It took me a few months of working with the agents to get really productive with it. The gains are significant. I write highly detailed specs (equiv multiple A4 pages) in markdown and dicate the agent hierarchy (which agent does what, who reports to who).
I've learned a lot of new things this year thanks to AI. It's true that the low levels skills with atrophy. The high level skills will grow though; my learning rate is the same, just at a much higher abstraction level; thus covering more subjects.
The main concern is the centralisation. The value I can get out of this thing currently well exceeds my income. AI companies are buying up all the chips. I worry we'll get something like the housing market where AI will be about 50% of our income.
We have to fight this centralisation at all costs!
This is something I think a lot of people don't seem to notice, or worry about, the moving of programming as a local task, to one that is controlled by big corporations, essentially turning programming into a subscription model, just like everything else, if you don't pay the subscription you will no longer be able to code i.e. PaaS (Programming as a Service). Obviously at the moment most programmers can still code without LLMs, but when autocomplete IDEs became main stream, it didn't take long before a large proportion of programmers couldn't program without an autocomplete IDE, I expect most new programmers coming in won't be able to "program" without a remote LLM.
That ignores the possibility that local inference gets good enough to run without a subscription on reasonably priced hardware.
I don't think that's too far away. Anthropic, OpenAI, etc. are pushing the idea that you need a subscription but if opensource tools get good enough they could easily become an expensive irrelivance.
There is that, but the way this usually works is that there is always a better closed service you have to pay for, and we see that with LLMs as well. Plus there is the fact that you currently need a very powerful machine to run these models at anywhere near the speed of the PaaS systems, and I'm not convinced we'll be able to do the Moore's law style jumps required to get that level of performance locally, not to mention the massive energy requirements, you can only go so small, and we are getting pretty close to the limit. Perhaps I'm wrong, but we don't see the jumps in processing power we used to see in the 80s and 90s, due to clock speed jumps, the clock speed of most CPUs has stayed pretty much the same for a long time. As LLMs are essentially probabilistic in nature, this does open up options not available to current deterministic CPU designs, so that might be an avenue which gets exploited to bring this to local development.
Local inference is already very good on open models if you have the hardware for it.
My concern is that inference hardware is becoming more and more specialized and datacenter-only. It won’t be possible any longer to just throw in a beefy GPU (in fact we’re already past that point).
Yep, good point. If they don't make the hardware available for personal use, then we wouldn't be able to buy it even it could be used in a personal system.
This is the most valid criticism. Theoretically in several years we may be able to run Opus quality coding models locally. If that doesn't happen then yes, it becomes a pay to play profession - which is not great.
I have found that using more REPLs and doing leetcodes/katas prevents the atrophy to be honest.
In fact, I'd say I code even better since I started doing one hour per day of a mixture of fun coding and algo quizzes while at work I mostly focus on writing a requirements plan and implementation plan later and then letting the AI cook while I review all the output multiple times from multiple angles.
The hardware needs to catch up I think. I asked ChatGPT (lol) how much it would cost to build a Deepseek server that runs at a reasonable speed and it quoted ~400k-800k(8-16 H100 + the rest of the server).
Guess we are still in the 1970s era of AI computing. We need to hope for a few more step changes or some breakthrough on model size.
The problem is that Moore's law is dead, silicon isn't advancing as fast as what we've envisioned in the past, we're experiencing all sorts of quantum tunneling effects in order to cram as much microstructure as possible into silicon, and R&D for manufacturing these chips are climbing at a rapid rate. There's a limit to how we can fight against Physics, and unless we discover a totally new paradigm to alleviate this issues (ex. optical computing?) we're going to experience diminishing returns at the end of the sigmoid-like tech advancement cycle.
You can run most open models (excluding kimi-k2) on hardware that costs anywhere from 45 - 85k (tbf, specced before the vram wars of late 2025 so +10k maybe?). 4-8 PRO6000s + all the other bits and pieces gives you a machine that you can host locally and run very capable models, at several quants (glm4.7, minimax2.1, devstral, dsv3, gpt-oss-120b, qwens, etc.), with enough speed and parallel sessions for a small team (of agents or humans).
[flagged]
Well, if you're programming without AI you need to understand what you're building too, lest you program yourself into a corner. Taking 3-5 minutes to speech-to-text an overview of why you want to build what exactly, using which general philosophies/tool seems like it should cost you almost zero extra time and brainpower