madbo1

This is exactly why privacy by architecture matters more than privacy by policy. The Netherlands trusted a policy ("Solvinity can't access the data") but the architecture allowed it anyway. The only real solution is cryptographic sovereignty systems where even the vendor mathematically cannot access user data, regardless of what US law says. Not we promise we won't look but we literally cannot look. Building something small in this direction a mesh network where identity is a BIP-39 seed phrase and messages are E2E encrypted at the protocol level,not the application level. The goal is that even I as the developer cannot read user messages. It's still early, but this problem you're describing is exactly why it needs to exist.

The direction seems very interesting indeed. Using optimization algorithms such as Augmented Vertex Block Descent in WebGPU marks a step towards doing proper numerical computations right in the browser and not just visualizations and/or ML inferences.

Perhaps, more importantly, this development makes optimization approaches more accessible by eliminating the need for special environments (CUDA, native compilation, and alike).

I am very interested in comparing the performance of these optimizations in comparison with native GPU implementations, especially in terms of memory throughput and precision limits.

This seems quite significant. While many people still think about AI as something cloud bound, there are limitations such as latency, cost and, most importantly, lack of control involved.

By moving AI agents into an execution environment where they work locally, one gets both deterministic execution, reduced latency and avoids transferring information to remote clouds all the time. In certain application scenarios, for instance, when building a personal assistant or implementing automation routines, this makes a huge difference.

The problem here is not only running the model locally – that seems increasingly easy to achieve, with developments like Ollama but also managing multiple agents and coordinating them in a manner that doesn’t require powerful hardware resources.

In case GAIA manages to simplify this process to make local execution of multiple AI agents feasible, this might very well lead to a transition from 'AI as a service' to 'AI as personal infrastructure'.

Pretty exciting stuff, really.

Reading this from India, where stuff like this is pretty much Tuesday business. But that’s not the problem; the problem is precisely the one hour of your life spent trying to figure out whether the issue is your DNS, your VPN, your configuration, or your programming. “The government in the country I’m accessing this from just decided to shut down my IP for the next two hours” rarely crosses your mind.

India has consistently been at the top of the number of Internet blackouts anywhere in the world for years (Access Now keeps track of this through its KeepItOn project). These tend to be brief and localized, triggered by something as mundane as an exam or protest or local incident. It’s such a routine occurrence here that there’s even a reflexive response: mobile data works differently from other connectivity types, so go with that, try new DNS settings, rely on Telegram instead of WhatsApp when the latter fails you, and always have a list of mirrors.

What’s fascinating about this case is that it’s identical except for who is pressing the button LaLiga, a privately owned entity, in place of the government.