• hrry

  • Post #1
  • (Edited)

Hey everyone! 👋

Have you heard of DeepSeek? 🖥️

It’s shaking the AI world (& stock market) at the moment and creating levels of noise that are reminiscent of OpenAI’s GPT-4 release. It’s wild; The big AI players are scrambling and this is going to be the tonic that the AI world needed to accelerate development.

❓ Why? Because we’re seeing open source catch up to closed source. Techies are typically advocates for using open-source (you can see how its built, you can host it in your own data centres, etc) vs. closed source, this is a signal that in the future AI ecosystem will favour Deepseek like technology over things that look like ChatGPT.

🌳 Sustainability is a huge plus here too. Companies consume so much energy to train and run AI which over time will have sizeable impact on the environment but, Deepseek has proven you don’t need nearly as much energy (or money) to power this technology. This is huge.

👨 And from my brief testing Deepseek is performing wonderfully, producing  impressive results without breaking the bank. I’ve found it’s incredible at reasoning and math, while I still prefer Claude 3.5 sonnet for coding & writing but, the advantage Claude 3.5 has is tiny!

So, what do you all think?

  • Have you tried Deepseek, and if so, what do you think?

  • How do you see LLMs reshaping industries in the next 5 years?

  • Are we moving too fast in generative AI, or not fast enough?

  • And, would you like to see more native AI features in Nothing OS? If so, what?

Would love to hear from you all – especially around your views of AI and Nothing hardware. Is it just hype, or are we going to see consumers get on board with the big shift that tech leaders are betting on?

    Related discussions

    hrry

    DeepSeek is indeed a game-changer! Open-source AI catching up to closed-source is huge for transparency, sustainability, and cost-efficiency. It’s impressive in reasoning and math, and close to Claude 3.5 in coding/writing. AI will reshape industries in the next 5 years, but we need to balance speed with ethics. Native AI features in Nothing OS? Absolutely—smarter, personalized experiences could be a big win. Exciting times ahead! 🚀 What do you think?

      hrry Thanks for this Harry - I’m still reading and forming opinions. Impressive how a new LLM product can cause such a stir, kicking up political and economic reactions! Let’s see how this all plays out!

        hrry a lot to learn abt this. Looking forward to watching some videos on this and comments here for more information.

        This stock price movement is crazy

          • Daniel

          • Post #5
          • (Edited)

          hrry

          Great thread! It’s mind-blowing how much a single tool has disrupted the landscape. Personally, I haven’t tried it yet since I find Grok to be quite sufficient, since it’s included with my X subscription.

          The current state of AI in smartphones doesn’t impress me much. For instance, image generation is a novelty that quickly loses its charm; AI translations are handy but not indispensable in everyday use; and the manipulation of genuine memories through AI-enhanced photography seems to deviate from capturing reality.

          I believe the concepts and points discussed in Nothing’s YouTube video titles “Let’s talk about AI” outline the ideal approach to embedding AI into the smartphones we use daily.

          “building a bridge from today’s app based user experience to tomorrow’s post app world ” - Carl Pei.

          One invaluable feature that AI provides that I use often is enhancing my writing, as demonstrated by this thread 😉

            Found this interesting message on X

            Don’t fall for false DeepSeek R1 news! Deepseek R1 made it to mainstream and there are so many wrong takes and false information.

            No, the training didn’t cost only ~$6M dollars; the compute for the base model (no RL) was equivalent to GPU hours worth $5.5M, excluding ablations, smaller runs, data generation, and any of DeepSeek R1 training.

            No, it’s not a side project (maybe started as one). DeepSeek is backed and owned by High-Flyer, a Chinese hedge fund; in 2020, they managed assets of over 7 billion dollars, and their talent includes Olympic medalists in mathematics, physics, and informatics.

            No, they don’t have just a few GPUs. They have ~50k GPUs.

            The real Deepseek R1 is 671B MoE model that needs > 16× 80GB Memory (16x H100s).

            Yes, the Deepseek R1 671B is really good! And they do great open source and science work > 2 years already

            There are 6 “distilled” versions. They are fine-tuned Qwen and Llama on 800k samples, (NO RL). Thats not “r1”. The smallest one with 1.5B (Yes, you can run locally, but it’s not near R1).

            Yes, the hosted version on chat(.)deepseek(.)com might use your data to train new models (as per ToS of DS).

            Yes, Open Science and Source will benefit everyone in the long term! Hugging Face is working on a fully open reproduction pipeline

            source

            I feel like with the advent of AI and it’s various guises and models. I have reached an era of consumer technology I feared…that I don’t know or understand it.

            There seems to be an oversaturation of AI at the moment, and this wealth of choice has left me feeling cold. I do not engage with AI, particularly…I mostly use Gemini on my Phone (2a) Plus, Google Assistant around the home and that is about it.

            Like others, I feel like I shall wait for this storm to calm and see what more knowledgeable heads write-up and produce some content about this.

            In terms of how Nothing can leverage AI, as it is the buzz word of the tech industry at the moment, I feel that their approach of making an assistant/interface friendly and relatable is a good first step.

            I need to feel that connection, somehow. It needs to have warmth, which in turn shall help me stay motivated to use its services.

            Good , but not make photos 🙂

            Sounds great. In any case the only way I know how to use AI is manipulation of photos and wallpapers. 😝

            I tried a few prompts but its very biased when it comes to China alone. I wouldn’t rely on an AI that has a bias. Sure, many will use it. But personally I would prefer ChatGPT over it.

              • GZZZ

                • Post #13

                hrry Yes, the main focus its trying to run one of those models locally on your phone, im looking for some tools to run locally with a cropped version.
                Mac and Ollama are rocking solid, privacy is key for the AGI future

                  good one

                  any thoughts on integrating this in the next big os update

                  I am eagerly waiting to see how the Nothing team integrates AI into our OS. 📱

                  its totally biased. might even manipulate the results. tried it but didnt like it. it even gives biased results.

                  I am impressed what Deepseek managed to accomplish, but after using their model, I came to the conclusion that it was too heavily biased. It had no Problem talking about bad American Presidents, but it did not answer anything about bad Chinese “leaders”. It also refuses to talk about anything related to tiananmen square massacre and praises the democrat rights Chinese people have. I hope, that you can decide which model you use in Nothing OS. (btw I don’t say other models doesn’t have these problems)

                  DeepSeek is a game-changer! Open-source AI catching up is huge for accessibility and sustainability. AI in Nothing OS? Smart assistants and AI-driven UI could be exciting.

                  I am using intensively perplexity assistant on my phone and ditched Gemini. I need answers and help not generated photos, videos etc…

                  From what I was reading deepseek can run to certain extent on a smartphone. Making an in-house assistant of it I guess will be the next step for the deepseek team.

                  I tried it carefully and to me it’s absolutely amazing. But I still go back to perplexity as it is text based and no useless gimmicks, can be used as an assistant and always gives the source articles the answers are based on.

                  In a nutshell, yes if it is to make my life easier I am all up for it, but if it is the Samsung way, pictures to doodles and other non sense 5 years old kindergarten stuff… thanks but no.

                  I want it to be an assistant, interacting locally with all apps and taking decisions on how to save battery etc… all in the background.

                  I guess I am asking a bit too much.