DeepSeek released every single technical detail about its models

Welcome back pals! Hope you had a great time in the weekend! Here’s what you need to know about AI today:

👉 DeepSeek released every single technical detail about its models

👉 Flora launches infinite AI canvas for artists

👉 Apple is struggling with modernizing Siri

and many more!

📧 Did someone forward you this email? Subscribe here for free to get the latest AI news everyday!

Read time: 5.1 minutes

What’s going on: In a series of posts on X marked by the “#OpenSourceWeek“ hashtag, DeepSeek has revealed all the technical details about the technology that enabled the rise of its two frontier models, V3 and R1. The company has demonstrated its commitment to transparency and community-driven innovation by releasing 6+ cutting-edge repositories on GitHub including FlashMLA, DeepEP, DeepGEMM, DualPipe, EPLB, and Fire-Flyer File System (3FS).

What does it mean: It seems like right now, the only company that is really “open” about AI is DeepSeek not OpenAI! They have open-sourced everything from the get go and basically shared all of the tiniest details of how they trained a world-class model with less than $6 million eventually leading to a $500b+ drop in Nvidia stocks value overnight a couple of weeks ago.

More details: 

• FlashMLA is an efficient MLA decoding kernel for Hopper GPUs that can reach up to 3000 GB/s of memory bandwidth when memory usage is the primary bottleneck, and up to 580 TFLOPS of computational throughput when compute power is the limiting factor.

• DeepEP is the first open-source communication library for Mixture-of-Experts (MoE) models.

• DeepGEMM is an optimized general matrix multiplication library (which is written in only 300 lines of code!) that can be integrated with popular frameworks like TensorFlow and PyTorch smoothly. It can reach up to 1350+ FP8 TFLOPS on Nvidia Hopper GPUs.

• DualPipe and EPLB are frameworks offering optimized parallelism strategies to optimize parallelism in distributed deep learning tasks.

• Fire-Flyer File System (3FS) is a distributed file system tailored for machine learning workflows that is heavily used in DeepSeek infrastructure and can utilize the full bandwidth of modern SSDs and RDMA networks.

• DeepSeek-V3/R1 Inference System uses cross-node expert parallelism to optimize throughput and latency for large-scale AI inference. Read this deep dive to learn more about how DeepSeek handles inference smoothly.

What’s going on: Flora is aiming to revolutionize the creative process for professionals with its AI-powered “infinite canvas”. The company is trying to address the shortcomings of existing AI tools that are designed by non-creatives for casual users rather than true artists. The infinite canvas provides a dynamic workspace where creatives can start with a simple prompt, like an image of a flower, and iteratively build upon it, exploring variations and details that are visually mapped out for clarity and collaboration.

What does it mean: Unlike platforms that prioritize quick generation over control or traditional creative software that can feel clunky and time-consuming, Flora positions itself as a powerful, intuitive tool tailored to the needs of professional creatives. Rather than developing new generative AI models, Flora integrates existing ones into a visual interface that allows users to generate and refine blocks of text, images, and videos, emphasizing the importance of the interface over the underlying technology.

More details:

• To demonstrate its capabilities, Flora’s alpha launch in August featured a unique art project: a live GoPro feed from it CEO’s head, stylized in real-time by AI for waitlist users.

• Flora doesn’t train its own models. Instead, it relies on others such as OpenAI’s Sora, Claude, Pika, Runway, Kling, Dream Machine, and Stable Diffusion.

• Backed by investors like A16Z Games Speedrun, Menlo Ventures, and angels from Midjourney and Stability, Flora offers a free tier with limited features and a professional plan starting at $16 per month.

• Interested in learning more and trying it for free? Visit their website.

🍎 Apple is struggling to rebuild Siri for adding generative AI and may not release a fully modernized, conversational version until 2027.

🐋 DeepSeek claimed its AI models could have a theoretical cost profit margin of 545%, based on potential revenue of $562,027 and GPU leasing costs of $87,072 in a 24-hour period, if all usage was billed at R1 pricing.

🧠 Sergey Brin, the co-founder of Google, urged Google employees to return to the office every weekday and work around 60 hours a week to help the company win the race to achieve Artificial General Intelligence (AGI).

📽 OpenAI has plans to integrate its Sora AI video generator into ChatGPT. While initially aimed at creatives through a dedicated web app, OpenAI now wants to reach a larger audience by making it accessible within ChatGPT.

📊 Google Sheets now has a Gemini-powered upgrade for Workspace business users, allowing them to analyze data faster and create visualizations using AI. Users can access Gemini via an icon in Sheets to generate insights like correlations, trends, and anomalies, as well as advanced visuals like heatmaps.

📞 Teleperformance SE, the world's largest call center company, is using Sanas, an AI startup, to neutralize the accents of its Indian call center employees for clearer communication with customers.

💰 Honor, a Chinese smartphone maker, plans to invest $10 billion in AI over five years to expand into AI-powered devices and strengthen ties with Google and Qualcomm.

✈ China has advised its leading AI entrepreneurs and researchers to avoid traveling to the US due to security concerns. Authorities fear that these individuals might inadvertently disclose sensitive information about China's AI advancements or be detained.

AI + Personal advice

Axiado - Machine Learning Engineer - Cybersecurity

Luminai - Software Engineer

PriceLabs - Data Scientist (Remote)

Lila Sciences - Machine Learning Researchers (Open-Endedness) - Open Level

Thank you for staying with us like always! If you are not subscribed, subscribe here for free to get more of these emails in your inbox! Cheers!