That’ll work until they create for themselves new competitors that are good enough at previous (normal) market prices. And you don’t need incredible high foundry tech for that, so it’s closing up on them eventually.

No it’s not, and part of that is the current legislative laissez-faire in the US that put its regulatory bodies on a hiatus. Under normal circumstances, this stuff should have been under much more scrutiny and regulations. I’m not saying that the state should control what LLMs do or who’s access to them, but they could very much tackle the deceptive marketing, environmental and societal impact, unsound financing, abnormal market consolidation, and mitigate the overall financial risk.

I still have mine, it’s a decent audiobook player for the kids, it even has a jack, a replaceable battery, induction charging and is so nimble. And did I mention the full keyboard? I hate that we followed Apple design ideals, for the sake of having today overpowered dumb, heavy , fragile and inconvenient consumption-only bricks.

Is that another overpowered smartphone SoC running a wasteful android with a thin UI and dumbed-down music player? If so, let me just repurpose an old smartphone and save the planet from some e-waste. If not, kudos.

Maybe I’m the weird one here, but when I look at fiio, and they are essentially smartphones (beefy Exynos or Qualcomm SoCs, running a bloated Android), minus some connectivity apps and a mobile antenna, I’d rather just use a phone, thanks. Most portable music players have this minimalistic appeal of devices that were engineered for just one single purpose and to do it well, and fiio certainly isn’t having that vibe.

Yes, the point is LLMs are AI.

So, we’ve gone full circle. LLMs is a sub-category within the collection of ML techniques within the collection of AI techniques. Point is, AI is just a term here. A label, that indicates nothing about the abilities of LLMs to exert any form of intelligence or reasoning. In other words, the whole field of AI could (should?) have been named “computational statistics” or “mathematics applied to numerical datasets” (or whatever else you want, really…), and LLMs would absolutely belong to “CoStats”/“MAND” fields, for the same reason we say they relate to “AI” today, it’s just that nobody would be silly enough to call them “artificially intelligent”.

ML is AI too. But sales didn’t call it that because AI had the reputation of “just brute force with some heuristics”.

What sales are you even thinking about? What do even presume the market for ML algorithms to be? Nobody was shopping for support vector machines as a service, or spending tokens on linear regression, or using convolutional neural networks via API before the current LLM craze. What the field is experiencing right now is unheard of. Never before did the private sector jump gun on a niche technique and spent trillions to package, anthropomorphise and market it as if “AI has finally been figured out, and we happen not only to own it, but also to sell it to you”. This deceptive rhetoric would be a much tougher sell if some early computer scientists hadn’t happened to name their field “AI”.

Was there even a point you wanted to make? I’m not sure exactly where you are heading with all this.

I mean, “machine learning” was a marketing term invented exactly to avoid the A in AI.

No, it’s not. ML is a sub-category within the collection of AI techniques that describes those algorithms whose behaviour is the result of fitting a training data-set to a pre-defined model by minimising an agreed-upon error function. For the longest of times, we were just calling that “statistics”, and many ML techniques and algorithms predate computers by centuries. Your mean-squares curve fitting? …qualifies as ML. That is to say, ML is all AI, but not all AI is ML.

LLMs are no different than function fitting with mean-squares. They are not magical, they are not black-boxes: they are fully described and completely predictable.

I mean, have fun re-defining artificial intelligence to suit your narrative. Here’s what the scientific consensus stands at: https://en.wikipedia.org/wiki/Artificial_intelligence

It’s also no secret that the companies involved in the LLM craze (OpenAI, Anthropic, Microsoft, …) diligently muddy the waters as a marketing trick to sell their product for more than what it is, that’s precisely what I’m calling out here: we can collectively do better/know better than that.

Also, LLMs are not AI (in the sense that most people would deem meaningful): there’s no reasoning involved, just a convincing illusion of it, served by extensive knowledge compaction and next token prediction. That is not to say that LLMs are useless, or that the sort of all-powerful AI people fantasize about cannot exist (I really don’t know, no people do). That is to say that it’s not for now, and that before it happens, we will see another long AI winter before the emergence of something fundamentally more convincing than LLMs.

Hard disagree, the more you look at it, the more this can be described as “a product to convince people that discuss technology online that this is a product for tons of people”. Most people need HDMI more than they need record-matching single-thread performance. Most people need more RAM because almost all they do is web-based and not MacOS native. Most people need I/O because they have mice, weird peripherals, and tons of usb drives, mostly USB-A.

That said, it’s a magnificent second or third device for the tech elite that’s already committed to the Apple ecosystem (having nothing on-device, whose peripherals and I/O needs are covered by more general-purpose computers) and who will absolutely brag that the performance of this lets them do lots of coding (using it as a metric for what a good computer is, in their eyes). Strangely both things can be true at the same time!

to be fair, that wouldn’t be my pick at that pricepoint. That single-core performance would still be unmatched, but I would have something with better IO, more RAM, and more life in it.

And thanks to the 8GB soldered RAM you’ve got, you will enjoy this ludicrous speed for about 5 minutes of mild usage, thanks to liquid glass, electron apps and modern web. Brilliant.

People don’t always need fastest possible computers for their tasks.

Not arguing with that, I am myself daily-driving a lenovo that was assembled 3 years before the last intel macbook was produced, so it’s not exactly new either.

Besides, Macs tend to be nice looking and their build quality is good.

This helps explain why anyone would want to buy an obsolete, irreparable, non-extensible laptop 6 years after its discontinuation. I don’t think it’s good advise, though :-)

What’s the point then? Those ancient intel chips are just slow and overheating and getting their ass kicked by anything by AMD from the last 5 years.

Element says lots of things.

You seem to be obsessed with optimising one resource at the expense of others.

If you want to push it and paint me as obsessed about something, then let it be this: providing this community with on-topic and reasonable advice

you’re only going to save a few MB of RAM.

This is false, and you should read once again my previous message illustrating why: on a decent “self-host”-friendly machine, the same software may work very well, or not at all, depending on whether the user would engage with very basic configuration. This goes beyond RAM (memory isn’t the sole shared resource), and I’m adamant that the alternative (which was “pretending that the problem doesn’t exist” turned into “throwing money at the problem”) is unreasonable.

On the other hand, if you’re doing it to learn more about computers then it might be worthwhile. This is a community of hobbiests, after all…

Or more importantly: the extent to which you can self-host out of sheer luck and ignorance like you suggest is very limited. If you don’t want to engage with a minimum amount of configuration, you might bump into security issues (a much broader and complex subject) long before any of the above has a material impact.

I’m saying this based on real world experience

And do you think I would spend my time engaging if that wasn’t from my own very “real world experience” of lessons learned the hard way?

Bringing-up “diminishing returns” as if this was an optimisation game also doesn’t do this justice. Take the typical “household FOSS package” with software names often brought up in here: a nextcloud instance, a photo-sharing service like immich, private instant messaging, a software forge, a subsonic-compatible audio/video streaming server, a couple php websites like wallabag and RSS aggregators.

An Intel Atom CPU and 4GB of RAM is plenty sufficient for all that, and will cost you single digit USD a month, granted you put the (one-time) effort to tune and balance those services. Would you run all the above from upstream’s docker files, I can guarantee you that you would deem this (perfectly fine otherwise) server underpowered for the task at hand (and would probably go for a 10th gen or so Intel Core CPU, quadruple the RAM and 3-6× the energy cost in the process).

And that’s the point I’m making here: a self-hosting community of tinkerers should (ideally) know better, for the ethics’ sake of keeping the process environmentally friendly, and not wasting other people’s money.

Do you have the data to back that up?

I mean, you are the one making the exceptional claim that unnecessarily running multiple instances of programs on a device with finite resources has no practical adverse effect. Of course, the effects can be more or less drastic depending on the many variables at play (hardware, software, memory pressure, thread starvation, cache misses, …) and can indeed be negligible in some lucky circumstances. The point is that you don’t call that shot, and especially not by burying your head in the sand and pretending it’s never gonna be a problem.

Effective use of computing resources requires tuning. Introduction of a new service creates imbalance. Ensuring that the server performs nominally and predictably for all intended services is a balancing act and a sysadmin’s job. Services whose deployment settings are set by someone with no prior knowledge of the deployment constraints can’t be trusted to do a good job at it (that’s the nature of the physical world we live in, not my opinion), and promoting this attitude promote the kind of wasteful and irresponsible computing I was on about.

Now, I’ll give you the link to this basic helper for tuning a PostgreSQL server: https://pgtune.leopard.in.ua/
Will you tell me what are the correct inputs for my homelab (I won’t tell you the hardware, the set-up, the other services running on it, the state of the system, etc)?
And later, when you will distribute your successful container to millions of users, what will you respond to the angry ones that will complain that your software is slow, to no fault of your coding, because they happen to pile up multiple DBs, web servers, application servers, reverse proxies, … on their banana SoCs?

I disagree. You are just entertaining the idea that servers must always and forever be oversized, that’s the definition of wasteful (and environmentally irresponsible). Unless you are firing-up and throwing-away services constantly, nothing justifies this and sparing the relatively low effort it is to deploy your infrastructure knowingly.

Precisely what pre-devops sysadmins were saying when containers were becoming trendy. You are just pushing the complexity elsewhere, and creating novel classes of problems for yourself (keeping your BoM in control and minimal is one of many others that got thrown away)