I am very sure it will be around for at least 5 years, and can you say the same about any job in particular? We are nowhere near self driving in a car really becoming reliable in all kinds of weather conditions, so do you expect that they will have AI up to snuff for routing planes full of hundreds of people in all kinds of weather and situations? I really don't think so myself.

For sure the demand will still be there, and it is a high rate burn out job. So if you can handle it and don't mind the stress and hours, I am sure you will have a job for quite a few years. Anything beyond five years, who knows in any job at all. But given the consequences if something goes wrong, I very much doubt that job will go away. You might get some more AI assistance taking some pressure off, but I am sure if someone gets fully trained up in the next 1-2 years, they very well could have a lifetime career. I can think of so few other jobs like that right now.

I first off agree it is a super high stress job with long hours, but I don't know I would call it underpaid. It's a six figure job with almost guaranteed overtime with federal benefits. I see the average salary is like around 150k with great job security and benefits, and the overtime will push your pay way above that. I mean who gets a six figure job that actually gives you overtime. Working in tech will require you to work all of those extra hours for just your base pay, and we are talking people under 30 here as they cap starting to 30 years old.....

Other people have mentioned the mac mini, but there is also the DGX spark from nvidia itself. Glad I got one when they first released it as it is much more expensive now, but it has a 20 core ARM, 128 gb of unified memory, and a blackwell GPU. I've run a ton of stuff on it, although I got it more to train experimental models which is really what it is designed for.

Yup....

https://www.bbc.com/news/artic...

Joined the fray of the new marketing technique of showing how powerful our AI is by showing how dangerous it is. Odd times.....

Read that with a sarcastic tone..... I am very much not a vibe coder given my 43 years of experience in the software industry.... I spend a lot of time working with my engineers to find the correct balance between the advantages AI coding can bring, vs the reality of what will happen down the road in production if you aren't careful.

You have to keep up, this is all new and unfolding right now.... It changes fast in this space and the people getting the best results use a combination of Codex and Opus 4.6, while Opus 4.7 so far is considered a flop. More tokens for less performance in the real world.

I take it you have never actually used the codex model then? It generates better structured and more intelligent code than Claude Opus 4.6, using significantly less tokens. Last year, Cursor was the darling of everyone and going to vibe code its way to the top. This market is far from decided....

I have not found this to be true at all, especially if you include cost in the equation and speed. Absolutely Claude Opus 4.6 is one of the best when it comes to large complex tasks involving many agents and files, for now. If I am refactoring code or reviewing code, Codex is significantly better in both cost, speed, and accuracy. Gemini 3 Flash is great for very quick/cheap smaller sections of more directed code at very low cost.

I hate these tools that lock you into a single model. I use a tool set that let's me pick and choose based on what I am doing. Heck, use codex to generate your large plan, and claude opus to then execute on it and you get great results if you are doing something very large. I honestly don't think there will be one model to rule them all, and I hope soon that a true multi model tool that automates the picking of the appropriate model (cross vendor, not just between anthropic models like claude code) will come out and bring some sanity back to all of this.

The other providers are not standing still at all, and there will be a new 'best' constantly. The differences between them will get smaller and smaller, and hopefully better and better SLMs will come out that we can host on more reasonable machines ourselves for targeted tasks.

The reality in the American west is very different. Personal water consumption is not even a drop in the bucket of the American west. It is all agriculture, and growing water hungry crops where there is not any water. 85% of the water in Utah and neary 90% of the water in Colorado goes towards agriculture. Then you look at the kind of agriculture, and it is totally unnecessary. In Utah there are companies growing alfalfa, a very water hungry drop, in the desert. They then stuff it in shipping containers and ship it overseas. A complete waste of precious water. Then you look at the irrigation techniques, and it is a lot of flood irrigation where they use enough water in a single day that would cover a town for a month. The water was divided up based on wet years many years ago. We need to redo that, and start charging for water to these large agricultural institutions based on a more realistic rate. Some things should be grown in the desert, others should not. Somebody not flushing their toilet is not going to change anything at all.

I swear we are all devolving our writing skills back to the day of hieroglyphics. Why, oh why would you ever use an emoticon or eggplant or whatever to instruct a computer to do something? I could perhaps see giving it a flow chart or something, but heck use BPMN, not an image. It's just not that descriptive for procedural operations.

I hear the complaint of not enough money all the time, and that somehow throwing more money at it will magically fix the problem. What I would suggest is to actually look where that money goes. It goes to a bloated administration that continue to suck up all of the resources, and not to the teachers on the ground. Further, that massive administration also micro manages the teachers who are now so constrained on what they can do, and you have parents who demand every single child is put on an 'Individualized Education Plan' that the teachers must also then track and adhere to. This is not sustainable

Look down the line a bit from this one and there is an article how students are hitting college with not even being able to halve a fraction. This is middle school level math at best. So now colleges are being forced to offer remedial level middle school math to make up the difference. Clearly this great system we have here in the US is an utter failure.

I remember in the 70s when I was a child. They came up with this great new math idea of a drawer full of exercises that you would take from and complete at your own pace (even that wouldn't be allowed today). A handful of us were done with the year's exercises in a couple of weeks. We were then given 'self study' which meant us grade schoolers being sent to the library to entertain ourselves. You can imagine how well that went. Fortunately, after grade school we still split into different tracks of math classes where many of us were moved on an AP track, so things got better. Today, they teach weird techniques and tricks instead of core math, teach just to the standardized tests as that is how the administration gets its money, and basically cater to the lowest performers. Adding money will only make this worse.

If I had children I absolutely would homeschool. Public school right now is fundamentally broken and does need a reimagining. That does not mean throw it out and dismantle everything, but someone does need to tip the apple cart and realize what we have now absolutely is not working. We can't deny the outcomes.

Yes! Finally someone who actually understands what this thing was actually designed for. I got one right out of the gate as I want to fine tune / distill expert models for eventual production use. My customers all run on CUDA, so CUDA is an absolute must for testing and development. Eventual deployment will be on much more realistic hardware hosted in the cloud. Developing in the cloud making lots of mistakes adds up fast dollar wise, and is honestly a bit clunky. Developing locally let's me make lots of mistakes. Heck, I kick it off and go do other things and check back later all for the one time sunk cost. I can also see that perhaps my reward model is not functioning properly and just abort it as it isn't going where I want to. A giant gaming PC would cost me more, and not fit in with what I am doing for the most part. I like a purpose specific machine for this work that has the full support of Nvidia and it's stack of software.

And to be clear, it absolutely can run quite a few models for experimentation, once again cheaply. I don't have to do anything but spin up a python virtual environment or use an existing UI based tool and play around. Sure, performance isn't great, but it let's me evaluate things on a very consistent and solid platform. I'm not going to run production inference on something like this for many reasons, just like I would not do that on a gamer machine.

For what it is worth I ran for 3 hours straight at 100% GPU utilization and saw no thermal throttling. It was putting out some decent heat for sure, but it was humming along. Also, I will note, I was doing this off grid at 10,200 feet on pure solar power in a smaller cabin. Try that with a gamer PC!

Actually, it is very reliable. I live off grid in an area with zero cell coverage. I use starlink for internet and also for wifi calling/texting as my primary service for over a couple of years now. That has always worked fine.

Just this year I signed up for and received the beta satellite texting service on my samsung phone. It works spot on every time for simple text messages. I have also always had inreach for emergencies, but am probably going to cancel it now. It is slow, takes a long time, and you have to have clear view of the sky. People also get confused on the weird message they get, and over a few messages I have to start to pay.

With Tmobile satellite messaging I get a signal even inside of my metal roofed cabin. Under fairly heavy tree cover, no problem. Was up at my water intake beyond reach of my wifi. In the past I would have to bring a radio with me to check in with my wife if the water looked good. Now, I was able to simply text back and forth. Unlimited texts, and with the normal read/received confirmation. Never mind just driving around out here I get texts where there previously was zero signal.

Now on the device side, obviously the inreach devices are more rugged, last a very long time, waterproof, etc. They win there. But the service is not nearly as reliable or good.

Interesting.... What did you find buggy about it? I have one and wrote my own python publisher to MQTT for it. It has been running 24x7 for 6 months at least now without a single glitch. My only beef with it is that the heating unit affects the temperature reading in the winter as it cycles on and off, so had to get a different standalone temp sensor as well to make it accurate in cold weather. You could see the jumps in the graph as it warmed and cooled itself.

In practicality from people using it, the distance traveled to space and through their network system means latency is in fact higher than fiber. Fiber also travels as the speed of light, albeit in fiber. People often see 15ms or less on good fiber connections, rarely do you see below 30ms on the starlink system. The signal does not stay in the starlink system down to varying pops.... You are always on the same pop that is closest to you. So like in my case in Colorado, I am always downlinked to Denver.