Comments by "nuqwestr" (@nuqwestr) on "Computerphile" channel.

@mb2776 Tolstoy's WAR AND PEACE is 1,500 pages, the Cliff Notes is only 30. There's your efficiency. But without the 1,500 pages there is no Cliff Notes.
7
@ NVIDIA stock increased 106% in 12-months, a drop in 5% in nothing, it's already headed back. A blip. Deepseek can run on a NVIDIA $300 chipset, too.
5
@ DeepSeek R1, General requirements: 7 billion parameter model: At least an NVIDIA RTX 3060 with 12GB of VRAM. 33 billion parameter model: At least an NVIDIA RTX 4080 with 16GB of VRAM. Between $300 and $800 consumer price vs $8,000 to $24,000 for A100s.
5
They said that about Red Hat in 1993, IBM purchased Red Hat in 2019.
4
NVIDIA only one player Groq 1. *Speed and Efficiency**: Groq's AI chip, known as the **Language Processing Unit (LPU)**, is designed specifically for AI inference. It delivers **exceptional processing speeds* with **low latency**, ensuring consistent and rapid data processing. 2. **Simplicity**: The architecture of Groq's chip is straightforward, which minimizes complexity and maximizes performance. This simplicity makes it easier to integrate into existing systems with minimal code changes. 3. **Scalability**: Groq's technology supports both single-core and multi-core configurations, making it versatile for a wide range of AI tasks. 4. **OpenAI Compatibility**: Groq offers seamless integration with OpenAI models, allowing developers to switch from OpenAI to Groq with just a few lines of code. 5. **Support for Open-Source Models**: Groq supports popular open-source AI models like **Llama 3.1**, making it accessible for developers and researchers working with these models. These features make Groq a strong choice for developers and organizations looking for fast and efficient AI inference solutions.
3
@ The stock was up 106% in last 12-months. Just a blip.
2
R1 can run on a chipset between $300 and $800, ChatGPT run on a chipset between $8,000 and $24,000, both can be purchased OPEN online.
1
Read the 1,500 pages of WAR AND PEACE once, distill it to 30-page Cliff Notes, distribute that billions of times instead. I don't like it. I'd rather question all 1,500 pages. Not all the time, but in critical situations, yes.
1
*DeepSeek data distillation* and *MPEG data compression* are related in that they both involve reducing the size of data, but they serve different purposes and use different methods. - **DeepSeek Data Distillation**: This process involves training a smaller, more efficient AI model (the "student") using knowledge extracted from a larger, pre-trained model (the "teacher"). The goal is to retain the performance of the larger model while reducing computational requirements and improving efficiency. - *MPEG Data Compression**: MPEG (Moving Picture Experts Group) standards, such as MPEG-4, use various algorithms to compress audio and video data. Techniques like **motion compensation**, **spatial compression**, and **advanced audio coding* are employed to reduce file sizes while maintaining acceptable quality. In summary, while both processes aim to optimize data, DeepSeek focuses on AI model efficiency, and MPEG targets multimedia file compression.
1
@the_bruce-z1w I don't know. My experience is limited to sound compression and the digital conversion from analog wave to digital particle/bit. The current sampling rate seems to produce a s/n value higher than vinyl records, and beyond which the human ear can differentiate. Digital sound has met the Turing Test. Time will tell if the number of parameters can be compressed/distilled to meet that same standard.
1