Hmm..

  pacman -Ss ollama | wc -l                                                                                                              
  16
  pacman -Ss llama.cpp | wc -l
  0
  pacman -Ss lmstudio | wc -l
  0

Maybe some day.

llama.cpp moves too quickly to be added as a stable package. Instead, you can get it directly from AUR: https://aur.archlinux.org/packages?O=0&K=llama.cpp

There are packages for Vulkan, ROCm and CUDA. They all work.

That doesn't make sense. Why would llama.cpp need to move any faster than ollama? For that matter, why not have a llama.cpp package and llama.cpp-git in the AUR?

what are you talking about? llama.cpp doesn't need to respect ollamas speed at all. It does not depend on it, it's the opposite of that.

The claim was that llama.cpp moves too fast to be in Arch's normal repos. But Arch does package ollama. Therefore, either 1. ollama somehow avoids the need to move fast, or 2. it moves at an acceptable pace when packaged.

Edit: Or perhaps put differently: If ollama includes a copy of llama.cpp and has a non-AUR package, why can't there be a non-AUR package that's just llama.cpp without ollama?

yay -S llama.cpp

I just installed llama.cpp on CachyOS after reading this article. It’s much faster and better than Ollama.

Then again...

  zypper --no-refresh search llamacpp | tail -n5 | wc -l
  5

Sometimes Arch has the software you want at the version you want, other times it doesn't but other distros do. That's why there's half a billion distros instead of just one.

I am quite confident that the following was NOT LLM:

> New users were signing up but not doing anything, they weren’t creating an org, a project, or a deployment, they just left an account sitting there.

Surely the LLM version is:

> New users were signing up but not doing anything; they weren't creating an org, a project, or a deployment—they just left an account sitting there.

It really depends on the LLM and the wrapper prompt. There are many other giveaways though - which I am not going to name to burn them.

You really should stop using LLMs to write messages complaining about LLM use though. (the "it depends" and the hyphen-as-emdash were dead giveaways).

/s

> the only artifact left is the diff

You also have code comments, docs in the repo, the commit message, the description and comments on the PR, the description and comments on your Issue tracker.

Providing context for a change is a solved problem, and there is relatively mature MCP for all common tooling.

Not to mention AIs predilection for copious and overly abundant comments.

"The Brutal Choice"

Is there an established name for this LLMism?

I don't need a "Reality Check" or a "Hard Truth". The thought can be concluded without this performative honesty nonsense or the emotive hyperbole.

This probably grates me more than any other.

> Not paying musicians anything at all if they don't have enough streams

... 1000 plays in a year?

We're taking a handful of people (Close friends? A proud mother? The artist themselves?) listening a few times a week.

If an artist has no following, and creates music that listeners consider substitutable for AI slop or low-effort shovelware, then they are hobbyists with no reasonable right to renumeration?

Outside of work, I'm a very sporadic coder. On some side-projects where I'm using Actions, I'll have an inspired few days of progress followed by completely idle weeks/months/quarters.

Losing free Actions doesn't particularly bother me, and I have no issue with paying what is most likely a negligible amount, but I don't really want to have a credit card on file which could be charged some unbounded amount if somebody gets into my account. I've shut down my personal AWS for similar reasons.

Is there any way of me just loading up a one-time $20? That will probably last well into 2027, and give me the peace of mind that I can just let it run. If my account's compromised, or I misconfigure something that goes wild, I am perfectly happy to write off that amount and have my incredibly-low-stakes toy projects fail to build.

Put a spend limit in GitHub and issue a chargeback if they ever bill you more.

Setup something like CircleCI that mainly relies on paid users of their main product, and has a free plan. Microsoft currently seem to be in the process of figuring out how to lower the costs of GitHub for free users, since I'm guessing they make their actual money on other segments and products.

There are several “virtual credit card” providers that allow you to generate additional cards, set limit on them like amounts and who can charge the CC. The availability varies per geography.

The problem with that is you might still get a huge bill if something goes wrong, then they try to charge it to your card at the end of the day/week/month/whatever, and it fails.

Now you still owe them the money, but haven't paid, so they tell you to pay on another card. If you refuse, they start debt collection against you and you could end up with your credit rating being affected, and maybe court cases and so on.

I want give the company an amount of money, then know that it's run out and I have to pay for more. You can set monthly limits (https://github.com/settings/billing/budgets), but if you are like me and have personal projects that you work on for a week or two a few times a year, that doesn't really work.

I know AWS, Azure, and GCP do allow for global caps. Azure has it with subscriptions for example. Not sure if it is only on recurring monthly basis. Having a pre-paid lump sum version available is nice but it would also open the door for denial of service if cash runs out. Maybe that is why it isn’t offered?

As someone who has had an occasional nasty AWS bill, the Ai providers using a pre-pay credit system is something I approve of, and would love to see everyone else offer.

I just want to say I found this quite an insightful comment. I similarly would love to use a pay-as-you-go pricing model as a way of safely trying out various SaaS services.

Unfortunately I feel it is not in the SaaS businesses interests, who want to replicate the gym membership model where the 70% who don't use the service are supplementing the other 30% who use it frequently.

Realistically you aren't their target market. They're targeting the enterprises who already have self hosted runners and aren't interested in switching to Actions minutes.

Australia is a Five Eyes country, with carte blanche access to data that the incumbent social media companies freely share with all the acronym deep-state authorities.

Could you elaborate further on how preventing a sizeable proportion of its citizens from communicating through these established spy-nets, causing them to disperse out to unpredictable alternatives they might not be able to control, increases mass surveillance?

That's definitely an interesting argument I haven't seen before.

I suppose it depends on how effective these types of measures actually are, and also on how many adults refuse to identify themselves. I would assume governments are more interested in spying on adults than under-16s, so the adults are probably more relevant here.

I hope you're right, though. Maybe there'll be a renaissance of smaller platforms. Probably not, but I can hope.

This legislation left it entirely up to the service providers to determine implementation, and so far they don't seem particularly motivated to disrupt my usage by asking me to prove my age.

My suspicion is that fairly simple heuristics of age estimation, combined with social graph inspection, are probably enough to completely disrupt the network effects of "social media" for kids, and achieve the stated objectives well enough that I never have to.

Maybe it turns out that I'm wrong, but why even risk it? If the true policy goal is extending mass-surveillance, why waste so much political capital on such a round-about approach which might yield nothing, or even set back your existing capabilities.

MyID (myid.gov.au) already exists, and could easily have been mandated, or "recommended", or even offered as a means of age verification now. But it wasn't.

So far I haven't been KYCd by anything.

Aside from YouTube I don't particularly engage with any of these often, but my Google, Facebook, Discord, Twitter, Bluesky, (current) Reddit, Slack, Telegram accounts all seem to be BAU without new requirements.

If the 80% of us currently holding unambiguously-over-16 accounts are exempt, and it only affects future over-16 users as they're onboarded, then it is a very blunt and very slow form of data harvesting which won't yield useful results until years/decades after all of the relevant decision-makers have moved on, retired and/or died. So this seems unlikely?

> Out of the five options available, only one is European (the one I am using). What I don't like is how I cannot add my own custom endpoint. What if I run Mistral locally (with Ollama, for example) and want to use that?

Set up your preferred self-hosted web interface (OpenWebUI or whatever, I haven't looked into this for a while), point it at ollama, and then configure it in Firefox:

browser.ml.chat.provider = http://localhost:3000/

At home I point this at Kagi Assistant, at work I point it to our internal GenAI platform's chat endpoint.

Out of curiosity, for the AI inept, how does this work? I can just point firefox at "https://kagi.com/assistant" and it can use it? Is that using MCP or is there some other standard interface for this?

Most of these AI providers use a similar kind of common query structure. OpenWebUI is a mostly consistent copy of ChatGPT so that's what the browser seems to default to when you configure something custom.

All the AI toolbar really does is open http://ai.url.com/some-query?prompt=${formattedPrompt} and display it next to the web page you have open.

The formatted prompt is something like "The user is on the page 'Stories about Cats'. The user wants you to summarize the following text: <text you have selected goes here>". You can configure your own prompt in about:config if you want, there are a bunch of examples here: https://github.com/mozilla-l10n/firefox-l10n/blob/main/en-GB...

There are prompts optimised for specific AI providers but the generic ones should work with any provider your choose.

When the web page opens at that URL, you're either going to get redirected to login and then redirected back, or the AI frontend will start executing the prompt.

It's functions as a mini browser window without a URL bar

I got different IDs in regular browsing vs my first incognito window vs my second incognito window.