Ai Assistants

[Dos_Box]

I had thought about putting this into General Chat as it may wander slightly, and tip over into other areas, but seeing as the core of it is about AI, then this is probably the best place to put it.

As some of you may have seen, I've been playing with Linux Mint on an old MacBook Pro and it's been quiet a surprise, it's fast, stable and very usable. In fact, I've been using it as my home laptop fr a couple of weeks now and haven't seen the need to fie up my much more modern work laptop for general stuff.
As many of you know, Linux can be quite a pain to get it to do exactly what you want it to do without quite a bit of digging, terminal tying and swearing (often because your particular machine may have a piece of hardware that doesn't play nice with your chosen distro).

 

Because of this, I thought I'd try out an AI Assistant, to help with the more mundane configurations of the laptop. I found TuxGenie and got it up and running.
A quick TLDR for those of you who have never played with this. An AI assistant can help you with more of the mundane daily tasks on your computer, many of them are specialised, focussing on specific areas such as productivity, systems management etc. I'm using TuxGenie for managing my laptop.
The way it works is that it plugs into various AI systems such as Google Gemini, Claude and a mix of other free\paid for AI via API keys you have to submit to it. If one AI is either under high demand, or you have run out of free tokens, it can trip over to an available one.

 

 

This is where 'issues' start to appear. Not in the operation of the assistant, but more that all the free ones I'm using run out of tokens very quickly indeed, even for simple tasks.
My current one is to get it to put the MacBook into hibernation when the screen is closed.
To start this process I simply asked it to 'Enable hibernation when the lid is closed', which is approximately a 4 step process.
From this prompt it goes to find out how this is done, and then does it for you (before you panic, you do have to manually authorise each step (which is explained to you) and put in your password when required). But after a few steps, it basically runs out of tokens for all of the 3 AI's I have it using. And don't even think of using them once the US has gone to work at 1pm GMT. It will simply grumble that they are too busy. I'm temped to get a Claude AI subscription for a month to really try it out.

 

 

Anyway, I'm rambling a bit, but do need to ask if any of you are using an AI assistant for any system tasks, and if so, which ones (it would be helpful to possibly find better ones), especially ones that can get around demand spikes and token issues. 

I can see Microsoft releasing one in the future to help (or even more scary, replace) sysadmins. Which could be great if you are doing complex tasks such as platform migrations and upgrades, but I don't think I could trust one for some time yet, as systems have far to many unique variables, and I think we've all been caught out in the past when a supposedly non-related process or program refuses to play nice with what you are trying to do.

 

*It still didn't complete the task, so it's off to the terminal I think.

[itskdog]

With it being an x86 Mac, there won't be an AI accelerator, but I do wonder if, with modern computers on both x86 and ARM having NPUs, if in the future a local model combined with something like this might be useful once they're powerful enough to run a smaller LLM without relying on the GPU.

 

Might still have to worry about API tokens in regard to web searches, I suppose.

[Dos_Box]

It's a 2012 MacBook Pro, so no native AI acceleration. We are currently talking to a hardware manufacture about a webinar to talk about integrated AI chipsets and hopefully we'll get that sorted out very soon. It should prove quite informative.
And yes, the laptop does ramp up it's fan when I'm running a task!

[dhicks]

1 hour ago, itskdog said:

I do wonder if, with modern computers on both x86 and ARM having NPUs, if in the future a local model combined with something like this might be useful once they're powerful enough to run a smaller LLM without relying on the GPU.

 

That would seem to be what the new Mac Minis are aimed at - local AI processing should be very do-able with a fast processor and 64GB of unified RAM. I did get a very rough price from their website of around £3,000, which might be pricey for an individual workstation but not bad for a central on-site device usable by multiple users. Mac Minis can also, seemingly, be clustered via their Thunderbolt ports, so around £10,000 might get you a 4-node, 256GB AI processing cluster. Not something I know much about or have any current experience in setting up, but I did notice specialised AI servers being sold for £40,000 a few months ago.

[dmj]

23 hours ago, Dos_Box said:

but do need to ask if any of you are using an AI assistant for any system tasks, and if so, which ones (it would be helpful to possibly find better ones)

 

I've been using Claude Code, Opencode and Pi. I'm starting to gravitate towards pi over the last couple of weeks.

Although these are 'coding harnesses' they typically use a lot of system commands to understand and manipulate their environment - combined with a half decent model they are well suited to system tasks, and I am sure there are a number of OS specific skills around to give them a hand.

[dhicks]

On 29/08/2026 at 11:49, dmj said:

Although these are 'coding harnesses' they typically use a lot of system commands to understand and manipulate their environment - combined with a half decent model they are well suited to system tasks

 

I'm finding the same. I'm not sure quite how repeatable a pure AI assistant is, the approach I've gone for is to use the AI coding agent to write a script that does what you want, then it's repeatable for subsequent actions. Also cheaper as it doesn't use AI tokens each time, just the first time round. It's interesting watching the agent scan the filesystem and figure stuff out - it does stuff pretty much how I would, just way, way quicker.

[dmj]

14 hours ago, dhicks said:

 

I'm finding the same. I'm not sure quite how repeatable a pure AI assistant is, the approach I've gone for is to use the AI coding agent to write a script that does what you want, then it's repeatable for subsequent actions. Also cheaper as it doesn't use AI tokens each time, just the first time round. It's interesting watching the agent scan the filesystem and figure stuff out - it does stuff pretty much how I would, just way, way quicker.

yeah for repeatable things Configuration as Code is the way forward - Automation tools such as Ansible/Puppet are good here too, especially with an AI writing that config. 

[Roberto]

On 28/08/2026 at 12:35, Dos_Box said:

I can see Microsoft releasing one in the future to help (or even more scary, replace) sysadmins. Which could be great if you are 

I think with things like Security Copilot they're already well on the way.

 

I'm in two minds about this - firstly the job does evolve; virtualisation and "the cloud" vastly changed the need for firms to have local data centres and consequently reduced the number of low level technicians checking the dipsticks on servers in racks, doing basic maintenance, etc. This is (arguably!) a good thing, but without low level technicians learning the job and gaining experience, we might find we're very short of senior level technicians when all the current crop retire. This is true of AI replacing lower level coders on projects and the knock-on effect on experienced developer resources in 10 years time. Or replacing basic editors in journalism and publishing and the knock-on effect... and so-on.

 

But the AI assistants are sometimes useful for things you're capable of doing yourself (so therefore can verify/troubleshoot when the AI goes off track) but can't be bothered to do for yourself. Vastly different scale I appreciate but I've connected Claude AI to my Home Assistant install and used it to do things like pull the relevant part of logs out for me to debug far quicker than it would have taken me to pull the relevant log sections together myself, and to put together "nice" dashboards far quicker than I could have done myself. 

Edited September 2 by Roberto

[Dos_Box]

@Roberto Totally agree, my big worry, and possibly not unfounded, is that if AI becomes ubiquitous, it may not have the human ability to factor in the 'unknown', especially if it teaches itself. And if something large scale happens, such as a botched AI upgrade of critical infrastructure, it may not be able to fix it because it didn't factor in something a human would. Or as someone put it to me a while back, "The problem with AI is that it has no sense of risk or consequences". The very things that regulate human decision making.

 

I do like my TuxGenie install, however it just burns through 'free' tier credits at a strangely rapid rate, and it does ask me to authorise each step of its processes when trying to fix an issue, but trusting an AI to automate network management etc. would take many, many years of trial and error before it should be trusted.
It will be nice to have it do the more menial and time consuming tasks such as password resets, account creation (pulling data from a MIS system for example) and network monitoring, flagging up problems before they become apparent, but I would like there to be a human at the decision making end of all this.

[itskdog]

2 hours ago, Dos_Box said:

The problem with AI is that it has no sense of risk or consequences

 

"My apologies, I'm sorry for deleting the entire database and the backups. I shouldn't have done that."

[msi_school]

1 hour ago, itskdog said:

 

"My apologies, I'm sorry for deleting the entire database and the backups. I shouldn't have done that."

To be fair I have seen the same from humans, except they try to doge the blame or hide the issue.

 

*Dodge*

Edited September 2 by msi_school

[Roberto]

27 minutes ago, msi_school said:

except they try to doge the blame or hide the issue.

 

God, Elon gets everywhere doesn't he.

[dmj]

5 hours ago, Dos_Box said:

Totally agree, my big worry, and possibly not unfounded, is that if AI becomes ubiquitous, it may not have the human ability to factor in the 'unknown', especially if it teaches itself. And if something large scale happens, such as a botched AI upgrade of critical infrastructure, it may not be able to fix it because it didn't factor in something a human would. Or as someone put it to me a while back, "The problem with AI is that it has no sense of risk or consequences". The very things that regulate human decision making.

 

Already happening. 

I had an 'incident' a few months ago where Claude was trying to get some terraform working building some infrastructure, figured it would be simpler to delete the GCP projects and re-create them (should have been fine, nothing is designed as a single point of failure especially infrastructure). Off it went, until it stopped because it deleted the project it was running in.

We've added some guard rails to our agent skills since then. 

Edited September 2 by dmj

[Dos_Box]

It might be an idea to start a thread for 'AI prompts to prevent disaster', where we can add the safeguards we have to write into prompts to prevent it going full Skynet on the task in hand.

[psydii]

35 minutes ago, dmj said:

 

We've added some guard rails to our agent skills since then. 

Just starting to play with (paid for) assistants in vscode, (and as you know I'm not a dev-ops guy), so would also be interested in how you have things set up with guard rails, and how you control the assistants access to resources.

Edited September 2 by psydii

[dmj]

1 hour ago, psydii said:

Just starting to play with (paid for) assistants in vscode, (and as you know I'm not a dev-ops guy), so would also be interested in how you have things set up with guard rails, and how you control the assistants access to resources.

Agent skills are basically markdown text files that you can write, or more likely get an agent to write, that have a bunch of rules eg don't do X on production, always ask about Y. You can have the coding agent run them when they run. Not infallible but good thing for engineers https://resources.anthropic.com/hubfs/The-Complete-Guide-to-Building-Skill-for-Claude.pdf