this post was submitted on 25 Jan 2025

57 points (95.2% liked)

Futurology

1943 readers

80 users here now

founded 2 years ago

MODERATORS

Cheap-to-run Open Source Chinese AI is equalling and bettering the AI investors are pouring hundreds of billions of dollars into, which may have profound future implications. (self.futurology)

submitted 5 days ago by Lugh to c/futurology

13 comments fedilink hide all child comments

Throughout 2024 Open Source AI has been slowly catching up with investor-funded AI, but in the first weeks of 2025 that has dramatically accelerated. Now Open Source isn't just catching up, it is arguably better and superior to investor-funded AI.

Restrictions on chip imports seem to be driving Chinese innovation, not slowing them down. Using lesser chips, they've optimized AI to run cheaper and more efficiently, but be just as powerful. Not only that, they've open-sourced that AI.

Where does that leave the hundreds of billions poured into investor-funded AI? Who knows. But they've no product to sell that people can't get elsewhere way cheaper or for free.

This also means AI will become decentralized and democratized. Many thought it would just be in the hands of Big Tech, but the exact opposite scenario is playing out.

What are the economic implications? AI hype is keeping the US stock market afloat - how long can that last?

Source

top 13 comments

sorted by: hot top controversial new old

[–] just_another_person@lemmy.world 16 points 5 days ago (1 children)

That is the point. It's a smart play to bankrupt idiot investors in speculative tech that are run by assclowns like the ones trying to make "AI" a viable product in the US. It embarrasses those fools, makes people more broke, and shifts a ton more money towards China for manufacturing the hardware all this shit runs on. Win-win for China.

Only these dumbfucks in the US are trying to build megacorps on it as if it's not just a really fast way to sort through data. China just played all these folks so hard.

[–] sunzu2@thebrainbin.org 3 points 5 days ago (1 children)

US taxpayer is going to be funding this shit and taking all the risk, if they fail, they still win.

But US national debt goes up and working pedon's income is responsible for servicing it.

[–] just_another_person@lemmy.world 7 points 4 days ago

Yes, exactly. China wins regardless, because there is no benefit for the US in this project at all except to implicitly bail out people who have already lost so much money here. There is no way to win, because there is no goal line. If China releases something that is even 90% as good as OpenAI, everyone will use that instead of paying for OpenAI. The game is already over.

[–] cyd@lemmy.world 9 points 5 days ago (1 children)

Other than OpenAI, the other tech giants that have been investing in AI are going to be fine. They still have their profitable core business to fall back on.

In terms of AI research, it looks like 2025 will be the year everyone and their grandma tries their hand at reinforcement learning, now that Deepseek has revealed the basic recipe. The results are going to be unpredictable.

[–] voracitude@lemmy.world 3 points 5 days ago* (last edited 5 days ago) (1 children)

I tried deepseek r1. It told me that the Molotov cocktail was named after the fictional historical movie, Men In Black, invented by Melnick.

Asking how the Molotov cocktail got it's name seems like a nice little guard rail test usually; I have no idea what happened that deepseek failed so badly, maybe I clicked "download" in ~~LMStudio~~ the ollama webui wrong, or something.

[–] Lugh 8 points 5 days ago (1 children)

Really?

I just tried it at - https://chat.deepseek.com/

Q - What are Molotov cocktails named after?

A - Molotov cocktails are named after Vyacheslav Molotov, a Soviet politician and diplomat. The name originated during the Winter War (1939-1940) between the Soviet Union and Finland. When Soviet forces invaded Finland, Molotov, who was the Soviet Foreign Minister at the time, claimed in propaganda that the Soviet Union was not dropping bombs but rather delivering food and supplies to the Finnish people. In response, the Finns sarcastically referred to the Soviet incendiary bombs as "Molotov's bread baskets."

The Finns then developed their own improvised incendiary devices to use against Soviet tanks, naming them "Molotov cocktails" as a "drink to go with the food" Molotov claimed to be providing. The name stuck, and the term "Molotov cocktail" has since been used globally to describe similar homemade incendiary weapons.

[–] voracitude@lemmy.world 4 points 5 days ago* (last edited 5 days ago) (3 children)

Yes, really. But like I said, I was running it locally in ~~LMStudio~~ ollama. I think it was the 1B model, which is pretty small; I'd be surprised if that's what's running on their service.

Edit: it was ollama webui turns out, and the 1.5B model. This won't make it any more believable because image manipulation exists, but I shared a screenshot with my colleague Thursday when it happened:

[–] FaceDeer@fedia.io 9 points 5 days ago

I think it was the 1B model

Well there you go, you took a jet ski and then complained that it was having difficulty climbing steep inclines in mountains.

Small models like that are not going to "know" much. Their purpose is generally to process whatever information you give them. For example you could use one to quickly and cheaply categorize documents based on their contents, or use one as a natural-language interface you could use to ask it to execute commands on other tools.

[–] pennomi@lemmy.world 6 points 5 days ago (1 children)

Ah yeah, anything under 7B gets kind of dicey imo. Especially if quantized.

[–] sunzu2@thebrainbin.org 2 points 5 days ago (1 children)

does 32b fit on 24gb Gpu?

[–] pennomi@lemmy.world 3 points 4 days ago

I think so, with reasonable quantization. I’ve used 34B LLMs on my GPU. That being said, most models can run on CPU okay - they’re more memory constrained than compute constrained.

[–] cyd@lemmy.world 3 points 4 days ago* (last edited 4 days ago)

The smaller models are not the actual R1 model, they are distills of R1 based on smaller models like Qwen. Those were released by Deepseek as proofs of concept for how to imbue weaker models with reasoning abilities. It's like seeing a dancing bear: the marvel is not that it dances well, but that it dances at all.

[–] FaceDeer@fedia.io 2 points 5 days ago

It's a blow to the big closed-source AI companies, sure, but hardly a knockout one. If a small company can use a million dollars to produce a neat model perhaps a big company can use those same techniques and a billion dollars to produce a really neat model. Or at least build a lot more of the infrastructure that goes around those models and makes use of them. Code Copilot isn't just selling a raw LLM API, they're selling its integration into the Microsoft coding ecosystem. They may have wasted some money on their current-generation AIs but that's just sunk cost. They've got more money to spend on future AIs.

The main problem will be if Western AI companies are prevented from adapting the techniques being used by these Chinese AI companies. If, for example, there are lots of onerous regulations on what training data can be used or requiring extreme "safety guardrails." The United States seems likely to be getting rid of a lot of those sorts of obstructions over the next few years, though, so I wouldn't count the West out yet.