Lol cool anecdote. Guess karma for hating on billionaires is more important than saving lives.

How is this different than "more" than the sum? Is the argument/claim that we can't figure out stuff via composition? If so, why not?

I like the networking perspective, but the ML perspective is such a loose analogy that it's hard to even judge. I mean, we've known forever softening constraints allows you to reach solutions otherwise unreachable, for one? There's a gulf of difference between succeeding at something deterministic by allowing failure vs. good pattern matching by optimizing over a rough landscape of examples.

> I like the networking perspective, but the ML perspective is such a loose analogy that it's hard to even judge.

Right. ML doesn't have to work well because it's used in situations where the cost of the errors falls on someone other than the service provider. Hallucinations require a business model where their cost is an externality, like pollution.

With an objective goal, such as tests or a spec or driving without hitting anything, to check the results, it's possible to do better, of course.

The Internet only works because fiber optic bandwidth is cheap. As someone who was working on congestion in the early days, I could see that congestion in the middle of the network had no known solution. If congestion could be pushed out to the edges, there were strategies, but there were no good solutions in the middle. And, in fact, the whole Internet would sometimes go into congestion collapse in the early 1990s, with the big peering points at MAE-EAST and MAE-WEST losing well over half of the packets. What saved the Internet was cheap long-haul bandwidth and big hardware-supported switches. This kept congestion at the fringes.

As a corollary, will we see a recurrence of congestion in the middle as FttH sees increased adoption? It's easy to believe that 10 Gbps ought to be enough for everyone, but history tells us that people will find a way to saturate any unused bandwidth (8K video with crazy bitrates, 1 TB video game installs, etc).

I'm not seeing how describing measures over possibility space as allowing for mistakes.

Seems like content reverse engineered from title.

Yeah, I didn't find his initial take very convincing, but he lost me at the followup:

> For most cases I don't think having explainability is worth the trade offs in capability. That'll be a good topic for a future post.

What’s objectionable about that, assuming the tradeoff is real and actually hurts?

"One time". Where have I heard that before?

I did undergrad research on amyloid aggregates in 2010, with Alzheimer's as the motivation. I remember papers coming out then making it pretty clear that there was no real mechanistic link between them. It was basically just a symptom. That was 16 years ago.

It's really hard to understand. There needs to be really loud batman sign in the sky type signals from some hero third party calling out objective product degradation. Do they use cc internally? If so do they use a different version? This should've been almost as loud a break as service just going down altogether, yet it took 2 weeks to fix?!

> ... we’ll ensure that a larger share of internal staff use the exact public build of Claude Code (as opposed to the version we use to test new features) ...

Apparently they are using another version internally.

No. There is no one solution. For some things Naproxen simply doesn't work.

Do you mean, compared to drugs in the same family that work via the same mechanism?

What about cox-2 like celoxib? Chronic pain for 15 years checking in.

Could you share your setup + workflow? How do you get it to not give shit layouts?

The raw Q&A is essential. I think Q & Q works so we'll because it reveals how the model is "thinking" about what you're working on, which allows for correction and guidance upfront.