Codewello

It seems there's going to be a new DeepSeek V3 model release, and it might be launching soon.
docs.unsloth.ai/basics/deepseek-v3-0526-how-to-run…

6 months ago (edited) | [YT] | 11

Codewello

What a great week! First, we got DeepSeek V3.1, which is currently my favorite model for coding
— and you can use it for free.https://youtu.be/Hlz93KRJv00

Plus, the Gemini 2.5 model, which is insane to think about!

8 months ago (edited) | [YT] | 19

Codewello

DeepSeek Open-Sources Six Libraries for LLM Development

DeepSeek has released six open-source software libraries designed to address challenges in training, inference, and data infrastructure for large language models (LLMs).



Significance:



Code Availability: Provides practical implementations of techniques previously described in DeepSeek's research.

LLM Optimization: The tools focus on improving efficiency in various aspects of LLM development.

Potential for Wider Adoption: The open-source nature of the release may encourage broader sharing of infrastructure tools within the AI research community.



Libraries Released:



Day 1: FlashMLA: An implementation of Multi-Head Latent Attention (MLA).

Optimized for NVIDIA Hopper GPUs (e.g., H800), it aims for high memory bandwidth utilization.

Uses a compressed Key-Value (KV) cache.

Supports BF16 and FP16 data types.

Paged KV cache with block size of 64.

Day 2: DeepEP: Focuses on Expert Parallelism within Mixture of Experts (MoE) models.

Day 3: DeepGEMM: A library for FP8 Matrix Multiplication (GEMM) operations.

Day 4: DualPipe & EPLB: Tools for pipeline parallelism and load balancing.

Day 5: 3FS & Smallpond: Systems for managing parallel file access and high-throughput data.

Day 6: DeepSeek-V3/R1: information on the operational side.

8 months ago | [YT] | 12

Codewello

For the past six months, it's been like a heavyweight title match between Google and OpenAI. Every time a Gemini model takes the LMSys leaderboard crown, OpenAI claps back with a model that just barely edges it out. But this time, Google pulled a classic "hold my coffee" move: they let OpenAI sneak in with their usual counter last week, only to drop an even shinier, stronger variant right after. It's like watching a chess match where every move is a plot twist wrapped in a fake-out!

1 year ago | [YT] | 6

Codewello

Thank you all so much—3K coder! We made it!

1 year ago | [YT] | 11

Codewello

Gemini tells grad student to "please die" while helping with his homework

1 year ago | [YT] | 4

Codewello

I ran a fun experiment using AI to predict the winner of the 2024 election! The AI gathered public data and analyzed it to generate results. Check it out here!

https://youtu.be/JObuWC0N5gg

1 year ago | [YT] | 3

Codewello

Sama apologizes for hyping up products, but recommends chatgpt plus and chrome search

1 year ago | [YT] | 5

Codewello

1 year ago | [YT] | 6

Codewello

OpenAI's brain drain continues! 🤔
Another genius jumps ship.
Why? My two cents: https://youtu.be/IZnMXcYLbh0 😎

1 year ago | [YT] | 5