If model distillation is illegal, then shouldn’t we also make training of copyrighted material illegal?
Right now, something is legal or illegal if you want it to be. All you have to do is bribe enough pockets in the justice system.
China is spending a fraction of what US companies are burning through. Profitability will be a lot lower bar. Plus they’ve been investing in renewable energy to power it.
I’ve also become more and more convinced that these colossal huge general models are not necessary for many useful tasks and smaller more focused models will do just fine for a lot of the workloads people will actually be willing to pay for. Particularly once these companies have to stop selling at a loss.
What’s convincing is that hobbyists are taking open source AI models and their own batch of training data in order to do small jobs. They need more LORAs and more massaging of the prompt, and they might take minutes or hours instead of seconds to generate, but they absolutely work.
It’s especially useful if you want to make slop of a category forbidden on the provider servers.
That would imply that AI does something profitable, which it doesn’t appear to
I would imagine there are countless cases where a mundane task requires human input in assembly, warehousing, … effectively Chinas whole export oriented economy. “Dark factories,” and the like, should only become easier. That’s not how the US orients itself, though.
From what I understand, China is also doing phenomenal on batteries and renewable energy sources. That’s going to make the cost of conducting research on Chinese infrastructure cheaper. Cheaper means more accessible, which accelerates innovation—right? I imagine this effect will become more evident over time.
They have to find a way to get the taxpayer foot the bill, or the grift is going to end
Or, people’s retirement funds via IPO. If you can change your plans investment strategy for your good and societies.
It’s lemmy.world, where China is always winning righteously.
I feel as though many are beginning to understand that AI is simply just another tool that has a place in specific use cases. It is not a magical technology which can be widely implemented.
AI isn’t a tool, it’s an umbrella term applied to countless individual tools. Just a couple of years ago automated language translation wasn’t AI. Now it is for whatever reason.
Tbf I think that is because systems that were classified as AI became the backend for a lot language translation
In software coding I must say copilot in vscode is pretty impressive and very useful
My homelab has never been running more smoothly and I’m learning tonnes with effectively bespoke tutorials for my exact setup. At work, if I spend weeks developing a high quality report, translating it into derivative documents (like work plans or policy documents) has never been easier to get started. It gets me unstuck in a few keystrokes with a “shitty first draft” that I can refine. But you have to start with good info - it can’t make something good out of nothing. I see value, but maybe different from what it’s being sold as.
Alright, I’ll bite. What’s going on with all these graphics with people with some sort of colorful grid pattern behind them lately?
It’s the style their design director Tag Hartman-Simkins uses for a ton of their articles
And they will soar ahead with their clean energy powering their datacenters too
Lol. Have you seen the US Olympic Math Team? The US is fucked lol. Qwen 3.6 30B is surprisingly performant and runs on a fucking laptop, don’t need to drink your water down or steal your land.
In the cold war the Soviets used to say things like this to trick the Americans into wasting time and money on stupid shit
Huh? Half a year ago China was apparently much better at AI than anyone else, now they’re just catching up?
I don’t think they said they were “better”, if I recall they did “almost as much with far less resources”. and really the bigger thing was they were releasing a lot as open source.
IMO I think that’s what the real tech giants are worried about. Honestly I think I’m actually seeing the real goal of AI… and the worse part of it is, AI is the secondary feature.
Sam Altmen made a comment on wanting to see basically compute sold as like a utility to people. Crucial is getting out of selling hardware to people. I’m leaning towards the idea that the ballooning hardware costs isn’t a side effect, it’s the primary intention. The idea of it is basically to move all of computing to a model more like Netflix, or say what Googles stadia gaming console was etc…
The goal is a future where personal PCs are a rare luxury like power generators. For normal use we get crappy thin clients to connect to cloud services with our software. Super government and corporate friendly. Everything super accessible by government and scannable for AI training etc… Encryption can basically become meaningless because they can log straight into the machines.
China’s a big threat because… they could encourage in house models. IE an agent that runs in your house, or companies run them in their server rooms, and most importantly they can be trained individually with your personal data and company data, without creating the privacy and security nightmares of wondering if chatgpt could share all of your personal information with a random dude or competing company. Plus if we are talking individual hardware, that hurts the actual overall goal of shifting all focus of all hardware manufacturers towards data center modules that consumers and even medium sized businesses can only rent.
Killing the consumer hardware market is hand in hand with killing 3d printing and chat control.
Sam Altmen made a comment on wanting to see basically compute sold as like a utility to people.
Meh, that idea is as old as Compute itself. In the early days, basically the 50’s into the 80’s, you had dumb terminals that only provided a connection to the actually powerful centralized hardware. Think mainframe although it went by different terms at different times.
Every decade or so someone like Altman regurgitates the concept and to date no one has been able to make it widely successful. For example Microsoft has been chasing it since the NT4 days when they licensed multi-session technology from Citrix and rebranded it as “Remote Desktop Services”. It’s still in use today by corporations but outside of particular industries with high demands for secure remote access you’ll almost never see it.
So the idea and actual implementations of it have been around for decades but it never catches on because it’s always more expensive, less performant, and less flexible than hoped for.
With that said if you cock your head and squint we kinda sorta reached it with all the browser and app based online services like e-mail, listening to music, etc. Kinda. It breaks down though in the exact place that it always has…when you want to create something more involved than a text document. Suddenly it’s back to it being more expensive, less performant, and less flexible than doing it on local hardware.
Missing the key point though… your general concept is absolutely right… it’s failed over and over again because it can’t keep up with the fact that it’s cheaper and better performance with what you can buy locally, on top of bandwidth access etc… But there’s 2 ways to make one product a more common option, option A, try to make the one you want them to use better, but with AI they’ve done something different, they are working towards making local systems worse.
Key point is, the direction of attack. Instead of making remote compute better, cheaper, faster etc… what if the attack is make local hardware, more expensive, harder to get, less reliable, or even illegal.
Right now, all of that is happening, on top of course of windows being more bloated, hardware costs are multiplying at a rate never seen before, even for outdated hardware. Netflix and spotify have already adapted the public into no longer buying physical medium for movies and music, and as people are also noting, governments making more and more regulations come in place on say OS’s needing age gates, 3d printers needing to report to the government etc… It’s literally an orchestrated perfect storm to be pushing people away from owning computers.
The point is they aren’t trying to “catch up” to local hardware, they are trying to hinder the ability to obtain local hardware.
Well, I bought 1 year of GLM’s coding plan for basically nothing ($50?) last year.
…I now see its many hundreds of dollars, and its API is so busy it gets throttled.
Seems like devs are catching on.
FYI, for anyone interested in this stuff, I would suggest 1 year of Xiaomi’s coding subscription.
MiMo 2.5 is fantastic, beyond what benchmarks show. It’s a genuinely useful general purpose model with better prose/world knowledge than others. It’s quite uncensored. It’s not just some deep fried agent, and it’s dirt cheap right now because no one has realized it yet (kind of like GLM 6 months ago).
And before you ask, I use local models too. I sometimes run a custom IQ3_KT of MiMo 2.5, in fact, which is how I figured out it’s good. self hosting is the way. Gestures at the forum we’re on
But for stuff where privacy isn’t so critical, it’s still nice to have a year of an unquantized, fast API of a huge model for like $50.
