering, the prototype is a stand-in for training data that looks like that prototype, we force (and know) how much the model will rely on that prototype for any token the model generates.

The demo in the post is not as granular because we don't want to overwhelm folks. We'll show granular attribution in the future.

This already happens in Europe and China, today.

Doesn't that cut both ways though? What prevents me from taking the code from dbt-fusion and AI-laundering it to include all the new features on a FOSS fork?

AFAIK AIs do not learn from the proprietary code.

What would stop them? AFAIK any and all data and content that can be obtained by any means is used in model training.

most models are built on theft anyway, especially if you take the GDPR viewpoint

Except the megacorps won't let you use their closed software for learning. They're fine to use others' copyrighted materials though.

There are plenty of US corporations who don't look further than a SOC2 report.

reverse engineering stays under "fair use" when it's for interoperability purposes

Which is an order of magnitude harder than reading the code.

Of course but the investments needed in obfuscations and the bugs introduced are orders of magnitude more important. It looks a lot like the struggle between ad providers and ad blockers - the ad blockers may be lagging behind but eventually they catch up. Same things with scrapping