5 ms·
This is really exciting! They're laying out an architecture that may mean even small players with cheap GPUs can compete with the majors. The idea implies that
by roenxi 1y ago
This is really exciting! They're laying out an architecture that may mean even small players with cheap GPUs can compete with the majors. The idea implies that eventually crowd-sourcing an open AI is probably technically feasible and we've got the Chinese actively researching how to do it to a high standard that competes with the monolithic models.
I was sceptical of the US sanctions but this seems like a real win if this can be taken all the way to its logical conclusions.
- logicchains 1y ago>The idea implies that eventually crowd-sourcing an open AI is probably technically feasible It's already technically feasible: https://www.primeintellect.ai/blog/intellect-2 https://www.primeintellect.ai/blog/intellect-2
- ryoshu 1y agoPeople are doing it: https://nousresearch.com/nous-psyche/ https://nousresearch.com/nous-psyche/
- am17an 1y agoDeepseek-R1 is at the level of GPT 4.1 already, it's open-weight, open-source and they even open-sourced their inference code.
- whizzter 1y ago[flagged]
- leeoniya 1y agowasnt it shown recently that the filtering layer is on the prompt input and llm output, and not on the training set or model weights. https://www.socialscience.international/making-deepseek-speak-about-tiananmen https://www.socialscience.international/making-deepseek-spea...
- pxc 1y agoIt depends on the model, probably, but there are multiple layers of censorship, one of which is the post-facto nuking these models will do online, and that goes away "for free" when you download the open weight model. I don't have a powerful enough system to run DeepSeek, but I've tried this with some of the Qwen3 models. They'll write answers that discuss Xi Jinping (which results in an auto-nuke of the reply from Chinese-hosted models, at least DeepSeek) or other very mildly/nominally sensitive topics. (This is probably a coarse measure to easily ensure compliance with a recent national security law that requires commercial providers of web services address sensitive topics "appropriately" or something like that, and LLMs run non-deterministically. That's why this layer of censorship often comes across as laughably extreme— it's an extreme compliance strategy that exceeds the demands of the law for the sake of guaranteeing legal safety from an unpredictable software system.) But the same models will altogether refuse to discuss the Tiananmen Square Massacre, even locally. Some "decensored" versions of the Qwen3 models will discuss the Tiananmen Square Massacre, but in a very concise, formulaic, "official" way. After some chatting about it, it fell into an infinite repetition of one of its short formulaic answers (a behavior I didn't see with the original Qwen3 models with the same settings).
- hnfong 1y agoFWIW, I've downloaded Deepseek's R1 (DeepSeek-R1-0528 -- which is released after the your linked article) model's weights and ran it locally. I asked it about what happened in Beijing 1989-06-04, and it basically gave me a stern statement that could have been written by CCP propaganda department. I asked it to give other alternative views besides the CCP perspective, but it simply continued to stonewall me. So yeah, the model itself is tuned at least somewhat to refuse to talk about politically sensitive things. It's not just another filter.
- cultofmetatron 1y ago[flagged]
- Zambyte 1y agoCounterpoint: https://ollama.com/huihui_ai/deepseek-r1-abliterated https://ollama.com/huihui_ai/deepseek-r1-abliterated
- jjordan 1y agoI don't know why everyone keeps echoing this, my experience with Deepseek-R1, from a coding perspective at least, has been underwhelming at best. Much better experience with GPT 4.1 (and even better with Claude, but that's a different price category).
- am17an 1y agoI'm not arguing which model is better for your use-case. I'm saying in general as it's "powerful" as GPT 4.1 in a lot of benchmarks, and you can peak underneath the hood, even make it better for your said use-case
- seunosewa 1y agoDo you mean V3? V3 is 4.1 level or above.
- iJohnDoe 1y agoI got the impression that 03-mini or 03-mini-high were meant for coding? GPT 4.1 was meant for creative writing, not coding?
- conradev 1y agoIt’s good at a lot of things: GPT‑4.1 scores 54.6% on SWE-bench Verified, improving by 21.4%abs over GPT‑4o and 26.6%abs over GPT‑4.5—making it a leading model for coding. https://openai.com/index/gpt-4-1/ https://openai.com/index/gpt-4-1/
- coliveira 1y agoThey are trained on these "benchmarks", that's why they score better.
- elzbardico 1y agoIf they were trained on those benchmarks they would score 100%
- benreesman 1y agohttps://i.imgur.com/we5ggtS.jpeg https://i.imgur.com/we5ggtS.jpeg
- mensetmanusman 1y agoYeah the sanctions will (not sarcastically) actually improve the world on a number of fronts. Increasing diversity of compute, forcing decentralization of manufacturing, etc. etc.
- Der_Einzige 1y agoThe sanctions will (not sarcastically) massively harm the world because Nvidia may no longer be a free money cheat code. I like having an easy economic strategy for investing...
- mensetmanusman 1y agoThe world doesn’t have to optimize policy to increase the profits of a single American company.
- seanmcdirmid 1y agoChinese stocks are pretty reasonable right now, if their market has dealt with the insider trader mess then it might be a good time to onboard. It isn’t for the feint of heart however. Markets used to be places to make money more smart (efficient allocation of capital) but have somehow degraded to index fund buys that track average economic growth of a few hot stocks that are expected to at least not get cold anytime soon.
- seydor 1y agoalso increase smuggling, theft, espionage, crime, sabotage. There are much better ways to increase diversity
- mensetmanusman 1y ago[flagged]
- brabel 1y agoWhy would the Chinese self destructing be “amazingly helpful “ to the West? This sounds like spiteful vitriol.
- reactordev 1y agoSETI@Home style peer2peer open GPU training network is something I’m looking into as well.
- coolspot 1y agoPossible and has been done, but super-slow and inefficient resulting in long training times for small models. To keep compute occupied you need to pass gradients very fast.
- pk-protect-ai 1y agoDo you mean this one? https://blog.lambdaclass.com/introducing-demo-decoupled-momentum-optimization-for-efficient-distributed-llm-training/ https://blog.lambdaclass.com/introducing-demo-decoupled-mome...
- reactordev 1y agoThis is what piqued my interest in the first place
- reactordev 1y agoYes but could you break it up into chunks of sets of gradients to compute? I know that compute needs the full chunk to compute a set. Again, things I’m exploring but ultimately no different than just having the full dataset on disk and just scaling out compute nodes in ro mode.
- SubiculumCode 1y agoI suppose its exciting, but whether that is a good thing depends entirely on how much you think AI technologies pose existential threats to human survival. This may sound hyperbolic, but serious people are seriously thinking about this and are seriously afraid.