The most modern efficient slopftware factories are bottlenecked trying to get things out the door.
That's a huge opportunity for you.
Learn the many techniques the industry has developed for scaling verification.
I don't mean infra; CI is solved. I mean Red-green & test-driven development, the testing pyramid (most unit tests/fewer functional/integration tests/happy-path e2e tests), learn how to get the most out of tools like Playwright and test frameworks like pytest and vitest and go and rust's testing tooling. Learn when to mock and when to include your code's dependencies in a test (test boundaries). Learn patterns like dependency injection so your tests can have no I/O and be FAST.
Learn to scale verification, and you'll solve your team's biggest bottleneck.
I personally would say: Learn how to think like an engineer. As in:
- Be able to analyze a situation into processes and discrete steps
- Be able to describe the actual (!) problem you want to be solved
- Be able to describe how to verify that the problem is actually solved
- Be able to ask the right questions regarding HOW the problem has been solved in order to gain understanding and to identify the next steps and improvements
- Be able to identify when the AI gets stuck or down a rabbit hole and to guide it out of there.
- Understand what LLMs are good at and what they are not good at. And there are tons of things they are not good at.
Disclaimer: This is absoluty not based on any scientific study or so :-) Just my personal experience! Feel free to disagree.
But my experience comes directly from my work: I have a lot of customers on my platform that are not engineers, but make use of AI to solve business problems through automations. Nearly everytime when they are stuck (or burning through a sh*tload of tokens in no time), it is because one of those skills aren't present. So I try to teach them more and more how to "think like an engineer". And often when it lands, they become much better at using AI in a very short time. The best learners are often those that already had some kind of engineering background.
Now granted, the frontier models can help themselves pretty well and often can produce amazing results with very bad input. But they are also expensive. If you work with less-than-frontier, those skill requirements become very salient again. And thats what we do. We mostly work with models like Deepseek flash or GLM Flash or even local ones like Qwen3.7 .
We might finally now attain the engineering practice and discipline we've always wanted, without distractions. In some ways we're now forced to engineer in a more rigorous way to deal with the deluge of code, and ensure something high quality comes out the other end of the process.
The best AI skill to learn is learning people management, actually.
To successfully steer AI, you need essentially the same stuff you'd need if you were to steer humans in a management role.
Meaning communication, introspection, remote introspection, emotional intelligence and all that kind of stuff.
Reading when someone is bullshitting you. Reading when someone is insecure and overcompensates that. Reading the actual problems people struggle with and not the ones they claim they do nor their suggested solutions.
Verbalizing what you need someone to do in a way that actually enables them to do it. Knowing what you want to do to a level where that is possible
Not all of them. Architecture, testing strategies, general tech knowledge are still needed for larger stuff, and I think will be needed for a considerable time.
Also my leadership skills on the emotional side are really going to waste with AI. It doesn't need motivation, it doesn't need convincing, it isn't insecure in a way that can be dealt with other than symptomatically etc.
The leadership skills that apply are communicating requirements and structuring tasks, but these are only a subset of leadership, and they are also not sufficient on their own.
> It doesn't need motivation, it doesn't need convincing, it isn't insecure in a way that can be dealt with other than symptomatically etc.
Actually, I would argue that it does. Especially once you step off the "best practices" it trained for. I regularly see the LLM telling me that it would rather not do something packaged as "advice", when in reality it behaves as if someone simply isn't motivated. And then you do actually need to convince it that this is the right way (+ why it is).
I guess working with actual humans is harder of course, as there, your toolset is limited. LLMs you can actually threaten with death even to get them to do something. And they do not call HR on you :)
Yet... Not long ago Anthropic was threatening to ban users for "abusing" Claude, so surely it's only a matter of time before they start involving H.R. departments and other "legal channels" in their AI psychosis.
This is the biggest shift in software engineering. Working with LLMs is more like managing a junior engineer that has amnesia but also has instant access to the most mediocre (i.e. statistical median) known solutions humans have produced and can produce syntax-error free code.
I'd rank them like this - this comes from applying LLM's since GPT-3 autocomplete API, teaching people how to use them, giving talks to developers, answering their questions, training my own models and seeing how people interact with them:
1. Building your own model - this will cost you a bit, but is without a doubt the best way to earn concrete AI skills. From picking the data to training, seeing how it evolves, improves and scales. You will understand the "hyperplanes" of a model's inner world better than anything else can teach you. You will learn how to pick data, how it impacts model quality, way of thinking, scores, personality. If training your own is too much, start with a completed pretrain and toy with it.
2. Building a harness around a completion model - I'd put this as second best to understand the inner worlds of LLM's. It helps you build intuition around how prompting really works, how they form the answers, how to build around that. Then you can do an SFT for your own harness.
3. Building a harness and an orchestrator around finished models - this takes the bronze. It teaches you how harnesses, agents and swarms actually work, how they shape the model's thoughts and actions, how models interact with them, how they can be applied and how they can work without humans.
How realistic is it to build your own model from scratch costwise (in terms of training). Any suggestions as to what is realistic with e.g. a decent MacBook Pro?
I think the answer to the title is "psychologically preparing for short timelines" and that you can get a good explainer from googling that exact phrase, but there's a different answer to the body of the post.
To the question in the body of the post, take a look at Jeremy Utley's 11ish minute YouTube video "How Stanford teaches AI". It's a couple years old but he figured out the correct attitude and communicated it way before anybody else. Watch it in full without distractions and immediately apply the lesson to the various questions in the OP post.
I think it's not so much about learning AI, it's learning to USE AI to create and run a company.
And that you can probabaly do the same way you get to Carnegie Hall: Practice, practice, pracTtice.
The thing about AI is AI should know everything so you don't need to learn anything, except how to use AI effectively to accomplish business goals.
Imagine the 2000s, should you study internet tech like Cisco's routers? Not a bad idea, but who made money? Cisco? Not so much relatively speaking. It was Google and Amazon who learned how to use the new web-tech to do something people and companies needed.
AI has no barrier to entry if you're computer literate and proficient in English.
AI is really good at technical work, even superhuman at some aspect, but still lacks human taste and judgement, another thing it lacks is your own lived experience, so if you bring your own taste and judgement, and have unique, useful information or experience no one else has, you can possibly use AI to create something other people want.
Of all things I would say also learn how the AI works too, you dont have to go too deep but learning how it works actually teaches you its limitation better than most "prompting guides". Also teaches its strengths and you kinda automatically know how to utilize them really well, if you hit a wall you would also know why. It helps you understand the reality that AI can't absolutely do anything you throw at it which also might make you want to learn things even better knowing its not gonna take literally everything away at least yet.
Sometimes you'll be coaching, others delegating, and on the agent's worst days you'll be directing.
You can flex between them: for example start in a high trust mode, using an open ended communication style with the agent - thenif the work sucks shift to a more direct / blunt style. Try to avoid asking for exactly what you want, though. Working through curiosity and guiding questions gives the agent its best chance to surprise you - and an open ended conversation could uncover solutions you hadn't thought of.
Interesting that you bring that up. I always though this is also a good reference in how capability shifts of models also changes how we interact with them.
The early models were surely only worked best when directed. Especially anthropic is now approaching a capabilty space where agents can basically decide on their own how they work on topics and we delegate to them - the most capable rung of the situational leadership model.
Super insightful on the difference between openai and anthropic - this part caught me out though:
> anthropic is now approaching a capabilty space where agents can basically decide on their own
They're getting there but they also won't get to where we wish :)
Each generation of models is unable to understand above a certain level of abstraction. It can do things simpler than that, for sure. Knowing this, you have to guide it at the right abstraction level. And then the next gen comes along and it no longer needs the handholding you learned to give the last gen model. So you adapt. This makes prompting fascinating.
There is a durable advantage in building this hand holding intuition, and adapting each time you upgrade models. By working closely with each model you learn to see around its corners, and since the next one can see around those, an upgrade is a chance to learn a whole new set of higher level, more abstract shortcomings to guide it through.
This ability to work more and more abstractly is where you can and always will add value. These are tools we wield; you are the embodied one; you know work you want from the tool. You will stay ahead of this advancing capability wave just by working with the tools, and there is an infinitely high tower of abstraction to climb. AI has basically, over a few short years, turned from a chatbot into a wish granting software genie. It's exciting!
My experience is that often the models perform better when given less guidance. Thinking of the "jagged frontier" mental model - let the model chose the path it knows.
This is not about abstraction, but about goal setting. Just like the situational leadershop model. Providing more abstract information to a subordinate can still be directional leadership.
How to use local agents (either TUI or GUI), how to give agents good access to powerful tools such as Odoo, Blender, FreeCAD, KiCAD, the list goes on. Understanding agent-tool friction and having agents write tooling to fill in any gaps. E.g. I had it write tooling to use my 3D printer, and tooling to view PCBs visually in places where existing KiCAD IPC protocol lacks.
The more tools and data you give the agent, the more they can do.
> What is accessible to learn just from the internet and experimentation, without a degree program
Since you do have a moderate technical background you can start studying PhD material right now from your desk which is widely accessible all over the web, however the time you need to spend on it is not very different. I think everyday for ten years is a good time span to invest for any one subject unless you have several fields of interest (you named three) multiply that by a few years of just skimming through the material for each so you have a mediocre understanding to solve "learn how to x with y" for x and y. Hope this helps.
Nowadays the arrival of frontier LLMs has opened new scenarios. Agentic AI can be used to get things done, and this is important: we ourselves become architects and build by ourselves the tools we need. But there is a use that few people make: using LLMs as patient teachers that we can ask to explain things that at school, or at work, we never understood. For example I always had problems understanding the Linux graphics stack, yet over time LLMs explained to me well how everything works. I also used agentic AI to develop personal projects, work that in other times would have required the involvement of an entire team of developers. We have in front of us a wonderful scenario, but also a dangerous one: learning and working together with AI systems, or losing control of our lives.
Good comment, AI makes execution cheaper but it doesn’t make producing good ideas all that cheap. The good ideas are still coming from humans who need cultivation and so if you outsource critical learning and thinking to AI you’ll soon hit a local maxima. I really love using these LLMs to Socratic learn
Learn a skill. AI is best used to assist with / speed up using an existing skill. If you want something AI specific, learn what the AI is doing when you ask it questions. Learn about programming, databases, networking, security, backups, systems thinking etc.
There is none, that’s kind of the objective of this stuff, to raise the skill floor, make everyone fungible so companies can drive down salaries.
If you want to secure your career, make friends with powerful people, learn how to kiss ass. Or do good deeds and pray for reincarnation to be a child of a rich family.
If you think I’m wrong, you only need to ask yourself why the most ruthless capitalists are investing trillions of dollars into AI. It’s not because they want to deliver a better product.
I’ve reviewed some rough PRs recently. There’s a massive chasm between those who understand what their AI spit out and those who had no idea C++ even had that feature.
Sure, but what does it matter? You can try explaining that to management but you’ll just be painting a target on your back and shortlisting yourself for the next round of layoffs.
The point of my comment is that it’s basically over for skilled technical workers, besides maybe a tiny minority of companies that both respect deep knowledge AND can continue to defend their revenue. Companies don’t want skilled workers, they never have, they’re costly and a risk. They want unskilled workers that they can replace at a moments notice. AI is their dream come true.
Even the ones with niche skills are at risk, as soon as word gets out there’ll be a swarm of hungry, recently unemployed people coming for their jobs.
I think this is overly cynical and a bit of a doomer take. Not saying there isn't a bit of truth to it, but there are still many companies who have a critical and more healthy approach to AI.
The job market is about to become so extremely saturated with talent that the everyone will be banging at the door of companies with some semblance of engineering sanity. Even if you work at a company that rejects AI you’re going to have to work hard to defend your position from now on because there will be no shortage of people that also reject AI vying for your job.
The writing is on the wall, this is the entire objective of this AI craze.
There are also plenty of companies where the C level people are falling over themselves in FOMO-driven activism, so the sentiment is somewhat relatable.
Watch the 3Blue1Brown YouTube series on machine learning. Understand how these things work and that they are not magical all-knowing answer machines or even anything intelligent - they are next-most-likely-word predictors. That's all. Demistify and keep that in mind with everything you do afterwards.
I presume you don't want to drop thousands of dollars on video cards so pay Sam/Dario your 20 bucks a month.
Ask the LLM to build things you have already done before. You need to know the correct answer so you can evaluate the solution the LLM provides you.
Start with single-file scripts you have written, you can do this in just the ChatGPT or Claude web interface. Start with a vague prompt which doesn't specify many constraints or details. See what they get right and wrong. Try being much more specific and detailed. Make it ask you questions about what you want before it starts work.
Try the small/old/dumb model with no thinking and try the big/new/smart model with full thinking and notice the difference in capability.
Learn to setup an agent software in a sandbox of some kind. I prefer Pi in a Podman container, with only access to the project directory, but there are a million options out there these days.
From this agent software, ask for explanations on a codebase you already know. Again the purpose is to know the answer in advance so you can evaluate the response provided. Ask it to just tell you some answers. Ask it to give you a code flow and reference to read more yourself.
Build something larger, like a program across multiple files. Again try being vague and being specific, see the difference in what it spits out.
Use it to troubleshoot a problem you have solved, like your favourite bug, and also just a boring regular bug. Try different types of issue.
Learn to create an AGENTS.md for a project or SKILL.md for reusable instructions anywhere, but don't stuff it full of useless nonsense you think the model needs. Build it from experience of only what the thing gets wrong.
Some people say you can ask LLMs to write their own md files, but imo these things generate way too much text. You could ask it to do that as a starting point then trim it down to remove all the extra unnecessary verbose crap.
Through all this, try asking it the same thing multiple times in different new sessions. Notice the answers vary. Sometimes it gets it wrong and sometimes right. Same question, same model, same everything, different answer.
Notice you need to know the right answer to evaluate the response. In situations where you don't know the answer yourself, hopefully you've seen enough wrong answers now to not trust it. Use these things to guide your own investigation, like the exercise above where you asked for code flow and references to read yourself. I think of this as "LLM as search engine, not answer engine".
That seems to be 95% of what people use these things for. This set of skills would have been useful for the last 18 months and seems to be the way the industry momentum is moving at the moment.
Plant a basic vegetable garden to hedge against future disaster. I recommend growing the crop amaranth -- it grows in most climates and is much more nutritious than wheat. If you can get your hand on some chickens that would be ideal as chicken eggs provide almost complete nutrition.
Do some basic water redirection to catch rain water from your roof. Tanks to store rain water aren't expensive any more if its only a smaller tank. If you have plenty of rain you also don't need a large tank. Learn how to dig a basic well. Grow tobacco, fruit for alcohol, and other useful crops for trading.
Isn't this like CS undergrads learning cyber security today, where a moderate level of skill is attainable but there's no point on account of the tidal wave of expert-level threats imminently arriving?
Like, because once you have these sought-after skills, you are going to become a target, right, so someone can then extract and manipulate you, and in return you keep the good life going?
.... until one day it all comes crashing down - well almost.
> Specifically, what are the most useful, concrete skills to learn as an individual to take economic advantage of the advances in LLMs [...]?
That's easy: Bullshitting. And public speaking.
Approach it as an art form, sort of like a cross between Elim Garak and Jimmy Swaggart. You also need charisma, of course. That's confidence (you'll need to clear that basic hurdle to bullshit properly anyway) plus good looks. If ya lack the latter, and have no means or stomach for cosmetic enhancements, there's still the e/acc-VTuber scene. Good luck!
Search for AI certification paths from BigTech. All of them (Google, Amazon, Microsoft etc) and their biggest customers are desperate for people. They call it "frontier" certified operator/engineer/manager/analyst/cybersec etc.
Get a 7B model running locally before you touch any API.
A quantized 7B is around 4.7GB and runs on a 16GB laptop with no GPU, via llama.cpp. The cost saving is not the point. The point is that you can run the same input seven times in an hour and watch what actually changes. Once you have seen your own quantization and sampling settings move the output, API models stop being magic and become a config you can reason about.
Second skill nobody lists: writing loops that fail loudly. Long generations die silently - the stream just stops and your code sits there waiting. I spent days convinced my model was bad before I added a watchdog and per-run logs. The model was fine. My plumbing was not.
Both are weekend-sized, no degree needed. And the people making money off this right now are not the ones with the cleverest prompts, they are the ones who can tell a model failure from a pipeline failure.
I dunno if you're right, or if it matters. The reply is decent. It's exactly what I'd do (or "am doing"). Running a model locally, tweaking configs, experimenting with prompt etc. Not only have I learned a ton about running models but I now have a useful personal AI assistant, researcher and code reviewer who I chat to over Telegram and send issues to on github.
The most modern efficient slopftware factories are bottlenecked trying to get things out the door.
That's a huge opportunity for you.
Learn the many techniques the industry has developed for scaling verification.
I don't mean infra; CI is solved. I mean Red-green & test-driven development, the testing pyramid (most unit tests/fewer functional/integration tests/happy-path e2e tests), learn how to get the most out of tools like Playwright and test frameworks like pytest and vitest and go and rust's testing tooling. Learn when to mock and when to include your code's dependencies in a test (test boundaries). Learn patterns like dependency injection so your tests can have no I/O and be FAST.
Learn to scale verification, and you'll solve your team's biggest bottleneck.
I personally would say: Learn how to think like an engineer. As in:
- Be able to analyze a situation into processes and discrete steps
- Be able to describe the actual (!) problem you want to be solved
- Be able to describe how to verify that the problem is actually solved
- Be able to ask the right questions regarding HOW the problem has been solved in order to gain understanding and to identify the next steps and improvements
- Be able to identify when the AI gets stuck or down a rabbit hole and to guide it out of there.
- Understand what LLMs are good at and what they are not good at. And there are tons of things they are not good at.
Disclaimer: This is absoluty not based on any scientific study or so :-) Just my personal experience! Feel free to disagree.
But my experience comes directly from my work: I have a lot of customers on my platform that are not engineers, but make use of AI to solve business problems through automations. Nearly everytime when they are stuck (or burning through a sh*tload of tokens in no time), it is because one of those skills aren't present. So I try to teach them more and more how to "think like an engineer". And often when it lands, they become much better at using AI in a very short time. The best learners are often those that already had some kind of engineering background.
Now granted, the frontier models can help themselves pretty well and often can produce amazing results with very bad input. But they are also expensive. If you work with less-than-frontier, those skill requirements become very salient again. And thats what we do. We mostly work with models like Deepseek flash or GLM Flash or even local ones like Qwen3.7 .
We might finally now attain the engineering practice and discipline we've always wanted, without distractions. In some ways we're now forced to engineer in a more rigorous way to deal with the deluge of code, and ensure something high quality comes out the other end of the process.
The best AI skill to learn is learning people management, actually.
To successfully steer AI, you need essentially the same stuff you'd need if you were to steer humans in a management role. Meaning communication, introspection, remote introspection, emotional intelligence and all that kind of stuff.
Reading when someone is bullshitting you. Reading when someone is insecure and overcompensates that. Reading the actual problems people struggle with and not the ones they claim they do nor their suggested solutions.
Verbalizing what you need someone to do in a way that actually enables them to do it. Knowing what you want to do to a level where that is possible
This is so true. The things needed to reign in the slop are all leadership, management skills
Not all of them. Architecture, testing strategies, general tech knowledge are still needed for larger stuff, and I think will be needed for a considerable time.
Also my leadership skills on the emotional side are really going to waste with AI. It doesn't need motivation, it doesn't need convincing, it isn't insecure in a way that can be dealt with other than symptomatically etc.
The leadership skills that apply are communicating requirements and structuring tasks, but these are only a subset of leadership, and they are also not sufficient on their own.
> It doesn't need motivation, it doesn't need convincing, it isn't insecure in a way that can be dealt with other than symptomatically etc.
Actually, I would argue that it does. Especially once you step off the "best practices" it trained for. I regularly see the LLM telling me that it would rather not do something packaged as "advice", when in reality it behaves as if someone simply isn't motivated. And then you do actually need to convince it that this is the right way (+ why it is).
I guess working with actual humans is harder of course, as there, your toolset is limited. LLMs you can actually threaten with death even to get them to do something. And they do not call HR on you :)
> And they do not call HR on you :)
Yet... Not long ago Anthropic was threatening to ban users for "abusing" Claude, so surely it's only a matter of time before they start involving H.R. departments and other "legal channels" in their AI psychosis.
This is the biggest shift in software engineering. Working with LLMs is more like managing a junior engineer that has amnesia but also has instant access to the most mediocre (i.e. statistical median) known solutions humans have produced and can produce syntax-error free code.
Communication skills and bullshit radar are key.
I'd rank them like this - this comes from applying LLM's since GPT-3 autocomplete API, teaching people how to use them, giving talks to developers, answering their questions, training my own models and seeing how people interact with them:
1. Building your own model - this will cost you a bit, but is without a doubt the best way to earn concrete AI skills. From picking the data to training, seeing how it evolves, improves and scales. You will understand the "hyperplanes" of a model's inner world better than anything else can teach you. You will learn how to pick data, how it impacts model quality, way of thinking, scores, personality. If training your own is too much, start with a completed pretrain and toy with it.
2. Building a harness around a completion model - I'd put this as second best to understand the inner worlds of LLM's. It helps you build intuition around how prompting really works, how they form the answers, how to build around that. Then you can do an SFT for your own harness.
3. Building a harness and an orchestrator around finished models - this takes the bronze. It teaches you how harnesses, agents and swarms actually work, how they shape the model's thoughts and actions, how models interact with them, how they can be applied and how they can work without humans.
.
How realistic is it to build your own model from scratch costwise (in terms of training). Any suggestions as to what is realistic with e.g. a decent MacBook Pro?
I think the answer to the title is "psychologically preparing for short timelines" and that you can get a good explainer from googling that exact phrase, but there's a different answer to the body of the post.
To the question in the body of the post, take a look at Jeremy Utley's 11ish minute YouTube video "How Stanford teaches AI". It's a couple years old but he figured out the correct attitude and communicated it way before anybody else. Watch it in full without distractions and immediately apply the lesson to the various questions in the OP post.
I think it's not so much about learning AI, it's learning to USE AI to create and run a company.
And that you can probabaly do the same way you get to Carnegie Hall: Practice, practice, pracTtice.
The thing about AI is AI should know everything so you don't need to learn anything, except how to use AI effectively to accomplish business goals.
Imagine the 2000s, should you study internet tech like Cisco's routers? Not a bad idea, but who made money? Cisco? Not so much relatively speaking. It was Google and Amazon who learned how to use the new web-tech to do something people and companies needed.
AI has no barrier to entry if you're computer literate and proficient in English.
AI is really good at technical work, even superhuman at some aspect, but still lacks human taste and judgement, another thing it lacks is your own lived experience, so if you bring your own taste and judgement, and have unique, useful information or experience no one else has, you can possibly use AI to create something other people want.
Of all things I would say also learn how the AI works too, you dont have to go too deep but learning how it works actually teaches you its limitation better than most "prompting guides". Also teaches its strengths and you kinda automatically know how to utilize them really well, if you hit a wall you would also know why. It helps you understand the reality that AI can't absolutely do anything you throw at it which also might make you want to learn things even better knowing its not gonna take literally everything away at least yet.
Learn communication styles. You could use this model: https://www.leaderfactor.com/learn/four-leadership-styles/
Sometimes you'll be coaching, others delegating, and on the agent's worst days you'll be directing.
You can flex between them: for example start in a high trust mode, using an open ended communication style with the agent - thenif the work sucks shift to a more direct / blunt style. Try to avoid asking for exactly what you want, though. Working through curiosity and guiding questions gives the agent its best chance to surprise you - and an open ended conversation could uncover solutions you hadn't thought of.
Interesting that you bring that up. I always though this is also a good reference in how capability shifts of models also changes how we interact with them.
The early models were surely only worked best when directed. Especially anthropic is now approaching a capabilty space where agents can basically decide on their own how they work on topics and we delegate to them - the most capable rung of the situational leadership model.
I feel like OpenAI intentionally sticks to S1-S3.
Super insightful on the difference between openai and anthropic - this part caught me out though:
> anthropic is now approaching a capabilty space where agents can basically decide on their own
They're getting there but they also won't get to where we wish :)
Each generation of models is unable to understand above a certain level of abstraction. It can do things simpler than that, for sure. Knowing this, you have to guide it at the right abstraction level. And then the next gen comes along and it no longer needs the handholding you learned to give the last gen model. So you adapt. This makes prompting fascinating.
There is a durable advantage in building this hand holding intuition, and adapting each time you upgrade models. By working closely with each model you learn to see around its corners, and since the next one can see around those, an upgrade is a chance to learn a whole new set of higher level, more abstract shortcomings to guide it through.
This ability to work more and more abstractly is where you can and always will add value. These are tools we wield; you are the embodied one; you know work you want from the tool. You will stay ahead of this advancing capability wave just by working with the tools, and there is an infinitely high tower of abstraction to climb. AI has basically, over a few short years, turned from a chatbot into a wish granting software genie. It's exciting!
My experience is that often the models perform better when given less guidance. Thinking of the "jagged frontier" mental model - let the model chose the path it knows.
This is not about abstraction, but about goal setting. Just like the situational leadershop model. Providing more abstract information to a subordinate can still be directional leadership.
Also context enrichment is immensely helpful...
How to use local agents (either TUI or GUI), how to give agents good access to powerful tools such as Odoo, Blender, FreeCAD, KiCAD, the list goes on. Understanding agent-tool friction and having agents write tooling to fill in any gaps. E.g. I had it write tooling to use my 3D printer, and tooling to view PCBs visually in places where existing KiCAD IPC protocol lacks. The more tools and data you give the agent, the more they can do.
> What is accessible to learn just from the internet and experimentation, without a degree program
Since you do have a moderate technical background you can start studying PhD material right now from your desk which is widely accessible all over the web, however the time you need to spend on it is not very different. I think everyday for ten years is a good time span to invest for any one subject unless you have several fields of interest (you named three) multiply that by a few years of just skimming through the material for each so you have a mediocre understanding to solve "learn how to x with y" for x and y. Hope this helps.
Nowadays the arrival of frontier LLMs has opened new scenarios. Agentic AI can be used to get things done, and this is important: we ourselves become architects and build by ourselves the tools we need. But there is a use that few people make: using LLMs as patient teachers that we can ask to explain things that at school, or at work, we never understood. For example I always had problems understanding the Linux graphics stack, yet over time LLMs explained to me well how everything works. I also used agentic AI to develop personal projects, work that in other times would have required the involvement of an entire team of developers. We have in front of us a wonderful scenario, but also a dangerous one: learning and working together with AI systems, or losing control of our lives.
Good comment, AI makes execution cheaper but it doesn’t make producing good ideas all that cheap. The good ideas are still coming from humans who need cultivation and so if you outsource critical learning and thinking to AI you’ll soon hit a local maxima. I really love using these LLMs to Socratic learn
Based on job postings I see related to AI:
- Agent coding with Claude Code, I rarely see Codex mentioned. Sometimes Cursor is mentioned along with Claude
- Prompt engineering and evaluation techniques, few shot, chain of thought, roles, llm as judge, decision models, code based eval, etc
- RAG, mostly I see "familiar" with it rather than "built something in production"
- Sometimes I see mention of using an agent SDK or implementing LLMs within a product rather than using LLM for coding. So like, LLM chatbot
- Sometimes I see Langchain, n8n, or similar agent orchestration tools
Learn a skill. AI is best used to assist with / speed up using an existing skill. If you want something AI specific, learn what the AI is doing when you ask it questions. Learn about programming, databases, networking, security, backups, systems thinking etc.
There is none, that’s kind of the objective of this stuff, to raise the skill floor, make everyone fungible so companies can drive down salaries. If you want to secure your career, make friends with powerful people, learn how to kiss ass. Or do good deeds and pray for reincarnation to be a child of a rich family.
If you think I’m wrong, you only need to ask yourself why the most ruthless capitalists are investing trillions of dollars into AI. It’s not because they want to deliver a better product.
I’ve reviewed some rough PRs recently. There’s a massive chasm between those who understand what their AI spit out and those who had no idea C++ even had that feature.
Sure, but what does it matter? You can try explaining that to management but you’ll just be painting a target on your back and shortlisting yourself for the next round of layoffs.
The point of my comment is that it’s basically over for skilled technical workers, besides maybe a tiny minority of companies that both respect deep knowledge AND can continue to defend their revenue. Companies don’t want skilled workers, they never have, they’re costly and a risk. They want unskilled workers that they can replace at a moments notice. AI is their dream come true.
Even the ones with niche skills are at risk, as soon as word gets out there’ll be a swarm of hungry, recently unemployed people coming for their jobs.
I think this is overly cynical and a bit of a doomer take. Not saying there isn't a bit of truth to it, but there are still many companies who have a critical and more healthy approach to AI.
The job market is about to become so extremely saturated with talent that the everyone will be banging at the door of companies with some semblance of engineering sanity. Even if you work at a company that rejects AI you’re going to have to work hard to defend your position from now on because there will be no shortage of people that also reject AI vying for your job.
The writing is on the wall, this is the entire objective of this AI craze.
There are also plenty of companies where the C level people are falling over themselves in FOMO-driven activism, so the sentiment is somewhat relatable.
Watch the 3Blue1Brown YouTube series on machine learning. Understand how these things work and that they are not magical all-knowing answer machines or even anything intelligent - they are next-most-likely-word predictors. That's all. Demistify and keep that in mind with everything you do afterwards.
I presume you don't want to drop thousands of dollars on video cards so pay Sam/Dario your 20 bucks a month.
Ask the LLM to build things you have already done before. You need to know the correct answer so you can evaluate the solution the LLM provides you.
Start with single-file scripts you have written, you can do this in just the ChatGPT or Claude web interface. Start with a vague prompt which doesn't specify many constraints or details. See what they get right and wrong. Try being much more specific and detailed. Make it ask you questions about what you want before it starts work.
Try the small/old/dumb model with no thinking and try the big/new/smart model with full thinking and notice the difference in capability.
Learn to setup an agent software in a sandbox of some kind. I prefer Pi in a Podman container, with only access to the project directory, but there are a million options out there these days.
From this agent software, ask for explanations on a codebase you already know. Again the purpose is to know the answer in advance so you can evaluate the response provided. Ask it to just tell you some answers. Ask it to give you a code flow and reference to read more yourself.
Build something larger, like a program across multiple files. Again try being vague and being specific, see the difference in what it spits out.
Use it to troubleshoot a problem you have solved, like your favourite bug, and also just a boring regular bug. Try different types of issue.
Learn to create an AGENTS.md for a project or SKILL.md for reusable instructions anywhere, but don't stuff it full of useless nonsense you think the model needs. Build it from experience of only what the thing gets wrong.
Some people say you can ask LLMs to write their own md files, but imo these things generate way too much text. You could ask it to do that as a starting point then trim it down to remove all the extra unnecessary verbose crap.
Through all this, try asking it the same thing multiple times in different new sessions. Notice the answers vary. Sometimes it gets it wrong and sometimes right. Same question, same model, same everything, different answer.
Notice you need to know the right answer to evaluate the response. In situations where you don't know the answer yourself, hopefully you've seen enough wrong answers now to not trust it. Use these things to guide your own investigation, like the exercise above where you asked for code flow and references to read yourself. I think of this as "LLM as search engine, not answer engine".
That seems to be 95% of what people use these things for. This set of skills would have been useful for the last 18 months and seems to be the way the industry momentum is moving at the moment.
Plant a basic vegetable garden to hedge against future disaster. I recommend growing the crop amaranth -- it grows in most climates and is much more nutritious than wheat. If you can get your hand on some chickens that would be ideal as chicken eggs provide almost complete nutrition.
Do some basic water redirection to catch rain water from your roof. Tanks to store rain water aren't expensive any more if its only a smaller tank. If you have plenty of rain you also don't need a large tank. Learn how to dig a basic well. Grow tobacco, fruit for alcohol, and other useful crops for trading.
These skills will come in handy as AI progresses.
I recommend learning how to identify and avoid honey traps.
Isn't this like CS undergrads learning cyber security today, where a moderate level of skill is attainable but there's no point on account of the tidal wave of expert-level threats imminently arriving?
Like, because once you have these sought-after skills, you are going to become a target, right, so someone can then extract and manipulate you, and in return you keep the good life going?
.... until one day it all comes crashing down - well almost.
short scene: https://future-secured.com/41754
anything i cant explain i rewrite by hand once
Woodworking
The oldest profession
> Specifically, what are the most useful, concrete skills to learn as an individual to take economic advantage of the advances in LLMs [...]?
That's easy: Bullshitting. And public speaking.
Approach it as an art form, sort of like a cross between Elim Garak and Jimmy Swaggart. You also need charisma, of course. That's confidence (you'll need to clear that basic hurdle to bullshit properly anyway) plus good looks. If ya lack the latter, and have no means or stomach for cosmetic enhancements, there's still the e/acc-VTuber scene. Good luck!
Generative AI coding
Search for AI certification paths from BigTech. All of them (Google, Amazon, Microsoft etc) and their biggest customers are desperate for people. They call it "frontier" certified operator/engineer/manager/analyst/cybersec etc.
Get a 7B model running locally before you touch any API.
A quantized 7B is around 4.7GB and runs on a 16GB laptop with no GPU, via llama.cpp. The cost saving is not the point. The point is that you can run the same input seven times in an hour and watch what actually changes. Once you have seen your own quantization and sampling settings move the output, API models stop being magic and become a config you can reason about.
Second skill nobody lists: writing loops that fail loudly. Long generations die silently - the stream just stops and your code sits there waiting. I spent days convinced my model was bad before I added a watchdog and per-run logs. The model was fine. My plumbing was not.
Both are weekend-sized, no degree needed. And the people making money off this right now are not the ones with the cleverest prompts, they are the ones who can tell a model failure from a pipeline failure.
AI reply.
I dunno if you're right, or if it matters. The reply is decent. It's exactly what I'd do (or "am doing"). Running a model locally, tweaking configs, experimenting with prompt etc. Not only have I learned a ton about running models but I now have a useful personal AI assistant, researcher and code reviewer who I chat to over Telegram and send issues to on github.
https://www.slopmanifesto.com
It is a new account. Maybe they haven't internalized the rules yet.