Latest news with #H100
Yahoo
7 hours ago
- Business
- Yahoo
Odyssey's new AI model streams 3D interactive worlds
Odyssey, a startup founded by self-driving pioneers Oliver Cameron and Jeff Hawke, has developed an AI model that lets users "interact" with streaming video. Available on the web in an "early demo," the model generates and streams video frames every 40 milliseconds. Via basic controls, viewers can explore areas within a video, similar to a 3D-rendered video game. "Given the current state of the world, an incoming action, and a history of states and actions, the model attempts to predict the next state of the world," explains Odyssey in a blog post. "Powering this is a new world model, demonstrating capabilities like generating pixels that feel realistic, maintaining spatial consistency, learning actions from video, and outputting coherent video streams for 5 minutes or more." A number of startups and big tech companies are chasing after world models, including DeepMind, influential AI researcher Fei-Fei Lee's World Labs, Microsoft, and Decart. They believe that world models could one day be used to create interactive media, such as games and movies, and run realistic simulations like training environments for robots. But creatives have mixed feelings about the tech. A recent Wired investigation found that game studios like Activision Blizzard, which has laid off scores of workers, are using AI to cut corners and combat attrition. And a 2024 study commissioned by the Animation Guild, a union representing Hollywood animators and cartoonists, estimated that over 100,000 U.S.-based film, television, and animation jobs will be disrupted by AI in the coming months. For its part, Odyssey is pledging to collaborate with creative professionals — not replace them. "Interactive video [...] opens the door to entirely new forms of entertainment, where stories can be generated and explored on demand, free from the constraints and costs of traditional production," writes the company in its blog post. "Over time, we believe everything that is video today — entertainment, ads, education, training, travel, and more — will evolve into interactive video, all powered by Odyssey." Odyssey's demo is a bit rough around the edges, which the company acknowledges in its post. The environments the model generates are blurry and distorted, and unstable in the sense that their layouts don't always remain the same. Walk forward in one direction for a while or turn around, and the surroundings might suddenly look different. But the company's promising to rapidly improve upon the model, which can currently stream video at up to 30 frames per second from clusters of Nvidia H100 GPUs at the cost of $1-$2 per "user-hour." "Looking ahead, we're researching richer world representations that capture dynamics far more faithfully, while increasing temporal stability and persistent state," writes Odyssey in its post. "In parallel, we're expanding the action space from motion to world interaction, learning open actions from large-scale video." Odyssey is taking a different approach than many AI labs in the world modeling space. It designed a 360-degree, backpack-mounted camera system to capture real-world landscapes, which Odyssey thinks can serve as a basis for higher-quality models than models trained solely on publicly available data. To date, Odyssey has raised $27 million from investors including EQT Ventures, GV, and Air Street Capital. Ed Catmull, one of the co-founders of Pixar and former president of Walt Disney Animation Studios, is on the startup's board of directors. Last December, Odyssey said it was working on software that allows creators to load scenes generated by its models into tools such as Unreal Engine, Blender, and Adobe After Effects so that they can be hand-edited. This article originally appeared on TechCrunch at Sign in to access your portfolio
Yahoo
8 hours ago
- Business
- Yahoo
Odyssey's new AI model streams 3D interactive worlds
Odyssey, a startup founded by self-driving pioneers Oliver Cameron and Jeff Hawke, has developed an AI model that lets users "interact" with streaming video. Available on the web in an "early demo," the model generates and streams video frames every 40 milliseconds. Via basic controls, viewers can explore areas within a video, similar to a 3D-rendered video game. "Given the current state of the world, an incoming action, and a history of states and actions, the model attempts to predict the next state of the world," explains Odyssey in a blog post. "Powering this is a new world model, demonstrating capabilities like generating pixels that feel realistic, maintaining spatial consistency, learning actions from video, and outputting coherent video streams for 5 minutes or more." A number of startups and big tech companies are chasing after world models, including DeepMind, influential AI researcher Fei-Fei Lee's World Labs, Microsoft, and Decart. They believe that world models could one day be used to create interactive media, such as games and movies, and run realistic simulations like training environments for robots. But creatives have mixed feelings about the tech. A recent Wired investigation found that game studios like Activision Blizzard, which has laid off scores of workers, are using AI to cut corners and combat attrition. And a 2024 study commissioned by the Animation Guild, a union representing Hollywood animators and cartoonists, estimated that over 100,000 U.S.-based film, television, and animation jobs will be disrupted by AI in the coming months. For its part, Odyssey is pledging to collaborate with creative professionals — not replace them. "Interactive video [...] opens the door to entirely new forms of entertainment, where stories can be generated and explored on demand, free from the constraints and costs of traditional production," writes the company in its blog post. "Over time, we believe everything that is video today — entertainment, ads, education, training, travel, and more — will evolve into interactive video, all powered by Odyssey." Odyssey's demo is a bit rough around the edges, which the company acknowledges in its post. The environments the model generates are blurry and distorted, and unstable in the sense that their layouts don't always remain the same. Walk forward in one direction for a while or turn around, and the surroundings might suddenly look different. But the company's promising to rapidly improve upon the model, which can currently stream video at up to 30 frames per second from clusters of Nvidia H100 GPUs at the cost of $1-$2 per "user-hour." "Looking ahead, we're researching richer world representations that capture dynamics far more faithfully, while increasing temporal stability and persistent state," writes Odyssey in its post. "In parallel, we're expanding the action space from motion to world interaction, learning open actions from large-scale video." Odyssey is taking a different approach than many AI labs in the world modeling space. It designed a 360-degree, backpack-mounted camera system to capture real-world landscapes, which Odyssey thinks can serve as a basis for higher-quality models than models trained solely on publicly available data. To date, Odyssey has raised $27 million from investors including EQT Ventures, GV, and Air Street Capital. Ed Catmull, one of the co-founders of Pixar and former president of Walt Disney Animation Studios, is on the startup's board of directors. Last December, Odyssey said it was working on software that allows creators to load scenes generated by its models into tools such as Unreal Engine, Blender, and Adobe After Effects so that they can be hand-edited. Error while retrieving data Sign in to access your portfolio Error while retrieving data
Geeky Gadgets
14 hours ago
- Business
- Geeky Gadgets
Unlock the Secret to Fine-Tuning Small AI Models for Big Results
What if you could transform a lightweight AI model into a specialized expert capable of automating complex tasks with precision? While large language models (LLMs) often dominate the conversation, their immense size and cost can make them impractical for many organizations. Enter the world of fine-tuning small LLMs, where efficiency meets expertise. By using innovative tools like Nvidia's H100 GPUs and Nemo microservices, even a modest 1-billion-parameter model can be fine-tuned into a domain-specific powerhouse. Imagine an AI agent that not only reviews code but also initiates pull requests or seamlessly integrates into your workflows—all without the hefty price tag of training a massive model from scratch. James Briggs explores how LoRA fine-tuning can unlock the potential of smaller LLMs, turning them into expert agents tailored to your unique needs. From preparing high-quality datasets to deploying scalable solutions, you'll discover a structured approach to creating AI tools that are both cost-effective and high-performing. Along the way, we'll delve into the critical role of function-calling capabilities and how they enable automation in fields like software development and customer support. Whether you're an AI enthusiast or a decision-maker seeking practical solutions, this journey into fine-tuning offers insights that could reshape how you think about AI's role in specialized workflows. Fine-Tuning Small LLMs The Importance of Function-Calling in LLMs Function-calling capabilities are critical for allowing LLMs to perform agentic workflows, such as automating code reviews, initiating pull requests, or conducting web searches. Many state-of-the-art LLMs lack robust function-calling abilities, which limits their utility in domain-specific applications. Fine-tuning bridges this gap by training a model on curated datasets, enhancing its ability to execute specific tasks with precision. This makes fine-tuned LLMs valuable tools for industries where accuracy, efficiency, and task-specific expertise are essential. By focusing on function-calling, you can transform a general-purpose LLM into a specialized agent capable of handling workflows that demand high levels of reliability and contextual understanding. This capability is particularly useful in fields such as software development, customer support, and data analysis, where task-specific automation can significantly improve productivity. Fine-Tuning as a Cost-Effective Strategy Fine-tuning small LLMs is a resource-efficient alternative to training large-scale models from scratch. Nvidia's H100 GPUs, accessible through the Launchpad platform, provide the necessary hardware acceleration to streamline this process. Using Nvidia's Nemo microservices, you can fine-tune a 1-billion-parameter model on datasets tailored for function-calling tasks, such as Salesforce's XLAM dataset. This approach ensures that the model is optimized for specific use cases while maintaining cost-effectiveness and scalability. The fine-tuning process not only reduces computational overhead but also shortens development timelines. By focusing on smaller models, you can achieve high performance without the need for extensive infrastructure investments. This makes fine-tuning an attractive option for organizations looking to deploy AI solutions quickly and efficiently. LoRA Fine-Tuning Tiny LLMs as Expert Agents Watch this video on YouTube. Advance your skills in fine-tuning by reading more of our detailed content. Nvidia Nemo Microservices: A Modular Framework Nvidia's Nemo microservices provide a modular and scalable framework for fine-tuning, hosting, and deploying LLMs. These tools simplify the entire workflow, from training to deployment, and include several key components: Customizer: Manages the fine-tuning process, making sure the model adapts effectively to the target tasks. Manages the fine-tuning process, making sure the model adapts effectively to the target tasks. Evaluator: Assesses the performance of fine-tuned models, validating improvements and making sure reliability. Assesses the performance of fine-tuned models, validating improvements and making sure reliability. Data Store & Entity Store: Organize datasets and register models for seamless integration and deployment. Organize datasets and register models for seamless integration and deployment. NIM Proxy: Hosts and routes requests to deployed models, making sure efficient communication. Hosts and routes requests to deployed models, making sure efficient communication. Guardrails: Implements safety measures to maintain robust performance in production environments. These microservices can be deployed using Helm charts and orchestrated with Kubernetes, allowing a scalable and efficient setup for managing LLM workflows. This modular approach allows you to customize and optimize each stage of the process, making sure that the final model meets the specific needs of your application. Preparing and Optimizing the Dataset A high-quality dataset is the cornerstone of successful fine-tuning. For function-calling tasks, the Salesforce XLAM dataset is a strong starting point. To optimize the dataset for training: Convert the dataset into an OpenAI-compatible format to ensure seamless integration with the model. Filter records to focus on single function calls, simplifying the training process and improving model accuracy. Split the data into training, validation, and test sets to enable effective evaluation of the model's performance. This structured approach ensures that the model is trained on relevant, high-quality data, enhancing its ability to handle real-world tasks. Proper dataset preparation is essential for achieving reliable and consistent results during both training and deployment. Training and Deployment Workflow The training process involves configuring key parameters, such as the learning rate, batch size, and the number of epochs. Tools like Weights & Biases can be used to monitor training progress in real time, providing insights into metrics such as validation loss and accuracy. These insights allow you to make adjustments during training, making sure optimal performance. Once training is complete, the fine-tuned model can be registered in the Entity Store, making it ready for deployment. Deployment involves hosting the model using Nvidia NIM containers, which ensure compatibility with OpenAI-style endpoints. This compatibility allows for seamless integration into existing workflows, allowing the model to be used in production environments with minimal adjustments. By using Kubernetes for orchestration, you can scale the deployment to meet varying demands. This ensures that the model remains responsive and reliable, even under high workloads. The combination of fine-tuning and scalable deployment makes it possible to create robust AI solutions tailored to specific use cases. Testing and Real-World Applications Testing the model's function-calling capabilities is a critical step before deployment. Using OpenAI-compatible APIs, you can evaluate the model's ability to execute tasks such as tool usage, parameter handling, and workflow automation. Successful test cases confirm the model's readiness for real-world applications, making sure it performs reliably in production environments. Fine-tuned LLMs offer several advantages for specialized tasks: Enhanced Functionality: Small models can perform complex tasks typically reserved for larger models, increasing their utility. Small models can perform complex tasks typically reserved for larger models, increasing their utility. Cost-Effectiveness: Fine-tuning reduces the resources required to develop domain-specific expert agents, making AI more accessible. Fine-tuning reduces the resources required to develop domain-specific expert agents, making AI more accessible. Scalability: The modular framework allows for easy scaling, making sure the model can handle varying workloads. These benefits make fine-tuned LLMs a practical choice for organizations looking to use AI for domain-specific applications. By focusing on function-calling capabilities, you can unlock new possibilities for automation and innovation, even with smaller models. Media Credit: James Briggs Filed Under: AI, Guides Latest Geeky Gadgets Deals Disclosure: Some of our articles include affiliate links. If you buy something through one of these links, Geeky Gadgets may earn an affiliate commission. Learn about our Disclosure Policy.
Yahoo
14 hours ago
- Business
- Yahoo
3 things Nvidia investors should look out for in its earnings call
The earnings event of the season is here. Nvidia (NVDA) will report its first quarter results for fiscal year 2026 after the market close today. Wall Street is expecting the AI chip darling to post another strong showing for earnings and revenue. Analysts forecast that earnings per share (EPS) will jump 46% year over year to $0.88. Revenue is expected to increase 66% to $43.3 billion. These are some sizzling growth numbers, though less sizzling than what Nvidia posted in 2024. "Nvidia is the one chip fueling the AI Revolution and the stock is not expensive. We see a $5 trillion market cap on the horizon for the Godfather of AI Jensen Huang and Nvidia," Wedbush tech analyst Dan Ives told Yahoo Finance. Nvidia's earnings report is about more than just the company, however. It's a bellwether for the AI boom, which is showing some signs of a slowdown. This time last year, Nvidia reported record quarterly revenue, up 262% year over year. Despite briefly becoming the world's most valuable company by market cap earlier this month, Nvidia shares are now slightly down year to date. The pullback reflects investor anxiety about the pace of AI adoption and whether Nvidia can maintain its explosive growth. On the Q1 earnings call last year, co-founder and CEO Jensen Huang said, "The next Industrial Revolution has begun." Well, now investors want to know where it's headed. Here are the top three things investors will be looking out for on Nvidia's market-moving earnings call. Nvidia's data center segment, driven by AI chips like the H100 and its Blackwell platform, accounts for the lion's share of its revenue. Analysts will be listening for signals about cloud providers and hyperscalers like Amazon (AMZN), Microsoft (MSFT), and Google (GOOG). Are they still placing large orders? Are there any hints of softening demand or delayed deployments? If so, that could rattle investor confidence in the AI infrastructure build-out and Nvidia's future. Read more: How does Nvidia make money? Nvidia has been benefiting from premium pricing on its high-performance chips. But with increasing competition from AMD (AMD) and custom silicon efforts from Big Tech companies, margins may come under more pressure. Investors will want to know if Nvidia is maintaining its pricing power or if it's beginning to feel cost compression, especially as it prepares for the high-volume rollout of Blackwell chips. Investors want to know whether demand is still accelerating or beginning to plateau. Any hint of slowing revenue or capital expenditures from key customers could prompt a reassessment of Nvidia's valuation. On the flip side, a bullish forecast — especially tied to Blackwell adoption or new enterprise use cases — could reignite momentum. Also listen for updates on supply chain capacity, chip-launch timelines, and geopolitical risks. Analysts may press Huang on the impact of tighter US export controls, which led to a $5.5 billion loss due to restricted chip sales to China. Huang has criticized the policy, warning it could accelerate China's domestic AI chip development. While China now accounts for just 5% of Nvidia's revenue, investors will be watching for any signs of further disruption. Don't forget: This isn't just Nvidia's moment — it's a checkpoint for the AI trade. Read more about Nvidia in the lead-up to earnings: Nvidia to report Q1 earnings as Middle East deals, export control reprieve boost stock How Nvidia "played a central role" in the $306 billion AI startup boom Why Nvidia's rise could signal bad news for climate goals Nvidia's bear case: Is the hype train running out of tracks? Big Tech's spending drove Nvidia's rise Brian Sozzi is Yahoo Finance's Executive Editor. Follow Sozzi on X @BrianSozzi, Instagram and on LinkedIn. Tips on stories? Email Brooke Sweeney is a senior producer at Yahoo Finance. Error in retrieving data Sign in to access your portfolio Error in retrieving data Error in retrieving data Error in retrieving data Error in retrieving data
Bloomberg
2 days ago
- Business
- Bloomberg
Silicon Data Creates First-of-Its-Kind Index for AI Chips
Silicon Data, a market-intelligence firm that focuses on the cost of graphics processing units, created what it says is the first daily index to track the specialized chips vital to powering artificial intelligence. The company's Silicon Data H100 Rental Index tracks the hourly cost of renting a GPU, according to founder and Chief Executive Officer Carmen Li.
