zamalek

That's a huge assumption, NVIDIA drivers are still hit and miss.

People who are willing to spend $71 on not having to build it themselves. That's the premium according to GNs best-effort like-for-like build.

Note that the performance of the RTX 5050 is completely irrelevant if it doesn't work. While some say that their NVIDIA rig is working, that's a risk that's probably not worth taking for something that is expected to "just work." The last thing a console user would tolerate is dealing with whatever mess NVIDIA has dreamed up next.

You need to build within the same constraints.

Also, it's $71 more thanjdiy according to GN.

Unfortunately many of these developers believe that kernel-level is mandatory (by virtue of believing that it actually works to prevent _all_ cheating, which we know it doesn't).

Out-sourcing does still involve humans, to be fair. If the premise is that "humans prefer human output" then outsourcing would still be preferred, even if it is preferred less.

> with Kerberos, etc, it all just works

I worked with customer's AD environments in the 2010's and I remember whiteboards of figuring out customer Kerberos config. "it all just works" is not my recollection of that 3-headed beast lmao.

I don't like anything Apple, I had an awful experience with MacOS (the laptop was great), iPhones and their 12 finger gestures (again, terrible software) aren't for me.

WiFi and worsening slowness woes with the Google TV/Chromecast pushed me into buying an Apple TV, and it is the single Apple product that I would glowingly recommend; that's saying something coming from an Apple hater. It is truly a fantastic product.

My biggest problem with Nextcloud is that it intentionally broke (at least twice by memory) on some updates, wanting me to run some oci command on the box. Remember that "for my family" also means "for my family when I'm gone." I never got round to having to find an alternative thanks to the divorce, but I'd consider Nextcloud a complete non-starter for this reason.

Find an alternative.

Is anyone doing text diffusion in latent space instead of tokens?

Dynamo is a fundamentally different DB to Postgres. If your problem fits into the dynamo approach (I'd argue that more problems do), then you should be using it. No all problems fit, though.

$1000 in a month? What on earth are you doing in CI? We're on Warp build (50% cheaper instead of 33%) with 6 people hitting it pretty hard across a few repos and only reach $150/mo. That's with building rust - a worse case than anything you're using for build time.

1. I'm not sure if blacksmith offers larger runners, but you if you're using them you have checked that bigger runners are worth the squeeze? A 2x runner does not mean a 2x faster build - I had a goal and sized CI according to that.

2. Caching. Nx can do this for your TS, provided your code is decomposed into packages.

3. At a previous job I had a post-run timings job. GH markdown supports Mermaid; I used a gannt chart to express it. I don't remember if the GH api supports getting timing for the current wf - so it may have been a second workflow.

The first is just a little legwork, the rest your agent should be able to do in 5 minutes.

> API is definitely being sold at a decent profit.

Where do you get this from?

Enterprise plans are being cancelled or limited all over the place (Uber, Microsoft). I doubt Anthropic would be leveraging a loss leader with their consumer plans, while catastrophically hemorrhaging customers on the enterprise.

They are either operating at a loss (possibly a minor one), or a minor profit (which is chasing customers away).

If they were comfortably profitable they wouldn't need to participate in the circular deal circus.

Zig and probably Pascal have the same advantage, can't speak for Ada as I've never built it.

C and C++? You've got to be joking. If the project provides static binaries, sure, but I don't want to have to worry about finding a necronomicon and summoning the correct kind of imp required to properly use whatever insane build system the project is using.

> and no one knows why.

I think we can make an educated guess as to why. Maybe we could get Claude to reverse engineer it to clean up its own mess.

If the figures look softer than expected, keep in mind that it's local lay-men buying in - while foreign career investors are selling out[1]. Broadcom's earnings call may have kicked it off: the CEO started off mistaking last year's figures, and had to start over with below projected numbers[2].

[1]: https://www.cnbc.com/2026/05/18/s-korea-market-volatility-ne... [2]: https://www.fool.com/earnings/call-transcripts/2026/06/03/br...

This the lobbyist's FUD, and SKG are so up-front with the reasonable constraints that it makes it pretty obvious that this is a paid piece.

Edit: oh, it's yours. Spend 5 minutes understanding exactly what SKG have said they are not asking for.

> and injects relevant context back into future prompts

It looks like this is left as an exercise for the student?

That plan smells AI generated:

> Gooey is in better shape than most ~140 KLOC Zig codebases — every directory has a mod.zig, namespaces are layered, and core/interface_verify.zig provides compile-time platform-backend checks. But the architecture has drifted in a few

I too have used AI to plan cleaning up its own mess, and this self-congratulatory prose is extremely consistent ("every directory has a mod.zig", whoop dee woo!).

In my experience, AI is largely incapable of fixing its own mess to an actually competent degree (and full disclosure: I still ask it to, not pointing fingers here) and it's probably due to it walking on egg shells around its own feelings. I've had to tell it to completely change course during cleanup at least 30 times this week.

Also: https://xcancel.com/mitchellh/status/2060088112257372610

I don't currently have a personal use-case for container services, but Quadlets are another example of systemd (and podman) beauty. It looks like someone has gone through the trouble of making the OS+home-manager modules: https://github.com/SEIAROTg/quadlet-nix

> with AI-generated content excluded from pre-training.

Though this is largely impossible these days, unless they pre-trained on pre-AI era data.

That is an absolutely massive amount of unsafe, is the plan to reduce it over time? The benefits of the Rust build system are reason alone to use this (as a direct competitor to the C version), but claims about memory safety are suspect no matter how confident the authors are about correctness.

I've was part of an acquire-to-remove-an-annoyance, including having the better product that didn't win for niche reasons - which vaguely matches the situation here. Yes, NVIDIA can wind down Groq (unambiguously a better product in this case) even though it doesn't make sense.

That being said, there's still a chance that NVIDIA engineering is in the process of stripping it for parts. Or the lawyers are - maybe they have too much momentum with GPUs and just want ASICs out of the market.

This kind of innovation stifling acquisition should have been blocked. NVIDIA is a serious monopoly threat.

I would argue that this form of disclosure is ethical in the face of Microsoft misbehaving. It's like mutually assured destruction - and in this case (it sounds like) Microsoft tried to cheat and thought they would get away with it.

Feeling consequences are how they are kept in line. Maybe next time they will think twice before (allegedly) treating a person like they did here, as well as the creative reasoning I recall them using in the past to reduce payouts.

First result: https://arstechnica.com/ai/2026/05/amid-mythos-hyped-cyberse...

There's also things that have since been discovered:

* Ralph Wiggum loops

* Simply not allowing an agent to stop its turn until all tasks are marked as done

* Sub agents over worktrees

* Context compression

> Probably more interesting

It is widely suspected that self-inflicted "bad news" ("Mythos is so dangerous we just can't give the public access to it") is nothing more than Dario's typical style of marketing - keep in mind that they have an IPO coming up, because he certainly factors that into everything he says in public (as is his responsibility, to be fair).

An alternative reason for delaying the model might not be "we are trying to make it safe." It could be "we don't know how to host this thing at scale, or cost-effectively".

GPT 5.5 has already been shown to be as adept as Mythos at finding vulnerabilities.

Finally, laymen massively underestimate the importance of the harness for model performance. OpenHands existed long before Claude Code, Claude Code changed everything because of the clever hand-holding it does. Mythos is definitely more than just a model.

The productivity is going into perverse incentives[1], e.g. we have improved (by which I mean "increased") token use. More PRs every day. More lines of code. All things we knew were shit-brained metrics a decade ago (obviously except token use).

We've also increased how much our coworkers need to read, or deal with. You can get an AI to make any point you want, so you can ignore the 5 humans raising alarms due to the 1 clanker you made say what you want to hear.

All numbers going up.

There are obviously people producing additional true value with it, probably, but that's almost certainly scarce.

[1]: https://en.wikipedia.org/wiki/Perverse_incentive

I have tried Nim meta programming (to make a tree vaguely like the one used by Zed), and the tooling support of pretty dreadful. I ran into multiple compiler crashes or simply unhelpful and confusing error messages, alongside LSP hangs.

Last I checked (which was a while ago to be fair), LLVM machine code quality still lagged behind GCC - so things should be slightly more interesting with the GCC back end.

There were also some bugs (hence disabled optimization passes) and missed opportunities from the lack of aliasing Rust precipitates - again, not sure where those sit - and GCC will have to play catch up here (unless there are other languages that exercise this part of the backend).

You'd have to have a model that has never seen children, which is perfectly fine, but I can't see any of those with the resources to train such a model form scratch actually doing so.