y42

Sire, wouldst thou perchance elaborate upon this ancient, forgotten technique, long buried in the mists of antiquity and shrouded in mysterious lore?

Still some way to go until we run "it" on our own machines. Even with current "mid to high end" hardware you barely reach the same quality and speed with a free LLM compared to what Opus or Astra delivers. In numbers: A 3k Euro Macbook with 32 GB RAM allows you to run a (hard guess only) up to 16 GB Model. Alone! Not speaking about sub agents here. Because you still need memory for your everday work and other programs, cache, context etc.

From my experience, there is no everlasting digital system, no matter how many fallbacks you add. So I wonder, how can you ensure, this still works after e. g. 1 year?

Yea.."art performance" kind of already catches it. I want to see what people can do with just HTML, JavaScript and CSS. Like back in the days, withouth subscriptions, algorithms, bloat.

It certainly sounds interesting, but then I wonder: After a couple of days or maybe weeks, even with a sophisticated knowledge management strategy, the knowledge of Wilmund is so big, that there won' t be much room left in the context window. Right?

The interesting part is this:

"After losing a bunch of customers to Codex"

Hot take: Right now it's quite simple to loose customers, like customer churn from A to B or back. That's the weakest spot, isn't it? No matter how complex the products grown and how deep they are integrated into our systems, common users can easily switch. Even heavy users, I argue. Just tell Codex to rewrite the existing Claude-instructions into their own. Even that may not be necessary, if you organized your work "agent agnostic".

And I dont see how that could change, that's why they try to offer tools that are even more integrated into our lives. Like "dots". But this also narrows down the use cases and client base, I argue.

When working with a whole bunch of sub-agents, like you probably would do when coding, 2-3 sessions at a time are probably not enough. Like I build very specific sub-agents for code review, ui review and so on. Within one project I work with at least 3 - 4 agents then. Using Opus on every task would not leave room for other everyday-tasks.

good to have a quick overview, though I am missing some kind of depth, like

>There are others — MLX-native runtimes, llama.cpp directly — but Ollama and LM Studio cover 95% of people reading this.

Sure, but... MLX native runtime have a lot of advantages right? and they are not necessarily more complicated like eg ollama. It's worth to mention them, too.

Coming to the next missing part: use cases.

Running it locally competes with dozens of freely available chat bots. And for most simple parts, seriously, you don't need that sophisticated payed AI service. But then: how do local model's perform exactly here: let's start with coding, image generation and more? Like stuff that I probably really want to run locally.

I don't know.... With all those approaches to run AI in your own premise, people need their own hardware. It makes them independent. Of course hyperscalers and AI labs don't like that, but I argue that consumers prefer independence over some costs of operations. If this would be the case, people wouldn't own cars.

I am quite familiar with Resolve..I am not a pro, and it's probably totally overload for my purpose, but I think it's easy to learn and even quite intuitive

I tried many others, but mostly all others are not stable enough or have an even worse UI/UX... so I wonder... is this different?

No rockets, no emoticons? Not real LI cringe.

Dont want to crash the party here, but I am still sceptic about all those on-premise-llm-approaches.

I think we strongly need something like that (shameless plug, I tried to build something around bitNet for the same reason: https://github.com/nickyreinert/bitNetRTR).

But at the end, all aproaches I saw, however genius they are: the actual results are always a mess. It's a better chat buddy, nothing else. It's e.g. far away from an decent coding assistants. I fine tuned Gemma with domain specific knowledge. Running it on a 16GB VRM GForce. Even then it's okai'sh but far way from a mind blowing experience. I ran some of the promised open source model on my 36GB MBPro M3, in Pi, Hermes, Continue. Can't compare the results to what Claude or Codex are offering.

You need at least something that's far away from consumer hardware, like those 7k'ish GForce machines with 96GB VRAM to get an idea of a good competitive model.

But... please, proof me wrong! =)

shameles self promotion: I posted similar thing a couple of days ago:

https://news.ycombinator.com/item?id=48773805

https://9.1-1-1.de

the threshold is not low, I would argue, it's just companies like to exploit it.

(I just wrote an extensive analysis about that https://nickyreinert.de/2026/2026-05-28-google-maps-reviews/

Security by Obviousness.

shameless plug, it's part of the Nerd Encyclopedia: it's also called "nasal demons".

https://nickyreinert.de/2023/2023-05-16-nerd-enzyklop%C3%A4d...

Having an M3 with 36 GByte I was under the assumption, that I can utilize like Qwen and similar models. It's quite easy to set up, you can use pi or hermes for CLI access, or "Continue" to use it in VS Code. You can choose between omlx, Ollama and even more to run the model itself. It's no rocket science, but the results are also not satisfying.

I use it occassionally for very easy tasks, fix typos or update meta data in blog posts. So yeah, it improves productivity. But coding-wise it's far away from Codex, Claude et al.

Shameless plug: It's not that fancy but follows the same principle:

https://institut-fdh.de/?2026-aya

There's also this well known page which does the exact same thing in a more ordered way:

https://browserleaks.com/

First: if you compare rates (salary, wages,...) you also always must consider rents, cost of living etc.

Second: You can't just pick Berlin for comparison.

Third: Take away foreign owend companies from Berlin, you get a cheap, dirty poor capital.

Most comments already explained it, however:

No! It's a common misunderstanding (imho) that software development is about "writing code" or knowing a programming language's syntax. It never was, it never will be. I experienced this misunderstanding:

I never studied CS or IT or anything related. All my skills are self taught, starting when I was a litle boy. When I was young, I had problems getting into this busines, because HR and recruiters where tasking me with "Write this <x> using this <y>". They wanted to see A grades from 6 years of a master in CS.

Well. I am not that guy that memorizes syntax or coding rules. My approach is very different - and so are many, I assume. And this is what software development is about: You get a problem and you find a way to solve it. At many stations in my professional career I was able to proof that. I "mastered" Pascal from the scratch, on the job. Same goes for VBA, Shell scripting, Python, PHP, JavaScript... you name it.

It may sound arrogant - but then you misunderstood what I am trying to say: The language, the syntax, the coding rules are just "tools" for me. I don't care what I use. I focus on the problem.

Isn't it the same as "Write this <x> using this <y>" you may ask? No. Because this expects a specific solution. Not saying this will not lead to a result. It just a very narrow view on the solution.

There are good software developers that can answer those questions "Write this <x> using this <y>". That's not the point. The point is: There are also great software developers who can't.

With AI you don't have to worry to memorize syntax or rules. You can focus on the problem solving process.

So, AI is changing how software will be created. It make it more efficient. And it changes how HR and recruites hopefully will look at this profession. It's not about A grades in a CS class.

I want to believe. A couple of weeks ago I fell into this "trap", they offered a similar thing. I subscribed to the Pro Plan. Had fun for a couple of weeks and then I entered frustration phase. I love the product, but I hate those up and downs. My rant made it to HN front page - which I am not happy of. I want the stuff I build to be seen on the front page.

shameless plug:

A series of Jupyter notebooks explaining the whole machine learning mechanism, from the beginning

https://github.com/nickyreinert/DeepLearning-with-PyTorch-fr...

and of course also how to build an llm from scratch

https://github.com/nickyreinert/basic-llm-with-pytorch/blob/...