eicker, eicker@lemmy.world

The interesting part is not whether Apple wins the biggest model race, but whether it changes the economics: If enough AI runs locally, every token avoided is cloud capacity nobody has to build. That is a very different business model from selling ever more cloud compute.

It’s easy to underestimate how different these facilities are: Traditional data centres are built to run millions of separate workloads efficiently. AI data centres are engineered to make tens of thousands of GPUs act like one giant supercomputer. The building may look similar, but the power, cooling, networking and economics are almost an entirely different industry.

The market is fragmented across the Member States, each of which conducts its own research. That is precisely what needs to change. Hopefully, this approach will work.

This is a significant step for Europe’s AI ambitions. Access to large scale compute has become one of the biggest barriers to building competitive AI models. If executed well, these gigafactories could give European startups, researchers, and companies the infrastructure they need to innovate at home instead of relying almost entirely on US or Chinese providers.

This is exactly how the DSA is supposed to work: Once a platform reaches sufficient scale and societal impact, it faces higher transparency and accountability requirements regardless of whether it’s social media, gaming, or AI. If anything, it would be more surprising if major AI platforms weren’t eventually included.

These are purely paper profits.

Unfortunately, it is highly likely to take decades before Meta loses its relevance: the network effects are enormous, and the vast majority of people over 35 will probably only leave Instagram and (from the age of 60) Facebook completely in exceptional circumstances. And I don’t even want to get started on WhatsApp…

Meta can absolutely pivot to renting compute for whatever model wins, but calling that an automatic death spiral feels premature: Companies with billions of users can survive technical setbacks for years. Betting against distribution has humbled plenty of investors before.

What strikes me as particularly risky at Meta is the sheer amount of power Zuckerberg wields: ultimately, it’s the same problem as with Musk, even though Zuck has behaved much more moderately so far.

Interesting that Microsoft is effectively running a portfolio strategy instead of betting on a single lab. Anthropic is already paying off on paper while OpenAI looks much more volatile. That alone suggests the AI race is far less settled than the headlines usually imply, even for the company funding both sides. And now the Chinese Open Weight AI models are coming onto the scene too, offering almost the same quality.

Pragmatic, but revealing. If Meta truly believed Llama would stay ahead, it would not be preparing for a future where competitors power its products. It also makes business sense for Meta because it already rents out compute capacity instead of using it for its own models: supporting multiple leading models could attract more customers to its infrastructure.

That’s a more realistic ask than hoping everyone slows down: If alignment methods can be made accessible to engineers without deep math, more people can stress test, critique and improve them. Knowledge spreads far more easily than voluntary restraint, especially when geopolitical incentives point the other way.

Interestingly, China doesn’t need this computing power, yet Kimi has now unveiled a competitive model all the same. I wonder how this will pan out?

So have they been getting better and better?

How much more effective have the large models become for you over the last three years? Do you think development might come to a halt in China, for example, if US providers were indeed to stop developing or releasing them?

Hehe… 🤣

Unfortunately, even with very long texts, this is no longer reliably possible, which is why some universities have recently stopped using such tests.

Reposted in the correct thread…