Iwan van der Kleijn’s Post

Since the cat is out of the bag on Hailuoai's MiniMax video model, and they're going to a public subscription model today, I figured it was time to pose the question; Is MiniMax a world simulation model? From my experience of generating over 3000 videos on the platform in the past 2 months, I'd cautiously say yes. Everything from physics, lighting, and object interactions is handled much more like a game engine than a "next frame" video engine like Runway Gen 3. Here's just one example of the things I've been playing with to test this theory, creating AI "practical" visual effects much like is done in tradition movie making. The way the objects, lighting and motion interact is unlike anything I've seen from other models to date with the exception of OpenAI's yet to be released Sora.

While LLMs continue to evolve in terms of their speed and accuracy and also in terms of their ability to address more and more complex problems as well as add cognitive features like multimodal AI, there is a parallel thread continuing that involves the use of these LLMs in complex projects by breaking the projects down to tasks and assigning them to different agents. Some of these agents will make API calls to LLMs and are called AI Agents. Their response becomes an input to other agents, and a project workflow is created for automation of complex processes consisting of multiple activities. The orchestration of these agents is an important part of the success of such projects, and orchestration tools are developing and evolving. In this post, I am attaching YouTube videos to explain two of the latest. One is Praison ai. The other is Open AI Swarm, which was released just yesterday. This process is very important in the industrial applications of AI, like the automation of complex processes and, in due course, real life systems that function autonomously. https://lnkd.in/dGvBZ6ty https://lnkd.in/dxQgvqZV

World simulator. If OpenAI’s video generation with SORA is simply extraordinary, it can definitely be stated that the videos are “simply” an outcome of a much more complex research (and generating videos was hardly simple!). Indeed, OpenAI is working on the physical simulation of the world. Stay tuned! “Our results suggest that scaling video generation models is a promising path towards building general purpose simulators of the physical world” #openai #ai #videogeneration #artificialinyelligence

OpenAI has introduced Swarm, an experimental framework intended to develop systems where multitudes of AI agents can jointly solve many complex problems much better. Unlike traditional single-agent models, Swarm takes center stage in orchestrating interactions among agents that allow them to scale efficiently by specializing. From big data analysis to complex simulations, Swarm offers scalability, flexibility, and parallel processing power for extending the state of the art in multiagent systems. See the full insights here: https://lnkd.in/gKWr6NRF

SORA by OpenAI could provide text to video, but the next iteration is here, HYPER REALISTIC 5D MOVIES GENERATED BY TEXT! If you think movie generation was impressive wait until you see this. A new AI model called E2D (Extensive Dynamic Diffusion) was released and it's insane. I took a look at some of the examples and couldn't believe what my eyes were seeing (had to use a tool in order to visualize the outputs). I invite you to take a look and let me know what you think. Kudos to the authors for releasing such a cool paper in such limited timeframe. What a time to be alive. PS Link to the outputs and the paper in the comments.

PyCon Italia 2024 videos are out. Here's Katharina Rasch and me talking about a few machine learning techniques for the rest of us - the GPU-poor. 1. Transfer Learning 2. Model Distillation 3. Quantization

GROQ: Setting the standards for #GenAI inference speed. 18x faster than anything out there. #LanguageProcessingUnit is a new type of end-to-end processing unit system that provides the fastest inference for computationally intensive applications with a sequential component to them, such as #LLMs #LPUInferenceEngine #llama2 #mistral #mixtral

Good taste? Call it with a proper name: design or lack of it.

The first-mover advantage in AI is fleeting, as competitors quickly catch up: ⚡ Just a year ago, no model could compete with GPT, but now Sonnet-3.5 is slightly ahead. 🌟 Sora, which seemed unbeatable 4 months ago, now has 4-5 clones with 70-80% quality. 💡 OpenAI's initial breakthrough spurred rapid advancements and cloning. 🛠️ Many companies have the technical skills but few can execute with strong will and good project selection. 🎯 First movers take significant risks, but their advantages are short-lived. #AIAchievements #TechInnovation #FirstMover ⚡ Sonnet-3.5 surpasses GPT, showing rapid advancements. 🌟 Sora clones like Kling, Luma, and Runway illustrate quick catch-up in AI capabilities. 💡 OpenAI's leadership catalyzed industry-wide progress. 🛠️ Execution and project choice are critical for sustaining innovation. 🎯 Despite the risks, first movers pave the way for rapid technological growth. Stay tuned to see how the AI landscape continues to evolve with new breakthroughs and competitive advancements! 🌐

Are you interested in a behind-the-scenes view of how AI works? Check out this 35-minute video with Samyam Rajbhandari from DeepSpeed! In it, he explains how DeepSpeed has played a significant role in advancing the training of large language models. GPUs do all the hard work for LLM’s, right? Well, they do need some help from tools like DeepSpeed. DeepSpeed allows researchers and practitioners to train larger models more efficiently, pushing the boundaries of what is possible in natural language processing and other domains. Join Samyam as he leads you through DeepSpeed's journey of addressing the challenges associated with LLM training and explains how they solved them. This video offers a piece of significant recent AI history that anybody working with LLMs or interested in the field won't want to miss. #AI #DeepSpeed #ailearning #aimodels https://lnkd.in/eZskJjRA