vhantz

I had never heard about SRE before and suddenly it's all I see. Modern problems (slop fest) require modern solutions (someone to babysit broken systems)?

  Site reliability engineering (SRE) is a discipline in the field of software engineering and IT infrastructure support that monitors and improves the availability and performance of deployed software systems and large software services (which are expected to deliver reliable response times across events such as new software deployments, hardware failures, and cybersecurity attacks).[1] There is typically a focus on automation and an infrastructure as code methodology. SRE uses elements of software engineering, IT infrastructure, web development, and operations[2] to assist with reliability. It is similar to DevOps as they both aim to improve the reliability and availability of deployed software systems.

https://en.wikipedia.org/wiki/Site_reliability_engineering

Maybe read TFA...

  > Fortunately, OpenAI has brought us one step closer to realizing Borges’s vision. Rather than generating every possible mathematical paper, they’ve found a way to probabilistically generate things that look correct.

Or at least ask your favorite LLM to read it, then give it your knee-jerk comment and ask it to check if it's already covered on the text. This saves you from wasting time using your own brain.

And still in no way warrants tens of thousands of pushes. These all overlap pretty broadly

Yeah that's all the same thing

Scaling what exactly? Anthropic doesn't release any new products except models. I'm sure you don't need tens of thousands of pushes every day to train models. Scaling fast is so far just their words, nothing else indicates it.

And how is this different than perforce?

What is all this code being used for? I'm 100% sure it's not going into their harness which has been stagnant.

Ha

Why would they? They would stand to profit since anthropic would have to take theor market share by doing the same for cheaper.

A company with (I'm guessing) less than 2500 developers need to make tens of thousands of merges every day? And they somehow had a team of 7 Claude baby sitters handle this entire thing? I use Claude everyday and I'm predicting this will turn into a major clusterfuck in not too long. Even taking into account the fact that they can burn virtually infinite tokens internally, this will not end well. If I'm wrong, I'll be glad.

  > Engineers are no less valuable, but their work is changing. All of the engineers who worked on this system would now rate Claude as better at writing code than them. But our engineers still supplied the invaluable understanding of the high-level inter-system components and the design-level tradeoffs, making informed choices around the problems we wanted to solve and the scale we knew we needed to achieve. Very often, we found that the role of the engineers was to tell Claude “no.” Agents will sometimes violate core system constraints to satisfy a pattern that should never have been allowed in the first place.

Well that's exactly what "writing" code always meant? It was never about what volume of code you can write. Better at writing code always meant better at writing good code... So how is Claude better than their engineers?

It's incredibly short-sighted to acknowledge that the baby sitters haf to tell Claude no and still not consider what subtle and nasty bugs managed to slip through. Well I guess we'll find out soon enough if that's the case. I'm betting when that's the case.

Unless you have a reasonable argument why a random AI would be trained on Microsoft's proprietary code, I don't see what you're basing any of this on. If you need the definition of clean room, go to Wikipedia.

Great ad for GLM 5.3 I have been an absolute buffoon paying for anthropic's stunted model s when I can access their same accessibility and more for cheaper. I'm switching

Is Microsoft selling the code of its proprietary products for AI trainings? Definitely news to me

Summary at the end is definitely an anti pattern too

How many token/s do you get with that setup?

It's a lot of different parts, the workflow itself is pretty straightforward. It is also all archlinux oriented because that's what I use everywhere. I'll try and clean it up over the weekend and make it available somehow. Where can I reach you once that's done?

Yeah that's the same thing I do. Except the credentials to upload or to interact with the infra in any way are not available during the build job in the workflow. Forgejo reuses the same GitHub conventions so initially they were all available there too

> The worse impact is that you have access to all the deploy keys, and on builds.sr.ht, there are deploy keys for sr.ht itself (probably not the case with other instances).

This is the most unexpected bit.

Circumstances made it so I got a new computer right around the time srht put out their anti-LLM policy. So I immediately started setting up a forgejo instance on my home network. With the latest waves of attacks on he AUR, I decided to keep that machine completely free from all AUR packages or helpers. So I also set up a package registry through the forgejo instance. With some work I have a workflow that takes a PKGBUILD and publishes an arch package that I can just `pacman - Syu` to install. Obviously I only put my own pkgbuilds and the few I have manually vetted there.

Even after manual vetting I wanted to lock things down as much as possible on the CI. So. I went with rootless podman, and some patching of arch's build infra just to get it to be able to run `pkgctl build`.

One of the main thing I looked out for was making sure no credentials were available anywhere during the jobs where the pkgbuilds (which can contain arbitrary commands) are being executed. I'm sure srht has to consider many more scenarios than I do. That will definitely make it harder to wall off credentials access. But it is still surprising that deploy keys are left laying around.

BTW this is not an attack on srht. I stopped actively using it because of their new policy. I still have a paid subscription that I do not plan to cancel anytime soon. I still have my public projects where I don't use LLMs there. I think opinionated maintainers and projects (Linus/linux, Poettering/systemd and Drew come to mind) are a net good for open source. Opinionated people make better software because they care imo. And those opinions don't have to be restricted to purely technical matters.

But anyway, it's a small team and there is a lot to cover, but they should get those leaking keys under control as much as possible.

Maybe you don't know how to read but my point is pretty simple. Even if all involved agents are conscious and intelligent, a stochastic process is not an intelligent one.

In short, even in the case of humans, which are universally (by humans) recognized as intelligent, such a process would not exhibit intelligent thought.

How would that process be the result of intelligent thought?

Presumably everyone in the chain would have their own idea of what the next word/token should be. They would each try to push it in that direction using their only lever, that single token. None of that guarantees that the output is syntactically correct, nor grammatically correct. Note that LLMs at least by the way they draw their tokens have that pretty much guaranteed. So that would be a regression even from what LLMs are capable of right now. But beyond that, even if it ends up being a correct sentence, it would not be a result of intelligent thought. Even if every agent in the process was intelligent, the process is not itself a use of that intelligence. Human beings can be part of purely mechanical processes, that doesn't make the mechanisms suddenly an exhibition of intelligent thought.

This also applies to society and human history itself as processes: “History is made in such a way that the final result always arises from conflicts between many individual wills, of which each in turn has been made what it is by a host of particular conditions of life. Thus there are innumerable intersecting forces, an infinite series of parallelograms of forces which give rise to one resultant — the historical event. This may again itself be viewed as the product of a power which works as a whole unconsciously and without volition. For what each individual wills is obstructed by everyone else, and what emerges is something that no one willed. Thus history has proceeded hitherto in the manner of a natural process and is essentially subject to the same laws of motion. But from the fact that the wills of individuals — each of whom desires what he is impelled to by his physical constitution and external, in the last resort economic, circumstances (either his own personal circumstances or those of society in general) — do not attain what they want, but are merged into an aggregate mean, a common resultant, it must not be concluded that they are equal to zero. On the contrary, each contributes to the resultant and is to this extent included in it."

https://www.marxists.org/archive/marx/works/1890/letters/90_...

What else could it be doing? That is literally the mechanism of how a model works.

English is at best my 3rd language. I have always found the huge discrepancy between pronunciation and spelling really nonsensical (keep in mind that I even speak French). But this article really helps see all that in a new light. Nothing is a-historical, not even phonetics. Newfound respect for the English language.

Still, large language models model language.

You don't need to prove to me that your intelligence is comparable to LLMs. I believe you.

To all commenters who think LLMs are intelligent and comparable to themselves, you don't need me to change your mind. I believe you.

What else are they modeling?

Building up strawmen against LLMs will only make the hypers look more reasonable. A language model responds to inouts exactly as a model of anything else would. That is enough to explain all the "intelligence" without believing a model is somehow a "new kind of mind". Those who think LLMs are intelligent don't know enough about models. And those who think they are useless don't know enough about models.

Claude Code's initial instructions to a model are a dump of 80kb of text. And none of it has anything to do with security. The "security" is handled by the evaluator model that sometimes denies some tool uses. I'm sure that in itself can be done very cheaply.

> study > anthropic.com link

Sure

> live *spiking* simulation