The FWA’s Post

DepthCrafter IMPRESSIVE Detailed & Stable Mono Depth Estimation https://lnkd.in/gjNbHYEx DepthCrafter from Tencent AI just came out. Using a diffusion model and some clever training techniques, it is able do impressive detailed depth estimation on videos with just a single camera. 0:00 Introduction 0:45 DepthCrafter Model and Training 1:23 DepthCrafter vs ChronoDepth, Depth Anything V2, NVDS 4:13 DepthCrafter vs Depth Anything V2 6:50 Novel View Rendering 7:57 Visual Effects

Double-Batch K-SVD (DBK-SVD) is a modified K-SVD algorithm designed to make neural network embeddings interpretable and modifiable by representing them as a sum of semantically meaningful basis vectors, each with a human-readable explanation. These embeddings, usually high-dimensional and non-interpretable, can now be decomposed into distinct components, enabling applications in image classification, clustering, and sentiment analysis. Using DBK-SVD and a Vision-Language Model, embeddings are analyzed through short natural language descriptions for each basis vector, improving interpretability and allowing for modifications at a semantic level, such as isolating style from content. Applied to a CLIP-based vision transformer, DBK-SVD demonstrates improved interpretability and specificity over sparse autoencoder approaches, yielding lower reconstruction loss and more diverse monosemantic concept explanations. This method also offers an interactive tool to inspect embeddings, paving the way for new approaches to understanding model behavior and inspecting information flow within neural networks. https://lnkd.in/ghBPS3kg

For AI enthusiasts eager to delve into interpretability, I highly recommend checking out "Exploring Gemma Scope: An Introduction to AI Interpretability and the Inner Workings of Gemma 2 2B." This demo offers a beginner-friendly yet insightful exploration of the Gemma 2 2B model, making the complexities of AI more accessible. Whether you're new to interpretability or looking to deepen your understanding, this is an excellent resource to get a look under the hood of the intricate world of AI.

How do you imagine your personal AI assistant? I picture mine as a #Jarvis. (Yes, I’m a Marvel fan 🤓🦸♀️) An assistant that doesn’t just answer questions but supports you in real-time. It anticipates your needs, keeps everything seamlessly connected: work, health, finances, and even your shopping list. The other day, I was listening to the “Top Minds in AI” panel at the Future Investment Initiative (#FII). It got me thinking about how AI is evolving. They talked about assistants that go beyond being reactive, stepping into the realms of personalization, prediction, and proactivity. So, how do you imagine your AI assistant in the near future? A Jarvis-style ally or something more minimalist? #ArtificialIntelligence #AI #Future #Innovation #jarvis #marvel #AIAssistants #personalization #prediction #proactivity #future #FII #digitalTransformation #tech #ironMan #marvel https://lnkd.in/e4r3nTSA

For Gemma users, can you explain the difference from RIG and RAG? https://lnkd.in/dm7qU4Sm #kaggle #gemma #AI #GenAI

I’m pleased to announce the release of our first offline multiple-choice Leaderboard on Hugging Face Spaces https://lnkd.in/ed2exWq7 🇰🇿 . Thanks to everyone who helped make this happen: Beksultan Sagyndyk, Sanzhar Umbet (💙), Kirill Yakunin, Sabina Abdullayeva, Madina A., Daulet Toibazar, Ardak Shalkarbayuli, Asset Karazhay This leaderboard evaluates the models against the Kazakh benchmarks we released earlier, helping the community better understand how different language models perform on Kazakh data and how deeply they understand the cultural component. 𝗞𝗮𝘇-𝗟𝗟𝗠 𝗟𝗲𝗮𝗱𝗲𝗿𝗯𝗼𝗮𝗿𝗱 Our leaderboard was built using references like EleutherAI's lm-evaluation-harness (https://lnkd.in/eCabpeA2) and Vikhr Models LM Eval MC (https://lnkd.in/eajqw6Xf). It features some of the leading models from major AI developers: - OpenAI: gpt-4o, gpt-4o-mini - Anthropic: Sonnet 3.5 - Google: Gemma and Gemini - Amazon: Nova - Institute of Smart Systems and Artificial Intelligence - Nazarbayev University (ISSAI): KazLLM - Meta: LLaMA - Yandex: YandexGPT4 (upcoming) 𝗥𝗲𝘀𝘂𝗹𝘁𝘀 - Current overall leader: GPT-4o leads with a 2% average lift over its closest competitor, Sonnet 3.5. - Current open-source leader: ISSAI’s LLaMA-3.1-KazLLM-1.0-8B achieves a 5% improvement over its origin LLaMA-3.1-8B-Instruct and a 1% over Google/Gemma-2-9b-it. - 27B+ model results are upcoming, stay tuned 🔜 We’ve made it simple for anyone to submit their model’s predictions to the leaderboard. The process is fully documented in our GitHub repository (https://lnkd.in/ea4299Se) and leaderboard page (submit button), which includes an easy-to-use evaluation script After running the script, you’ll get a JSON file with the benchmark results. Simply submit this file to the leaderboard. 𝗨𝘀𝗲 𝗰𝗮𝘀𝗲𝘀 This leaderboard goes beyond model comparison—it potentially could be a valuable resource for local tech companies and entrepreneurs dealing with sensitive or uncensored data. These organizations often need in-house open-source models but struggle to identify the right fit for their needs. By providing a clear performance comparison, we aim to support their decision-making process and foster the growth of AI solutions tailored to Kazakhstan-specific needs. 𝗡𝗲𝘅𝘁 𝘀𝘁𝗲𝗽𝘀 - Delivering more complex cultural and business benchmarks to evaluate models on deeper, real-world use cases - Releasing an offline evaluation arena with ELO rating computation to bring even more transparency and competition to Kazakh LLMs. - Releasing custom instruct datasets, further enhancing the resources available for fine-tuning and evaluation. We’re also actively looking for 𝗶𝗻𝘃𝗲𝘀𝘁𝗼𝗿𝘀, 𝗰𝗼𝗻𝘁𝗿𝗶𝗯𝘂𝘁𝗼𝗿𝘀 𝗮𝗻𝗱 𝗰𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗼𝗿𝘀! Whether you’re an individual or an organization, if you share our passion for advancing Kazakh NLP, we’d love to hear from you. DM US, we need you!

Linear development =High Weight of the Past relative to the Future. Exponential Development=High Weight of the Future relative to the Past. AI Compute is Doubly Exponential.

Ai is a beautiful thing.

I always enjoy discussing Artificial Intelligence (AI) in the context of K-pop because the way K-pop fans see AI art challenges a lot of stereotypes on them and the K-pop industry. For example, a Brazilian fan went viral on X by commenting that they'd give anything to know who's behind the K-pop AI group PLAVE, especially after seeing one of the members flirting with Hyoyeon in this video. A comment as simple as that is enough to prove that even though AI art can indeed captivate people, there will always be a desire for human connection at the heart of fandoms. One of the things that K-pop fans actually enjoy the most is seeing the relationships and dynamics that play out inside or between groups, something that AI idols cannot bring to the table. 😅 More thoughts of mine in: ◾ ... this post on AI music in K-pop : https://lnkd.in/dr8jRtNf ◾ ... my interview for Terra in which I discuss the implications of AI for Intellectual Property in K-pop (in Portuguese): https://lnkd.in/ejRc9CrY https://lnkd.in/dzRyFfGm

Hello folks! 🚀 Excited to share my latest article on Medium: "An In-Depth Exploration of the Vision-and-Language Transformer (ViLT) Model." In the rapidly evolving AI landscape, the ViLT model stands out by seamlessly integrating vision and language using a pure transformer-based architecture—no convolutions or region supervision needed. In this article, I delve into the model’s key components, discuss its limitations, and provide hands-on implementation examples, including both pre-trained and custom-trained approaches. Whether you're diving into multimodal AI or just curious about cutting-edge models, I believe this piece will offer valuable insights!