> "LLMs expose the literacy gap between great and mediocre coders. If you can't see how bad the generated code is, LLMs will only make you not see it faster."

I wholeheartedly believe that if the general trend in SWE goes without hiring new grads/self-taught people (who can still use AI-assistance in their work), we're gonna be seeing huge outsourcing to AIaaS companies and generally become tremendously dependent on them when existing pool of seasoned programmers goes extinct and the work doesn't intrigue training new people.

Although there is cheap, open source models running locally or available at scale on reasonable prices without compromising programmer integrity.

Genuine question: Will you also reject the results and fixes regarding the current PHP CVEs until there are handcrafted versions? And - a question from another perspective - would you be open to using Mistral in a self hosted environment?

Good article.

I don't use LLMs in my personal work too. The reasons are like yours.

If you care about your privacy, security, skills, personal feelings and personal growth, you mustn't use AI or LLMs.

I work on what should be pretty simple things (3D models of projects for CNC machines) and based on flet.dev recently announcing 1.0 and a quick test one-shotting a convincing protototype of a drawing program description which heretofore every such system had failed at miserably, decided to proceed with it.

I got a quick prototype of a rabbet box generator:

https://studio.flet.dev/run/bicQWc3Aul/

but the geometry was all wrong and ultimately required that I describe the calculation of the position and dimensions of each element in exhaustive detail in order to get a usable DXF (the wireframe preview is still all messed up) which eventually worked to make a box:

https://community.carbide3d.com/uploads/default/original/3X/...

I was also drawing the design up in an interactive CAD/CAM program, and modeling it in BlockSCAD/OpenSCAD (and dropping into BelfrySCAD to figure out rotation):

https://www.blockscad3d.com/community/projects/2169316

(and speaking of rotation, at this time, the sliding and hinged lid options open on opposite sides in the preview...)

Since then, I've made a very simple generator which simple does a box decorated with lines, arcs, and circles on each side --- my intention is to dis-assemble that and use it to write a Domain Specific Language implementation for creating such designs and previewing them and writing out DXFs --- which is a long-winded way of saying I've been finding this tooling frustrating to use, but potentially useful to make rough drafts of programming templates which can then be used as a starting point for more traditional coding.

Eventually all of this will go into:

https://github.com/WillAdams/gcodepreview

I’m not disagreeing with the author, but if you look at his dealbreakers: most of them will be resolved over time (quality, maintainability, environmental impact & even politics), they are not so much against LLMs as they are against the current versions of them.

The only arguments against AI itself are about cognitive decline and “I like coding” While there is nothing wrong with those preferences, it’s akin to saying “I like building furniture with only hand tools, and not doing it by hand leads to decline in your ability.”

> I just personally feel squeamish about having my work being done by rather skeezy companies like OpenAI, Anthropic, or, god forbid, Cursor/SpaceX.

Local models exist. Your data doesn't have to leave your computer.

I've struck the balance of using it at work but not at home.

At work I still read through every line in the diff (unless it's a personal util/non-production facing/critical etc). Even with this I feel less 'in touch' with my work code than I do with my home code.

> Entrepreneur at heart

The post sounds more like a programmer at heart, and that's totally fine.

To me, being an entrepreneur is more about building products, solving problems, and creating business value. LLMs help to deliver bug fixes, new features, and entirely new products faster.

Disclaimer: I use LLMs for support and deep dives into existing code, not for creating new substantial code that I intend to publish.

Quoting from the article:

> With these subsidized token prices currently, I bet we're seeing less and less AI usage encouragement in companies in the future.

I see a different trend, and for a different reason. Employees are encouraged to use open-weight LLMs hosted on-premise, or, at least, in the "correct" geographical area, in preference to models from big providers. The motivation is mostly compliance with GDPR and NDAs. And when this happens, complaints like this get ignored:

> Your local model wrote a wrong root cause analysis for the customer ticket. It's a broken tool, worse than nothing. Claude did it right. We really need either access to frontier models or, alternatively, reasonable expectations on your side about what is doable with the tools you provide.

And that's bad.

I'm not writing a blog post just for this, but the man is right. For me, the only thing I expecte from a LLM is a StackOverflow/Google that actually works and maybe some static code analysis.