zmmmmm

there is hypothetically impossible and practically impossible

if AI tools get good enough that they can exploit ingrained hardware properties such as subtle timing and race conditions and swarms off them become cheap enough to deploy so that they can persistently exploit low probability events we could reach a state where I think it's actually practically impossible.

> None of this was free. Through the summer, much of our engineering effort went into tests and tooling, and customer-facing changes per engineer stayed between 17 and 30 a month

This is the reality I think for a lot of orgs and is absorbing a lot of the productivity gains. The gap between vibe coding a green field app from nothing which is inherently "AI native" and implementing AI tooling into complex orgs with large amounts of legacy infrastructure, integrations, ingrained practices etc is enormous. For us, it has been a big effort to enable agentic coding over a significant footprint of our codebase. And all this is happening while the plane keeps flying, with near zero tolerance for defects in critical systems. There is a lot more to it than you will understand if you just look at twitter or Hacker News.

This could all go multiple ways but one of them is we end up in a place where technical security is just not possible to implement any more. What does that world look like?

Presumably it means much more behavioral security and inherent friction built in to everything we do. Maybe high value transactions will just inherently have to take multiple days to do, just to slow down things enough that human reaction time can equal what LLMs can execute?

On the technical side, I can see zero trust becoming completely saturated through every layer. The passkey you authenticate with isn't opening a gate to a trusted tier within which components trust each other - even within the same software component. Every internal function call at every layer doing hardware signature verification against continuously minted private keys (since we assume exposure, the lifetime can't be longer than milliseconds now).

It's really interesting to think through what a fundamentally robust security environent does look like if you bake in an assumption that technical security is impossible.

Yes I actually agree with you on the support question.

Quest Pro is a good example though because yes, they provided OS updates. But they never released anything like the required software to address professional use cases that they suggested would happen when they launched it. That's where I harbor the doubts about support. I just don't think Meta "gets" professional / business use. They aren't built with that DNA as a company.

It's going to be interesting to see this compete with the Meta VR glasses. They are essentially identical for productivity except the Aura has physical light pass through while the VR glasses use external cameras to provide pass through. The Meta VR glasses will have better resolution I think, but then the Aura has access to the whole Android ecosystem - so that means real implementations of Microsoft Word etc.

But the real choice is the unfortunate one of who is least likely to abandon the whole thing when it doesn't take off and leave you with an un/poorly supported device. Both companies have terrible track records on this.

i might be overly optimistic, but I'm assuming it'll do the tiling by itself with the above. As long as there's not an algorithmic reason to do it a specific way, in theory you could just optimise it for the target hardware.

Seems like a nice win even if it just does the example they have:

    void matmul(Matrix2DFloat A, Matrix2DFloat B, Matrix2DFloat C, int size) {
      for (@Parallel int i=0; i<size; i++) {
        for (@Parallel int j=0; j<size; j++) {
          float sum = 0f;
          for (int k=0; k<size; k++)
            sum += A.get(i,k)*B.get(k,j);
          C.set(i, j, sum);
        }
      }
    }

But the elephant in the room is the Vector API. Is that going to end up like logging - a standardised API nobody uses because it came 10 years late to the party and isn't as good as what was already established?

And why confuse everybody by calling it a VM when it ships as a JAR file that runs in an existing VM?

I think most people in the tech sector and especially driving the AI hype have absolutely no clue how difficult it is to implement change in complex organisations. There are a bunch of technical reasons but half the reasons are not even technical - social, legal, etc. You have to shift the practices of humans and they have incredible inertia. Like decade level change response period. And during the change period they are less productive. Expect a decade type period before "productivity" impact manifests and even then it won't really be measurable directly because people will be doing different things.

Since the comments are going to be flooded with people reactively assuming it is straight up directly accessing them, here is the actual qualification:

> [they will not use for ads] ... unless you choose to make data from these services available to other Google services that support personalized ads

So it is more reasonable than I think most people are assuming.

Trump has shown zero loyalty to anybody he doesn't have a future transactional benefit from. It's amazing the loyalty he seems to generate given how utterly unreciprocating of it he seems to be.

To an extent yes. But there is no previous version of humanity that incorporates every perspective of every viewpoint ever expressed in writing before. It's obviously derivative of its training material but that doesn't stop it having a novel aspect.

perhaps that is the original sin that leads to our ex post facto rationalisation tendencies. the moment we became prospectively rational we would be unable to content with our own forthcoming deaths.

You have to think the current administration is completely and totally owned by foreign interests at this point.

The question is, what do you do as the new administration coming in? Presumably you assume every system and every person is compromised all the way down, like a compromised server - clean wipe and re-install.

One would assume that, like any good hacker, the current set would be heavily focused on ensuring they can survive a reboot / wipe. The question is, what would they be doing to achieve that?

Critically, in the intended approach for this, Claude itself (the harness) runs outside the sandbox. Their approach means you trust Claude is perfect, just the commands it runs need isolation. And to some extent, you trust its judgement on what and when needs the isolation applied.

I know there are distinct camps here, and the long term story may evolve - but for now I just want the whole harness inside a sandbox - ideally, a full VM - and THEN you can further sandbox things it runs if you really want different tiers within that.

it looks beautiful

I am sceptical of these things however since they are a well worn path that so far has taken off only to a limited extent. In reality, you get bored very quickly in these environments, and it's impossible to implement impressive visual fidelity in worlds that users build themselves. Horizon Worlds is probably the marquee example of that. Am curious what the differentiating angle is here to previous efforts?