The issue is anything that is of value requires some level of detail, of complexity, and that is only of interest to people that know that specific complexity, and it is a pain point for them. Now they'll care. Everyone else? Lost them. So, the marketing challenge is to find some aspirational complexity that people wish they knew, and how that can be solved with AI, and without turning that thing into a trivial nuisance, but a valued skill. That logical series right there is, well, too much for far too many.

When LLMs and ChatGPT first came out, it struck me as obvious and dangerous to a deep thinker or a knowledge worker the answering capacity. So, from my initial use I did not ask them questions, I have always "done my own work" and then asked the LLMs to criticize that work. This has been an exponential ladder of learning, and my cognitive growth is personally noticeable. I'm not hesitating to scribble out calculus and work it out, as I need for my work, where in the past I'd have found some other way because I felt uncomfortable with my tip-of-my-tongue calc skills. Don't ask AI, do your own work and ask for criticism, and them improve your own work yourself. This creates a learning ladder that you will climb.

That's nice except when you work somewhere where more and more developers are pushed to pump out slop generated by AI as fast as possible. So far I am not there yet but I have plenty of friends in the industry who are basically 'not allowed' to code manually anymore.

It was never safe, he's exposed the system's design was never intended to be safe for anyone but those in charge.

This. People are locked in their 401k and penalized when taking it out says a lot about what it is.

People should honestly read Killing the Sacred Cows it’s an eye-opener for anyone invested in 401k.

A "401(k)" is not a monolithic entity. In practice, most employers offer a choice of funds, with the most popular being a year-targeted fund that rebalances between equities and bonds as you get closer to retirement. Having said that, you can probably dump your entire portfolio into government bonds, small cap stocks, or euro futures.

> …says a lot about what it is

Doesn’t it say that it’s a retirement fund, intended to be saved until retirement age? The 10% penalty is little more than a wrist slap level deterrent, too. It’s usually like ~1 year of returns. Not a huge deal if you need to dip into it.

(There’s plenty to criticize about the whole 401k system of retirement accounts. But these criticisms seem misguided)

I mean put it in context of the OP.

People putting retirement funds in a pile of companies that often have little impact on local communities they live in.

They’re changing laws to fast-track sketchy IPOs, putting hard earned money at risk why? So we can send people on a death-mission to Mars?

Point being, they are doing what they will with other people’s money and won’t suffer the consequences. Removing the checks and balances is exactly how financial disasters happen.

You can move your 401k money between several funds at any time.

Exactly right, there's even ones so conservative they market themselves as cash equivalent. Basically zero gain/loss in those funds. If you're so worried then go login to your 401k and change it.

“Basically zero gain/loss in those funds”

Right so what’s the point then? Doesn’t that prove the point: zero gain/loss with a 10% withdrawal penalty to boot.

And your proposed alternative? Please provide at least 50 years of historic returns

If I could pick from any possible retirement plan, I'd want in on the UK pension system that's guaranteed to beat inflation and earnings growth. Until the money runs out, at least!

Because it's not an actual investment and can't run out. Like US Social Security and many other national schemes, the UK is pay-as-you-go. Money coming in is immediately paid out.

Any funds lying around are supposed to be for temporary imbalances, but became significant due to a major demographic imbalance: the Baby Boom.

But they're not significant. The National Insurance Fund is supposed to keep a minimum overfund of £24B (at current spending).

It's now at £79B. It's significantly overfunded.

You can buy i-bonds in the US. You are limited in how much though. They are pegged above CPI. You never hear about them because no one makes money on it. And maybe it isn't that great of an investment.

Historic returns do not predict future returns.

Na, all it will take is a shift in perspective. As humanity becomes more and more Asian, those names will trigger a "who?"

It'll then just be Confucius, Cao Cao, Oda Nobunaga. Doesn't really refute the point.

Pretty odd to say Mohammed a) isn't Asian, and b) isn't widely known across the continent, because he is, not just in western and central Asia but even in places like China, Jakarta and Mindinao. Islam is expanding rapidly in Japan now.

Mohammed was Asian. Julius Caesar had huge bits of his Empire in Asia. Most of Alexander the Great's empire was in Asia.

There needs to be a realization of how important communication skills are to develop and possess. The act of disagreement has skill levels that do not trigger emotional responses, and cause cross understanding to occur. Learning how to convey understanding and gain understanding from others becomes more and more important in a landscape of rapid change. Which we are collectively terrible at, with most companies being miscommunication circuses, with all the stress that generates, needlessly.

Is there an AI Coding Agent application structure emerging that is more or less universal across llm models? Is anyone collecting and writing on how to understand this architectural style?

> Is anyone collecting and writing on how to understand this architectural style?

Are we on the same site? Is anyone writing about anything else?

The blog post this discussion is for is one of the first in depth discussions I've seen of how these coding agents work. Most posts cover how to use them, not their internals and how they operate.

The pattern across Claude Code, Codex and Cursor does seem to be converging: gather context, make a plan, execute, then verify.

What feels less standardized is how much control the user gets between those stages. Settings like showClearContextOnPlanAccept and disableAutoMode are interesting because they expose that boundary between “agent decides” and “human reviews before execution.”

That seems like the part where different coding agents will continue to feel very different in practice.

As costs come down, these dials will shift too. There are many configs which give control simply for the sake of saving tokens.

Around the same time I first read On The Road, my wiser than should be possible mother said "Oh, you need to read this too" and it was "Off the Road: Twenty Years with Cassady, Kerouac and Ginsberg", written by Carolyn Cassady. Rips the band aid right off of those sexist abuser of innocents, those utter assholes. They write great literature, and the fact that they expose their own terrible ethics bare, but surrounded by non-condemning language is the trick. They never hid their nature, but America never realized what they were praising, not really. Which is all so American!

Tried both Firefox latest and google chrome's latest desktop browser and does not run in either.

One of the first homework assignments when I learned C back in '83 was after a long lecture on how the string functions are fundamentally broken, and the class introduction to writing C was fixing all of them.

My memory growing up is that making your own C library was basically an inevitable rite of passage for any aspiring programmer.

And then your own custom allocator that would be fitted for your algorithms and vastly faster than malloc.

Yeah, it's a shame we never got something like boost for C. Every company I ever worked for had its own common C library solving these problems.

It's a shame we never got a package manager for C (or C++).

EDIT: perhaps I should have been clearer; by not having one early on, we now have multiple competing package managers, with no clear winner. Responses prove that point.

Although vcpkg is probably the most popular, I’m a fan of https://conan.io/center

Never used this? https://vcpkg.io/en/

I worked at a shop where we used Boost in a C++ code base that the only use of C++ was the harness to use Boost. After that, it was all C, object-styled C, as that code base started before C++ compilers were not a template overlay on C.

Any reports from people using their coding agent(s)?

I'm running Qwen 3.6 27B Q5 K M GGUF on a Tesla P40 and koboldcpp using pi.dev as the harness, I gotta say I am impressed. Took some setup and configuring but I already have some code it has made commited and pushed. It can be slow on my hardware at >50k tokens, but the fact I bought this one P40 for like $150 back when the LLM trend started I can't complain. (I have a second one too but I couldn't physically fit the card in my server unfortunately.)

The setup I had to do was important and I had to compile koboldcpp with a few special params for my hardware, I mostly just had Claude figure it out. I don't remember everything I did now but it was very slow and would often stop mid task, it seems it was mostly a parsing issue. It made the model seem broken/dumb, but once I had all that settled I actually am able to use this how I use Claude Code. Disclaimer, I am pretty explicit with requirements, I imagine this fails more when you leave it to figure out things on its own but for my flow its pretty rad.

Currently setting it up as an automated agent now to pull Trello cards, create PRs for them, and move the card to be reviewed.

Command I am using to run: python koboldcpp.py \ --port 61514 --quiet --multiuser --gpulayers 999 --contextsize 262144 --quantkv 2 \ --usecublas normal --threads 4 --jinja --jinja_tools --jinja_kwargs '{"enable_thinking":true, "preserve_thinking":false}' \ --skiplauncher --model /data/models/Qwen3.6-27B-Q5_K_M.gguf --smartcache 5

Qwen recommends to preserve_thinking: true for agentic/coding workloads.

Thanks!! I had disabled that previously while debugging, I can confirm this is helping accuracy from what I can tell so far. (And speed since the cache is preserved more often!)

Use the MTP models which 2x token generation speed, for example: https://unsloth.ai/docs/models/qwen3.6#mtp-guide

Very interesting I'll have to check this out thank you. This is why I love HN.

I'm using the pi-mono coding agent (open source, free) without any extensions and very simple prompts. The 3.6 27B model (BF16, 250k context) uses 67GB VRAM on an RTX PRO 9000.

It's very capable on almost any coding task I've thrown at it, and very good for easy-to-medium hard scripts, new code bases.

It struggles on some complex tasks in larger code bases, e.g. using to debug and fix bugs in llama.cpp it gets close to working code but often introduces errors. For such tasks its still very useful as a search/explore tool and drafting fixes.