abalashov

Yes, but we have far fewer reasons to think otherwise, since we don't know anything--if you're talking about Americans.

Hard to say. Advertisements have at least a pretense of persuasive intent. Propaganda mostly operates in the absence of contrary information.

There is way too many mixed marriages etc. for any reasonable Russian to believe that their formerly-closest East Slavic neighbour has turned into a Fourth Reich.

Exactly. The distortions in American public life are quite a bit more intricate and artful.

Recognising that there is propaganda != being deeply savvy about what the propaganda is. However, I've found it to be an article of faith in America that the US just doesn't engage in that kind of thing--less so in recent years, as the crisis of belief and confidence in institutions has ricocheted, egged on by the disinformation and demagoguery of the Trump years.

As a Russian, I deeply appreciate this. In Russia, it seemed obvious to even the most parochial peasant that there was propaganda, while in America, the vast majority of the society not only fails to consider the possibility, but is cholerically allergic to the very idea.

Edit: recognising that there is propaganda != knowing what is and isn't propaganda. That's all I meant.

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.

Try the OpenRouter route!

https://openrouter.ai/moonshotai/kimi-k3

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.

Thank you! This looks like it might be the ticket.

Hey, there has to be a legitimate use-case somewhere. Thank you!

Does it?

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.

Okay, that's genuinely funny. Sadly, it might be a little subtle for the folks who need to feel the mockery most.

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.

It's almost like it's inherent in the definition of LLMs.

It's really, _really_ high time we dispensed with the idea that this is "AI". Nobody said they're not useful, but "AI" they are not.

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:

https://medium.com/@JerryCuomo/why-ai-gets-the-strawberry-qu...

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.

Although I agree with you, in fairness, there are some lively controversies in the world of cognitive science and philosophy of mind about whether this meaningfully differs from human thinking at sufficient scale.

The general idea is that the building blocks of "coherent models" and "processes" within the churning of the human intellect are also, in so many words, prior art and existing concepts, and so, while the human mind is not a text model, a sufficiently large and sensorily multimodal neural net would not be too different. Neural nets are, after all, inspired by what we understand of human cognition -- they'd say.

This is an astute observation. I think it reflects a larger and longer-running strain in the relationship between technocracy and the humanities, though, of which this latest iteration is just even more choleric and rote. The plumbers of capitalism always seem to have had deep contempt for the arts and the humanities, not in the least because they didn't do too well at them in school or didn't understand how philosophy relates to making money, or something.

This has led to some rather fantastical conclusions on both sides, however. On one side, there's an almost sadistic "revenge of the nerds" glee at the notion that these airy-fairy, frou-frou, and "feminised" liberal arts majors will finally crumble before the stochastic parrot machine god, and on the other side, a no less comical notion that after "AI" ushers in utopia, then high-brow artistic and literary pursuits will be all that remains for us to do.

Fair enough, but I'm not sure this is how the general population uses LLM chatbots, nor how the highly qualified always use LLMs.

On the contrary, I believe most people use them precisely to find out more about topics of which they know vanishingly little, much as they'd have used Google before LLMs.

This is my position, as well, and I have switched exclusively to Chinese models in support of this view (and cost and because they're pretty awesome).

It's quite trivial to show that an LLM doesn't have underlying intent, and that it can only emit direct textual convolutions of its training and not combine tokens in truly novel ways. This is the very thesis of the world-model folks, e.g. LeCun et al, that LLMs are a general intelligence dead-end because they lack any inner concept of the world around them, and do not reason from that.

Furthermore, LLMs clearly do not "reason", despite the marketing around this term; their "chains of thought" are the nothing more than the result of having been trained on explicitly verbalised multi-step processes. There are many cases where the putative result arrived at in the <think>chain of thought</think> does not match the result emitted.

Whether they are "better" at software development than I am greatly depends on whether one is asking them to retrace worn technology paths that are well-represented in their training--in effect, to copy prior art--or to do something in quite obscure technology, or something quite novel altogether. (However, I will happily concede that most everyday business programming involves neither.)

Still, if LLMs were actually intelligent, let alone superhuman in the sense you suggest, then we would expect major scientific breakthroughs to be raining from the sky. If, say, an Einstein, could transform physics with only the knowledge gleaned from a human's feeble capacity to retain the literature of the time, I'd expect LLMs, who retain orders of magnitude more information with far greater fidelity and precision, to have offered at least a small slither of evidence of their superhuman capabilities.

I would also expect the objective progress and capabilities of this galaxy brain to be accelerating, not substantially slowing down as it has. GPT-2 to GPT-3 was truly a quantum leap, GPT-3 to GPT-4 was a substantial jump, GPT-4 to GPT-5 was meh, 5+ is basically unimportant, and so it goes for the other models. There are, of course, holes plugged and benchmarks where these evolutions have been, in various niche ways, consequential, but in the plainspoken meaning of model capability, the low-hanging fruit of pretraining was clearly exhausted quite some time ago. The carnival has been running on "agentic" / MCP / RAG / tool-use fumes since. This is moderately impressive and adds quite a bit of runway, but intelligence it is not.

I do think it's easy, in this technology discussion bubble in which we dwell, to overestimate the centrality of LLMs to the arc of developments in our time.

They'll be important, but I don't think they'll be _that_ important, because the rest of society and the economy don't move at the speed of SV. Instead, they'll be overtaken by other, more traditional categories of events, ruptures and dislocations.

Moreover, folks will eventually realise that while they are very impressive derivative databases of knowledge, they're not at all "AI" -- well, not the "I" part, anyway -- as the concept is traditionally understood. There's not any "I". It emits convolutions of its training, and it does so very impressively, and that can even be harnessed by agents to connect them to levers, servos and richer information sources. It's nifty. But it's just not intelligence. It's more of a kind of queryable database than a robot.

Once that realisation diffuses more widely, I think it'll turn out to be a more prosaic and underwhelming development than is presently hypothesised, either here or by the press. It doesn't mean many business and managerial class folks won't try to squeeze everything they can out of so-called AI, but the idea that this can effectuate truly widespread labour displacement will probably quiet down considerably. (The valuations that depend on this assumption may collapse more abruptly and less gracefully.)

The challenge is staying solvent until then. :-)