I don't know that this starts with LLMs, though LLMs are certainly--at least for some people--a way to take their ADHD tendencies further.
It's been obvious to me even since I was a computer-addled kid, in the 90s, that the way computer interfaces and multi-modal notifications affect the nervous system, the capacity for sustained attention and patient reading, and so on is pernicious. The impact has both cognitive and moral valences that reinforce each other, because the instincts reinforced by computer use turn into technocratic, anti-humanistic worldviews, and even cosmologies.
I imagine it's easy to read something Ludditic into this critique, but I mean it quite a bit more matter-of-factly, and from the vantage point of a tech professional who is as passionate about coding and computers as anyone else on HN. Still, it's sobering to consider the ways in which being so computer-bound since age 9 or 10 has altered the development of my habits of mind, my nervous system and my cognition, and speculate about might have been gained and lost.
I have an MBP M4 Max 128 GB and have encountered these very limitations, though many are averted by using Qwen3.6-35B-A3B. I'm also not sure there's a magical harness out there that obviates these limitations.
Oh, okay. Didn't know that, as I haven't tried K3 myself -- I'm quite content with Kimi K2.7-Code for the moment, which OpenRouter serves me to via US-based inference providers. Supposedly.
Obviously, not relevant if your reason for being here is to ferret out the excitement about K3 specifically. But if this is your first time trying Kimi, try K2.6 or K2.7! If my experience is any indication, you'll be blown away by those, too.
I just pay metered usage via OpenRouter. I've spent no more than $10-$20/mo, but my usage is, by the standards of so-called "agentic coding" these days, very light and confined.
I doubt my workflows match the profile of "daily driver" as you're conceiving of it. I'm mostly in the IDE chat box asking for a hand here and there, but no "long horizon" / big vision / architecture things. Occasionally, I'll solicit a general opinion, quite abstractly, which isn't very token-intensive. My contexts are tiny and I aim to keep them that way.
The only token-intensive thing I do is occasional analyses of open-source codebases along the lines of "what is the actual behaviour when message X comes in and is malformed like this?", or something else similarly far-reaching. For that, I use DeepSeek's Reasonix, which they savvily dropped at the exact moment the sting of Claude API pricing became a sensation. It's a very cache-oriented harness, and delivers on its promise of being very economical. I doubt I've spent more than $10 with DeepSeek over the months I've used it. When I want a harness to inhale a codebase and give me an answer, Reasonix has been the go-to for a while now, and it's great. DeepSeek-V4-Flash is enough for many simple things, though I just run it on DeepSeek-V4-Pro because it's so cheap that I don't care.
Yeah, but the trillion dollars are not concentrated in the Linux kernel project, but instead distributed all over the planet. That's not how the AI lab economy works.
I've been using Kimi for coding tasks for close to six months now, and haven't looked back. I'll periodically try something on Claude to make sure I'm not missing anything, but I've been very happy.
I just do the OpenRouter thing. My use of LLMs is narrow enough that cost is a negligible consideration either way.
You would think it would be that straightforward. However, accurate synchronisation on GPS or temporal attributes would be required.
Judging by the paucity of software to do this, historically, it is not a straightforward problem, or all the devices involved don't generate all the data points required.
The real mess is when you have 26 clips from a long event to string together. It can easily take a day and a half to make a 3 minute montage out of that.
As a cyclist (and former racer), I still want to know how to capture videos with telemetry overlays (speed, power, HR, etc) from my head unit in a straightforward way. NorCal Cycling's videos - https://www.youtube.com/@NorCalCycling - are an excellent demonstration of this at work.
Yes, I've done the Garmin VIRB Edit thing, which is the very approach recommended by Jeff (NorCalCycling) in his tutorial videos on the subject. It feels like something out of 2005. It is incredibly labour-intensive and imprecise unless you're fortunate enough to be in relatively short criteriums where you've got the battery runtime to just record the whole race. Most real-world events and rides require one to turn the camera on and off at certain moments, which then requires _hours_ of stitching together clips and correlating them to GPS fixes from the head unit (in the FIT file), and quite imprecisely at that.
There has to be a more 2026 solution to this. All you need to do is correlate the footage to the FIT data points by timestamp, in the temporal domain.
If Garmin came out with one, it would absolutely annihilate this space. To the best of my knowledge, there is no competition that offers anything turn-key, though perhaps the best of my knowledge has not aged well and by now there is something. It's maddening.
My use of "worldly knowledge" might have been sloppy. I really meant "real-time worldly knowledge".
The American chatbot apps are just very polished on live web search and tool use, and the models are very eager to do it. Chinese models are perfectly capable of that, but you need to bring your own MCPs (at least, if you want anything beyond WebFetch) and steer the model toward greater eagerness to use them.
I switched to exclusively Chinese models, mostly Kimi, many months ago. I'll still ask Claude questions that require ambitious real-time web search / worldly knowledge, but for just about anything else, the Chinese models have been so good that I haven't looked back.
> the concepts just don't stick because I don't actually do the work
Yes! And this is the part of the de-skilling puzzle that is completely unaddressed by AI boosters.
Maybe LLMs are a force multiplier, but there still has to be some force to multiply, and I don't think a lot of folks ask the question of how that force is actually cultivated. This nebulous, airy-fairy notion that humans add "architecture" or "taste" doesn't tell the story of how, concretely, they came to have it. It seems to me there is no escaping that it came from typing the code.
Like you, I have a much better grasp of code and API surfaces I've physically typed than things LLMs have emitted and I have reviewed, in a conceptual sense, but which I could not have typed myself then, nor can type now.
Underrated and paradoxical element of this, to which, as far as I can tell, LLM boosters offer no solution:
> And without my knowledge I would have no clue if what it's trying do is correct, or safe.
I would contend you got the knowledge by typing the code yourself, that there's no other way to get it, and that if you stop typing the code yourself--and the slogging that entails--you'll lose the ability to prompt LLMs effectively.
It's not that I think the physiokinetic aspects of typing as an input mechanism hold some metaphysical distinction, but rather the level of engagement it forces with the code, and the units in which it does so. I'm not aware as yet of any viable replacement for that.
It's easy to trade on decades of software engineering experience with LLMs: with sufficient experience, everything goes around and comes around, almost any pattern is recognisable, the gratification is immediate, the benefits are now, while the costs and disasters are down the road.
However, the technology world is not static, and if you don't keep up with new frameworks, libraries, languages and other tech in that physical-mechanical "mind-body-keyboard" way that typing--or something substantially close to it--accomplishes, you will lose the ability to navigate that world fluently. To say it's just another abstraction layer and that the world didn't crumble after compilers is to miss something quite essential about how LLMs differ from compilers or high-level languages. The disengagement with the process of physically programming something quite specific will take down with it the ability, over time, to formulate useful prompts and competently review the output.
This is exactly the right answer. The supposed "rationality" of capitalism can ruin us before we get a chance to dazzle the world with our contrarian insights.
One hardly needs to soar to the exalted heights of "mysticism" or any sort of "ineffable qualities" to come into contact with the limitations of LLMs.
1) Models do not perform the functional equivalent of reasoning at all. When we reason, we don't simply babble out textual derivations of prior examples of "reasoning" to which we have been exposed, arrive at a conclusion, then occasionally state an altogether different conclusion while pointing at the largely irrelevant reasoning to substantiate it.
2) Models have real-world, not-at-all-mystical functional constraints that are directly relevant to the production of everyday economic work in which one attempts to involve them. Their inability to extrapolate or maintain clear mental models leads to staggering, head-scratching mistakes that even a very feeble and developmentally awry human intelligence would not make.
A basic, if well-worn example that was widely discussed in the last year or two:
However, this is emblematic of a much larger idea: the LLM doesn't have any idea what a letter is or what you're asking it to do. This isn't a question of "ineffable qualia"; when it doesn't know what something is at any essential level, it can't competently solve problems related to it. One bumps up against this in everyday programming and all the time.
Also, what is "mystical" about my demand for the kind of scientific progress--no, forget that, any scientific progress--that a functional superhuman intelligence would yield? I am not a "mystic", either; I want functional results, show me the functionality.
http://www.evaristesys.com/ http://www.csrpswitch.com/