that journalists will be reductive to the point of complete naive fabrication in an attempt to explain technology in terms that allow people to keep the narratives they had about it previously, because they've willingly selected for audiences who don't want to know any better...

We already knew that management often doesn't understand what it pertains to manage, and that it doesn't see a strategic issue with this. That's nothing new.

you think the goal of research is to identify only peak performance...?

It's worse than you thought.

And in addition, from a... control systems perspective... you can never guarantee that the system wont respond in a way that is offensive or potentially break security protocols. It also isn't possible to make it understand the implications of its failures.

Every instance of it being put in front of customers unsupervised will end in disaster sooner or later.

Also a number of other functions.

This is why people think you're trying to hurt them when you're trying to hurt them: You explain that only the parts you personally find relevant have meaning, the rest can be discarded.

I know you probably meant this as a quip, but no. Stop.

A rich guy who has spent years working out the morality.

Maybe this article isn't the place for this discussion but

Why do people presume that there will ever be a time when LLMs are not subject to prompt injection?

The belief that these systems will eventually be what people imagine them to be in narrative fiction runs throughout writing on the technology. It suggests the majority of people writing about it have absolutely no idea how it functions.

Yes by the way, I know what I'm saying is obvious to people who do understand how it functions.

This is one of those forever problems. It's surprisingly difficult to get it right.

yeah i see how the framework is set up, it does make sense to separate it out.

As you say, the divisible workload is the key... but that makes me wonder why it's necessary in this particular implementation.On the other hand, if you're enjoying the work and someone finds it useful... well that's fine then.

From my perspective there's an ongoing conversation around HPC and other types of distributed computing by people who don't know what the words computational complexity even mean... and so I'm used to having to produce comparative figures every few months when some faculty person or someone from business services/central IT rhetorically asks me why we can't just do it all in the cloud.

It can get pretty frustrating.

As unrelated point... I'm not sure Cuboid is the best name, but then name collisions are utterly normal in my line of work.

Lets just say

At least it's in Ruby and not something that AWS will bring to my researchers in 6 months to tell them that they can do all their simulations on AWS and don't need the local HPC cluster.

They'll find out that AWS are wrong real fast, but I'll have to clean up the mess after.

The standard solution in the actual supercomputing world is the venerable Message Passing Interface, of which the most accessible implementation is probably OpenMPI.

Note that this distinct from OpenMP, which is a multi-cpu threading shared memory library.

All of these solutions need to be written into the applications that use them, interchangeably with any other threading systems they may use. The quality of the integration varies greatly, and can heavily affect processing outcomes... but also the way they're set up needs to use sane compute topography, primarily (but not exclusively) for latency reasons.

People make it sound like this is fire and forget and none of it is. Tuning your network to support this kind of operation is non-trivial, and doing it automatically over non-specified cloud topologies without fast interconnect is going to slow it down to the point where you never get results.

Of course, this assumes we're dealing with some indivisible task that requires synchronization across processing. Other use cases do exist.

I work in research.

Don't waste your breath explaining your belief that research is compartmentalized like this is some kind of corporate dystopia story with brilliant science guys in lab coats being the magical drivers of progress and that's why I don't know. It doesn't work like that.

Transhumanism is a story.

As much as I want it to be true (for want of a better word), the closer you get to it the more the narrative nature of the concept becomes unavoidable.

Hype is not appropriate as a driver of this kind of technology.

It's becoming increasingly visible that there's no relationship between having money and having a grip on reality, let alone having an ability to identify good engineering practice.

That claims ownership of the LAION-5B dataset used to train Stable Diffusion.

Something tells me ignoring copyright is an essential part of his business model.

+1

Plus, it smacks of blaming a perceived easy target for development problems you can't solve because management asked you to do something really weird and stupid, and then gave you an impossible deadline... ...And then told the public you'd invented 'negative latency'. I might try to be sympathetic to the guy, but... he's making a spurious claim to justify a problem with a service that will always have problems.