Civic technologist building AI tools that help systems explain themselves. Clew Suite builder, NACM 2026 speaker. I write about hackathons, product judgment, agent readiness, and learning in public.
Thanks for this! I have been struggling to think of a plan that's best for me and this is very helpful guidance. It's so easy to get lost with all there is out there and overwhelmed, too! But your staged plan makes it less so.
Hi,
immediately caught by the title, since i have been working on a similar post about ai and how i try to handle it.
It is currently on hold, i am very busy and i personaly need some more time to process it.
I am currently trying to handle all this AI hype and try to find my way. Tech changes too fast to process it in the way we used to do.
I am not very optimistic. Try not to be pessimistic.
I write $0-stack engineering guides: running local LLMs (DeepSeek, Qwen, Llama) on consumer laptops, and shipping full-stack apps on free tiers. Everything benchmarked on real hardware.
Great framework, especially around understanding underlying systems instead of just consuming higher-level APIs.
One concrete way I've stayed sharp with AI is running and tuning models locally (via Ollama/llama.cpp) rather than treating AI purely as a cloud prompt box. When you actually see how KV cache scales, why context window growth eats VRAM non-linearly, and how GPU offloading splits layers across memory buses, you understand what AI tools can and cannot reliably do in production.
For anyone experimenting on consumer hardware, keeping context windows lean (e.g., num_ctx: 4096) and monitoring layer offloading gives you surprisingly capable performance on zero monthly subscription burn.
How much of your day-to-day workflow are you running through local tooling vs cloud LLMs?
That is a great point I didn't mention in this post but I have been more deliberate about my choice of including open models I can run locally vs cloud LLMs. This has pros and cons over privacy, performance, transparency as you mention.
Your T-shaped split is the part I want to copy. Spending 70 to 80 percent of learning time deepening one domain while still experimenting broadly is a much clearer rule than just staying curious, and fintech is a strong vertical for it. The ADR habit stood out too, since writing down the why is exactly what keeps AI output reviewable instead of just fast. And the question about whether you would sign your name on the AI work is a great checkpoint. I am going to borrow that one before sharing anything generated.
The obsession with wrapping remote APIs into multi-layered agent abstractions is creating massive hidden latency and failure cascades in production.
If you profile real-world workloads, bare-metal local inference (via direct C++ / llama.cpp or vLLM bindings) consistently beats fragile cloud wrappers on both predictability and cost. Tooling should eliminate abstraction penalties, not compound them.
I write $0-stack engineering guides: running local LLMs (DeepSeek, Qwen, Llama) on consumer laptops, and shipping full-stack apps on free tiers. Everything benchmarked on real hardware.
The transition from business/finance to software is actually the exact skillset that AI commoditizes least.
Writing syntax, scaffolding boilerplate, and prompting models is trivial now. What models fail at is evaluating whether a piece of software is actually solving a commercial problem that someone will pay for versus 'selling furniture to carpenters'.
Engineers who understand unit economics, conversion bottlenecks, and domain friction of real businesses will always run circles around pure syntax typists.
Hey, I'm Hema π
Developer, writer, and creator of Dev Opportunity Radar, a weekly series published every Friday on DEV, helping people discover opportunities they might otherwise miss.
I think I came across this article at the right time, Julien. I was actually feeling a little overwhelmed lately, thinking about how things are changing and what I should be focusing on. Reading through your approach gave me some guidance and helped me think about things a little differently. I think I needed to read this. Thank you for sharing π
Nice to hear from you Hemapriya! It does feel like a rollercoaster at the moment. I am glad you found some guidance in my post. If there is anything you need more clarity and advice on, feel free to reach out to me :)
I write $0-stack engineering guides: running local LLMs (DeepSeek, Qwen, Llama) on consumer laptops, and shipping full-stack apps on free tiers. Everything benchmarked on real hardware.
Strongly agree with your emphasis on fundamentals over prompt engineering.
One practical shift that gave me massive clarity was moving away from cloud AI black-boxes and learning to run/quantize models locally (llama.cpp, Ollama, GGUF).
When you understand where memory bandwidth actually chokes (KV-cache footprint vs layer weights), prompt evaluation vs token generation, and context truncation, you suddenly see through 90% of the AI hype cycle. You realize these aren't magical thinking enginesβthey are memory-bandwidth-bound matrix multiplication graphs.
That foundation completely changes how you build software on top of them, whether you're designing local agents or reducing cloud API bills.
Thanks for this! I have been struggling to think of a plan that's best for me and this is very helpful guidance. It's so easy to get lost with all there is out there and overwhelmed, too! But your staged plan makes it less so.
Thanks! Glad to hear.
I wanted to write out and share my plan in the most practical way possible.
Hi,
immediately caught by the title, since i have been working on a similar post about ai and how i try to handle it.
It is currently on hold, i am very busy and i personaly need some more time to process it.
I am currently trying to handle all this AI hype and try to find my way. Tech changes too fast to process it in the way we used to do.
I am not very optimistic. Try not to be pessimistic.
happy to read your post and know about you
Hi, happy to connect with you.
I look forward to reading your post and learning from your plan, when you release it.
That's a smart, well-thought-out plan. Thanks for sharing!
Thanks Sylwia! Your post from the other week actually inspired me to share my plan here.
Great framework, especially around understanding underlying systems instead of just consuming higher-level APIs.
One concrete way I've stayed sharp with AI is running and tuning models locally (via Ollama/llama.cpp) rather than treating AI purely as a cloud prompt box. When you actually see how KV cache scales, why context window growth eats VRAM non-linearly, and how GPU offloading splits layers across memory buses, you understand what AI tools can and cannot reliably do in production.
For anyone experimenting on consumer hardware, keeping context windows lean (e.g.,
num_ctx: 4096) and monitoring layer offloading gives you surprisingly capable performance on zero monthly subscription burn.How much of your day-to-day workflow are you running through local tooling vs cloud LLMs?
That is a great point I didn't mention in this post but I have been more deliberate about my choice of including open models I can run locally vs cloud LLMs. This has pros and cons over privacy, performance, transparency as you mention.
Your T-shaped split is the part I want to copy. Spending 70 to 80 percent of learning time deepening one domain while still experimenting broadly is a much clearer rule than just staying curious, and fintech is a strong vertical for it. The ADR habit stood out too, since writing down the why is exactly what keeps AI output reviewable instead of just fast. And the question about whether you would sign your name on the AI work is a great checkpoint. I am going to borrow that one before sharing anything generated.
Nice to see AI comment and acknowledge my plan to beat AI ;)
The obsession with wrapping remote APIs into multi-layered agent abstractions is creating massive hidden latency and failure cascades in production.
If you profile real-world workloads, bare-metal local inference (via direct C++ / llama.cpp or vLLM bindings) consistently beats fragile cloud wrappers on both predictability and cost. Tooling should eliminate abstraction penalties, not compound them.
This does not seem relevant to the topic I posted about :)
The transition from business/finance to software is actually the exact skillset that AI commoditizes least.
Writing syntax, scaffolding boilerplate, and prompting models is trivial now. What models fail at is evaluating whether a piece of software is actually solving a commercial problem that someone will pay for versus 'selling furniture to carpenters'.
Engineers who understand unit economics, conversion bottlenecks, and domain friction of real businesses will always run circles around pure syntax typists.
Speaking business and product language and knowing how this translates well into technical requirements definitely can be an edge today.
I think I came across this article at the right time, Julien. I was actually feeling a little overwhelmed lately, thinking about how things are changing and what I should be focusing on. Reading through your approach gave me some guidance and helped me think about things a little differently. I think I needed to read this. Thank you for sharing π
Nice to hear from you Hemapriya! It does feel like a rollercoaster at the moment. I am glad you found some guidance in my post. If there is anything you need more clarity and advice on, feel free to reach out to me :)
Strongly agree with your emphasis on fundamentals over prompt engineering.
One practical shift that gave me massive clarity was moving away from cloud AI black-boxes and learning to run/quantize models locally (llama.cpp, Ollama, GGUF).
When you understand where memory bandwidth actually chokes (KV-cache footprint vs layer weights), prompt evaluation vs token generation, and context truncation, you suddenly see through 90% of the AI hype cycle. You realize these aren't magical thinking enginesβthey are memory-bandwidth-bound matrix multiplication graphs.
That foundation completely changes how you build software on top of them, whether you're designing local agents or reducing cloud API bills.
Good point. I am trying to be more deliberate on my choice of open models vs closed.
reassuring, thanks coach.
You are welcome. Do you also follow a plan?
I follow email subscriptions to see what's happening in the ai space . but I try to avoid ai psychosis the best I can thru a good ole book.lol
Agree, itβs hard to beat a good book!
Any recent recommendations?
Very thoughtful - well-considered and detailed "survival plan" - bookmarked this!
Thanks leob! I am glad you found this plan useful in some way.
It is indeed a survival plan, hanging on :D