The joys of Google AI

[marjohn56]

i have just spent a very interesting number of hours with Google AI. It has built,along with my input a Powershell script, or should I say a number of them that monitor my main windows server,  3 docker containers running on two VMs on that machine, a monitor for Caddy which is running on my QNAP, 5 Unifi access points, and a script the monitors what I think are the critical services on Opnsense, i.e. the gateway status, Unbound, these are checked by using the API. If the main server goes down which runs most of the scripts, a script running on the QNAP will send the alarm. Most of the notifications are sent via email, but what happens if the internet/Router go down, in those cases it gets sent using ntfy to our phones and an always awake tablet. Very impressed with Google AI, I have never written Powershell scripts, but they are all working perfectly. I'm glad I've retired!

[meyergru]

From one retiree to another: On a side note, just because that is probably not too obvious for everybody (I learned this just now): Try the paid versions of any cloud LLM - they are order of magnitude better that just the free crap.

I found over the course of the last two years that the free stuff did not get better, but even seemed to get worse. This may have been the result of the companies trying to reduce non-paid efforts. I found that answers came too early and were imprecise to a level that they caused more harm than good. Switching to a paid version and selecting a better model immediately changed the game. In their current shape, cloud LLMs can really help speed up things. I could do most of that stuff myself, but it would take me much longer.

[keeka]

Very impressed with Google AI, I have never written Powershell scripts, but they are all working perfectly. I'm glad I've retired!

I worked on and off in surveying, IT operations and ecommerce. I can honestly say I derived most job satisfaction from coding and task automation. Now AI can do all of it. I really cannot see it ending well but that's age based skepticism, I hope!

[OPNenthu]

Any advice from the graybeards on what you see as worth investing time into learning now?  How would you advise a colleague who, for example, had just entered middle age and was let go from a company where they had worked for many years and now find themselves lost?

As I see it AI is still in flux and not stable enough to make long term plans around.  Had AI not come into the picture, I would have maybe advised the colleague to learn a new in-demand programming skill such as web development, say, but now programming is not even valued.  The "value" seems to be in managing the AI and taking personal responsibility for its outcomes, while dealing with increasing demand for output.  It's an idea that Cory Doctorow coined as the "Reverse Centaur," where now humans are in service to the machine rather than vice-versa.

What are the good jobs now and in the foreseeable future?

[meyergru]

That is an interesting and difficult question - I am grateful I do not have to ponder that out of sheer necessity.

What I see now is that LLMs have some specific advantages over humans. For example, the more capable models can keep a much larger amount of information in their working context at once - something I have found much less useful with smaller local or free models.. They can therefore do tasks "in memory" for which I would need pencil and paper, or several intermediate steps. They can also access information on the Internet much faster than I can.

So I basically use them as tools that can be guided and directed by me - and much more efficiently and cheaply than a team of human assistants ever could.

For people doing intellectual work, I would argue that the important thing is to develop the skills needed to make the best use of these tools. That being said, it still takes a lot of expertise and problem-solving ability to direct them properly - at this point, they still need expert guidance and supervision.
My best guess would be that the gap between good and bad engineers will probably widen. There may be fewer people doing this kind of work, but those who remain in demand will be highly skilled and very productive.

So I would invest less in learning one particular new technology, and more in strengthening problem-solving skills: learning how to frame a problem, break it down, question assumptions, move between different levels of abstraction and determine which problem actually needs to be solved. Domain expertise certainly helps, but these skills transfer remarkably well between domains.

Quite often, the important question is not how to perform a task correctly, but whether it is the right task to perform in the first place. Those decisions should not be left to managers alone, and AI is certainly not suited to making them either.

That is something I have been practising for decades. Story of my life. Even without the AI part.

[mooh]

So I would invest less in learning one particular new technology, and more in strengthening problem-solving skills: learning how to frame a problem, break it down, question assumptions, move between different levels of abstraction and determine which problem actually needs to be solved.

The current large language models have moved beyond producing output tokens based on stochastic processes and some input. Most LLM interfaces provide insights into the reasoning of the model. They identify unclear questions and ask for input. They also identify sub- problems, ponder about solution strategies and why they choose one over the others, etc. It can be quite amazing, really, sometimes even more interesting than the answer itself.  Problem solving strategies are just another "language".

If I was a young person today I'd probably become a craftsman, assuming that robotics still has 20 years to catch up.

And when I come home after a long day of building the next AI data center I will enjoy the discussions in the OPNsense Science Fiction and Philosophy forum.

[meyergru]

I believe that, in some respects, they have come very far, but I think we are talking about two different things. I certainly expect an AI to outperform me at individual tasks — otherwise it would not be much of a tool. What I do not expect it to do is decide what my goals should be or which goals are worth pursuing.

Asking clarifying questions, decomposing a problem or comparing solution strategies neither requires nor indicates that kind of agency. Those are extremely useful capabilities, but they still operate within a goal and context supplied by somebody else.

I tend to regard current AIs as collections of remarkably powerful, but uneven, cognitive capabilities rather than thinking entities pursuing goals of their own. If they ever became better than us not merely at solving problems, but at deciding which problems ought to be solved and why, then we would indeed have a rather different problem.

I also do like the "OPNsense Science Fiction and Philosophy" wording, and maybe robotics really is further behind than "thinking", so that might be an alternative. There are still everyday tasks that even the most advanced robots struggle to perform reliably, while a six-year-old does them without thinking — such as picking up a key, orienting it correctly and opening an unfamiliar lock.

[OPNenthu]

[...] I will enjoy the discussions in the OPNsense Science Fiction and Philosophy forum.

You found the discussion engaging, or at least the opportunity to project subject matter expertise.  Why this bit of shade at the end?

[marjohn56]

AI, no matter how clever it is, is still subject to the acronym GiGo. For a free system Google AI is pretty impressive, and it can write scripts and create web pages faster than I can. 🤣

What it also is very clever at, is making suggestions at how I can add and improve things. It wasn't my idea to add all of the monitoring, but I said that I have a windows server running two VMs that were running Windows, and that on those VMs were immich and Jellyfin etc, it said waste of resources, let's run them all in Ubuntu server VMs. So it walked me through the migrations using docker containers, then there was the Caddy implementation, originally on Opnsense, it didn't think that was the best and most secure way, so a new vlan was created to pipe the traffic to Caddy on a spare port of the QNap, and on and on. It's completely changed my system. It's all now running very smoothly and the servers using less resources, it's a win win. I also end up with a nice web page showing this.