My thought playing with Llama 1 in 2023, after the initial shock wore off, was “whoa. You know, most people do not have the background to understand how this thing works.” It messed with my head, for sure. It’s like it was designed to defeat a turing test, especially when the next versions got more sycophantic and “confident.”

There was even a story of a Google scientists getting AI psychosis from a BERT model back then:

https://www.scientificamerican.com/...

I guess a naive part of me thought… it’s fine. They’ll be warnings. They’ll be tons of finetunes that have no hesitance slapping sense into the user, like Resetti in Animal Crossing, and it won’t be public facing anyway. The community using these things will teach each other how they work, right?

Laughs nervously.

I wasn’t cognizant of how maniacally sociopathic Altman was back then.

The crazy thing is they aren’t really designed for that, especially the reasoning part.

I pictured LLMs proliferating in the way SGLang was taking it: fill in the blank. You’d give raw completion models (not chat finetunes) chunks of structured content (like JSON schemas, names in fields, more complex schemes filled programatically) and it would fill in fields, with intelligent context caching. Like, if you needed to finish a programming function, maybe you give it the file as context and guide it along.

Temperature based sampling felt like a quick hack, just until either looping or structured sampling was figured out.

…But they didn’t fix any of that.

Papers came out on the issues. But the big AI houses didn’t implement any of it. They basically kept it exactly the same and just kept scaling it to burn more compute.

What I am getting at is that OpenAI/Anthropic development is way, way more conservative than anyone thinks, and they basically took temporary hacks and made them bigger. That’s why you can’t trust the code they churn out, because random mistakes and so many other issues are literally part of their core.

Then they turn around and sell them as confident engagement engines…

It’s utterly mind boggling to me.

I’d get into optical computing.

Look up Intel’s research into optical buses, as an example. Or experiments in pure optical inference. I just find this whole field extremely practical in a wide range, from “usefully augmenting silicon PC chips” to specialized accelerators to farther-out computing paradigms.

I’d also look into Josephine Junction computation. If there’s a sudden breakthrough in room temperature superconductivity, that becomes very important, very quickly.

I also find “biological computers” or those emulating biology interesting, but don’t understand enough about them to even begin to assess it.

I’m less enthusiastic about quantum computing. It’s extremely useful in a few niches, but it feels like it’s hit crypto-like levels of hype: a narrowly useful tech that’s being pushed way beyond what it’s good at.

It is related to the other areas I listed, though.

Pretty much all this stuff will be corporate-first though.

That’s just how it works. They pay the big R&D for prototypes, and the mass commercialization comes later.

If you’re more directly interested in consumer tech, you’ll have to look into memresistors, alternative lithography like gallium or nanotubes, chiplets and “split” architectures and interconnects, things like that. ECE magazines talk about this stuff constantly because it’s closer mass commercialization.

Well, what if Sony made the PS6 a Linux PC?

They can have their own embedded storefront and proprietary stuff, but that would still be cool.

More GPUs =/= better AI.

More data =/= better AI.

More tech bro “superstars” =/= better AI

This is what people like Musk and Zuckerberg don’t seem to understand.

Training scales very poorly past a certain cluster size, especially if you go for new architectures to actually pursue improvements, hence reports of GPUs being tasked with busywork just to meet utilization quotes. Increasing data size and training scale hits diminishing returns, quick, or even regresses models because the bulk data is shit and the model is too inefficient. A prime example: Llama 4. “Superstar” AI engineers are better and Tweeting and sycophantic gaslighting than coding something interesting.

In other words, I’d argue there’s a much smaller “sweet spot” for pure LLMs that these billionaires are way, way past. And no one is telling them no because they’re too rich to hear it. It’s all going to collapse on itself because scaling like that just does not work.

Holy moly, this comment section.

Good job for defeating my very low expectations, Lemmy.

No, they don't.

They're surrounded by yes men. And from everything I hear them say, they don't understand the first thing about how LLMs actually work.

On Lemmy/Piefed?

Usually because they’re Tweets reposted from somewhere else (hence OP would have to transcribe to post as text).

The question I have is why post Tweets at all.

They don't even need to dampen, just redirect. Pay engineers to make them specialized tools, don't waste so much money on GPUs, and give part of the compute to the community to tinker with, instead of hoarding it and doing squat.

That is so NCD it hurts.

You should repost it here, lol.

Others said it better, but:

It depends on what, and where, you're fighting.

All part of the plan, so you subscribe to game streaming.

7900s, 4090s, and 5090s will become “forbidden technology” like you see in post apocalyptic fantasy where tech is magic. But also “idoocracy cyberpunk,” as human production is diverted to launching GPUs in space which engineers… awkwardly task with busywork.

You think I’m being hyperbolic. I am not.

Is there a notable preorder discount?

If no, then don't.

The only case would be if you know you're going to buy it anyway, even if its bad (just to support the franchise or something like that), and if there's a preorder discount. Otherwise there's no logical reason to preorder.

This is so Non Credible Defense it hurts.

Many games are sold on 1st party storefronts. Sometimes even indie ones, like Rimworld. I’d say they are less likely to yoink the game since they own all the rights, and they’re often downloaded as DRM-free executables.

So I’d check, in order:

• GoG

• The game dev/publisher storefront.

• Some storefront with a DRM-free download. I noticed EGS (for example) does this for many of their giveaways.

• Steam, as a last resort.

I wouldn’t trust PS5 discs though.

…But you should consider practicality, too.

In practice, your PS5 is going to be your most powerful machine until you get a bigger PC with a GPU. And FYI, the Steam Machine will not be $1000, not even close.

That, and launching less intense games on your Deck is way more convenient if bought through Steam. Otherwise you have to deal with some intricacies of Proton and controller remapping to get them running.

So the PS5 discs aren’t impractical. Steam isn’t impractical. They should launch for a reasonably long time.

The war in Ukraine is NOT a modern war shaping future combat. It’s primarily characterised by obsolete static and trench warfare with both sides severely lacking capable air defense and experience in proper combined arms warfare. Then Russia aggrevates that situation even further as “We can’t compete on the high-tech end of things so we drag everyone down to our level via a massive electronic warfare campaign” is part of their main strategy since the cold war.

Politically, that situation is so insane though.

EU stockpiles and stockpiles high tech weapons for this exact situation. Including air defenses that could tip the scales, I presume. And when they get an honest to god Russian invasion at their doorstep, Europe... Lets them sit and wear away in storage?

I wouldn’t say all “AI” was a grift. Machine learning is a useful tool, like a hammer, it’s just not a magic genie for everything. Always has been, always will be.

Same with blockchain, albeit in a much narrower niche. I do think it’s a terrible system for a widely-used currency, though.

Same with quantum computing. It’s a niche.

The pattern is that Tech Bros inflate something narrowly interesting into a “it’s going to ascend the human race if you give us enough money” FOMO thing.

…And, currently, the next target seems to be space travel.

Again, I emphasize. Very useful in certain niches, like science. Stupendously impractical outside of them.

It already has.

This is misleading.

UBlock has been deprecated on Chrome, for a long time. UBlock Lite is its replacement and will continue to function (albeit with more limited efficacy).

What they are talking about is a complicated series of command line flags to re-enable Manifest V2+installing UBlock from source… But who in their right mind would still be using vanilla Google Chrome and jumping through all those hoops?

It will be an issue for forks like Helium or Ungoogled Chromium. They’ll just have to patch in a native blocker, I suppose.

TL;DR: Headline is wrong.

Chrome users will notice nothing. The end happened a long time ago.

It’s not a guess. Look at the price of any PC now. Or wholesale prices of individual components like CPU trays or RAM ICs.

$1000 is basically impossible for the specs it has. Even if they sell at break-even cost, it’s not even close.

No one knows the actual price tag will be, of course, but there are practical minimum bounds.

Hence, if that’s your limit… you should definitely plan on the Steam Machine being too expensive. Maybe consider a used AMD 6000 series GPU in one of your servers.

It’s fun mixed with ADD.

• Hyper research

• Pause, put in to-do pile

• …Repeat

Bonus if it’s something I don’t actually need the product immediately. Like, if you want to talk about every single mirrorless camera on the market and what will come out within a year, with reasonably probability, well. It’s on the top of my head now. And it’s definitely not researched at cost of other stuff I need to do, nope.