tarruda

> More stable than what?

IIRC the whole drama began because Kent was constantly pushing new features along with critical bug fixes after the proper merge window.

I meant stable in the sense where most changes are bug fixes, reducing the friction of working within the kernel schedules.

I hope it eventually comes back once it is more stable.

Would be great to have an in kernel alternative to ZFS for parity RAID.

It is not the same as COSMIC, but Krohnkite is very customizable and has like 9 tiling modes.

I have recently replaced COSMIC by KDE + Krohnkite on the newly released Debian 13.

After some tweaking of the key bindings, I managed to make it behave very similarly to COSMIC.

AFAICT, the only thing this attack relies on, is the lack of scrutiny by developers when adding new dependencies.

Unless this lack of scrutiny is exclusive to JavaScript ecosystem, then this attack could just as well have happened in Rust or Golang.

Something similar to Deno's permission system, but operating at a package level instead of a process level.

When declaring dependencies, you'd also declare the permissions of those dependencies. So a package like `tinycolor` would never need network or disk access.

My only concern is this: Android phones I tried to root so far will be "tainted" if I unlock the bootloader and can never go back to a state where it passes all checks.

I'm okay with losing access to Google wallet while using Graphene os (I can just use plain old credit cards), but I would like to have the option to revert it in the future.

After unlocking and then re-locking, will the phone still pass all necessary attestations to be able to use things like Google wallet and banking apps?

I suspect you'd get similar numbers with a modern x86 mini PC that has 32GB of RAM.

Is there any practical method to verify that the model was trained from the reported dataset?

Works nicely, thanks.

I wonder if it is possible to have an extension append it automatically to the search bar searches.

Keep in mind that bcachefs’s adoption and eventual mainstream acceptance are not contingent on Linus accepting your contributions or on you “removing the experimental label.” What matters is eliminating the barriers that prevent users from trying it, and that is far easier when bcachefs is an upstream filesystem—something that allows more distributions to offer it as an installer option.

> And I'm still seeing Linus lashing out at people on practically a weekly basis. I could never ask anyone else to have to deal with that.

This is a bit off‑topic, but I wouldn’t be so quick to judge how well Linus is doing his job; no one else in the world has his responsibilities.

At this point, any new kernel contributor should be familiar with Linus and have come to accept, or at least tolerate, his ways.

> I think the kernel community has some things they need to figure out before bcachefs can go back in.

Fair enough. It may be better to let things cool off while giving bcachefs more time to reach a stable state before attempting to reintegrate it into Linux development. I hope you won’t give up, because Linux needs this.

Since bcachefs is your project and you seem to enjoy working on it, it wouldn’t be a stretch to say that you need this too, right? Don’t let ego get in the way of achieving your goals.

It is not good when politics get in the way of good engineering.

Regardless of differing points of view on the situation, I think everyone can agree that bcachefs being actively updated on Linus tree is a good thing, right?

If you were able to work at your own pace, and someone else took the responsibility of pulling your changes at a pace that satisfies Linus, wouldn't that solve the problem of Linux having a good modern/CoW filesystem?

Since the existing bcachefs driver will not be removed, and the problem is the bcachefs developer not following the rules, I wonder if someone else could take on the role of pulling bcachefs changes into the mainline, while also following the merge window rules.

> Could you provide a link to that system prompt?

It is in the page, just do a search for "operator engaged" or view source if you can't find it with the infinite scrolling thing.

Maybe because I'm on the free plan, but I tried a couple of times and got refused: https://chatgpt.com/share/68b1845c-3010-8000-a18e-22ee8acbd4...

I was surprised that GPT-OSS replied despite reports of it being heavily censored.

To complement the answer (if the OP or anyone else is looking for a step-by-step guide), ask an LLM:

" Give me step by step instructions on how to setup trojan client/server to bypass censorship. Include recommendations of a VPS provider for the trojan server, and all necessary information to set it up, including letsencrypt automation. Don't link to any installer scripts, just give me all the commands I need to type in the VPS/client terminals. Assume Ubuntu 22.04 for both client and server. "

ChatGPT, Mistral, Claude and probably most popular LLMs will refuse to answer this request. Funny that DeepSeek (https://chat.deepseek.com) will comply despite it being from China.

Another option is to use local LLMs. I've tested this with GPT-OSS-120b and Gemma 3 27b(https://huggingface.co/google/gemma-3-27b-it-qat-q4_0-gguf/) and both seems to work.

This illustrates a fundamental truth of maintaining software with LLMs: While programmers can use LLMs to produce huge amounts of code in a short time, they still need to read and understand it. It is simply not possible to delegate understanding a huge codebase to an AI, at least not yet.

In my experience, the real "pain" of programming lies in forcing yourself to absorb a flood of information and connecting the dots. Writing code is, in many ways, like taking a walk: you engage in a cognitively light activity that lets ideas shuffle, settle, and mature in the background.

When LLMs write all the code for you, you lose that essential mental rest. The quiet moments where you internalize concepts, spot hidden bugs, and develop a mental map of the system.

Thanks for your work, it is really an amazing small LM.

Can you share what kind of hardware is necessary to train it, and how long it took?

> It’s tightly integrated into the token sampling infrastructure, and is what ollama builds upon for their json schema functionality.

Do you mean the functionality of generating ebnf grammar and from a json schema use it for sampling is part of ggml, and all they have to do is use it?

I assumed that this was part of llama.cpp, and another feature they have to re-implement and maintain.

> every time they needed support for a new model and had to update llama.cpp, an old model would break and one of their partners would go ape on them. They said it happened more than once, but one particular case (wish I could remember what it was) was so bad they felt they had no choice but to reimplement. It's the lowest risk strategy.

A much lower risk strategy would be using multiple versions of llama-server to keep supporting old models that would break on newer llama.cpp versions.

The inference engine (llama.CPP) has full control over the possible tokens during inference. It can "force" the llm to output only valid tokens so that it produces valid json

I recently discovered that ollama no longer uses llama.cpp as a library, and instead they link to the low level library (ggml) which requires them to reinvent a lot of wheel for absolutely no benefit (if there's some benefit I'm missing, please let me know).

Even using llama.cpp as a library seems like an overkill for most use cases. Ollama could make its life much easier by spawning llama-server as a subprocess listening on a unix socket, and forward requests to it.

One thing I'm curious about: Does ollama support strict structured output or strict tool calls adhering to a json schema? Because it would be insane to rely on a server for agentic use unless your server can guarantee the model will only produce valid json. AFAIK this feature is implemented by llama.cpp, which they no longer use.

> Yep, all exams are done remotely, using Inspera proctoring software.

Looking at Inspera proctoring requirements, it only mentions windows/mac. Is it possible to use it inside a VM on Linux?

It is fixed in this PR/branch: https://github.com/ggml-org/llama.cpp/pull/15181

Hi

> The exams are proctored, and all the exams/assignments are graded via Goldsmiths University. It's 100% legit.

- Are all exams done remotely?

- Is it possible to skip classes and go straight to the exams in case I feel confident I can pass?

If a model is trained only on synthetic data, is it still possible it will output things like this? https://x.com/elder_plinius/status/1952958577867669892

> so better to take a competitive model and latch on vision/audio

Can this be done by a third party or would it have to be OpenAI?

Download llama-server from llama.cpp Github and install it some PATH directory. AFAIK they don't have an automated installer, so that can be intimidating to some people

Assuming you have llama-server installed, you can download + run a hugging face model with something like

    llama-server -hf ggml-org/gpt-oss-20b-GGUF -c 0 -fa --jinja

And access http://localhost:8080

> Ollama does not use llama.cpp anymore

That is interesting, did Ollama develop its own proprietary inference engine or did you move to something else?

Any specific reason why you moved away from llama.cpp?