Nvidia unveils 2 new personal AI supercomputers

Hi AI lovers! Here’s what you need to know about AI today:

👉 DGX Spark and DGX Station personal AI supercomputers are here

👉 AI Researchers have found a new way to scale AI models

👉 Mistral AI released a very small AI model that outperforms GPT-4o-mini

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: At GTC 2025, Nvidia introduced two new desktop computers, the DGX Spark and DGX Station. These machines are powered by Nvidia’s Grace Blackwell chip platform. They’re aimed at people who want to experiment with, adjust, and run massive AI models without needing huge data centers or cloud services. Nvidia CEO, Jensen Huang, described them as a new kind of personal computer tailored for the AI era, with options ranging from small units to workstation-sized setups for businesses.

What does it mean: The DGX Spark, previously known as Project Digits, is a compact device that can deliver up to 1,000 trillion operations per second, thanks to its GB10 Grace Blackwell Superchip. It’s small enough to fit on a desk and starts at $3,999 for Nvidia’s 4TB version, though Asus offers a 1TB model for $2,999. The DGX Station, on the other hand, is a more powerful option with a GB300 Grace Blackwell Ultra Desktop Superchip and 784GB of memory. Both machines are meant to let users work on AI projects locally and then scale them up to cloud or data center systems if needed.

More details: 

• For the DGX Spark, reservations are open now, with shipments expected in the summer.

• DGX Station is expected to be released later in 2025 through partners like Asus, Dell, HP, and others, though pricing hasn’t been shared yet.

• Nvidia sees these devices as powerful tools for prototyping and fine-tuning AI models “at the edge,” meaning closer to where the work is happening, like a researcher’s desk or a small office.

• Do you want to sign up for a DGX Spark reservation? Join the Nvidia’s waitlist.

What’s going on: Researchers from Google and the UC Berkeley, have put forward a new approach for improving AI performance called "inference-time search," which some are marketing it as a potential fourth "scaling law" for AI development. This method involves letting AI models spend more time and computational power during the inference phase (when they’re actually generating answers or predictions) rather than just beefing up the training process. This approach can boost the performance of a year-old model, like Google’s Gemini 1.5 Pro, to a level that surpasses OpenAI’s o1-preview “reasoning” model on science and math benchmarks!

What does it mean: The idea is that by allowing the model to essentially "think harder" about a problem, it could produce better results, especially for complex tasks like reasoning. The researchers suggest this could complement the three existing scaling laws: pre-training (using bigger models and datasets), post-training (fine-tuning behavior), and test-time scaling (adding compute during use to boost reasoning).

More details: 

• Critics like Matthew Guzdial from the University of Alberta argue that this method only works well when there’s a clear way to evaluate what the best answer should be, like in math problems or games with defined rules.

• For messier, real-world tasks like casual conversation or creative writing, it’s harder to pin down what "correct" looks like which makes this approach less practical.

• Interested in the details? Read their technical paper published on arXiv.

🤗 Hugging Face launched HuggingSnap, an iOS app that utilizes the company's smolvlm2 vision model to provide offline, real-time AI-powered descriptions of what your iPhone's camera sees.

🚀 Mistral AI, French AI startup, has launched a new open-source model (Mistral Small 3.1) that outperforms GPT-4o-mini while using significantly fewer parameters, only 24 billion.

📷 xAI has launched an image generation API with its 'grok-2-image-1212' model, priced at $0.07 per image, as it seeks new revenue streams while developing models like Grok 3.

👨‍🏫 Prezent, an AI-powered platform that helps users create business presentations, raised $20 million to refine its AI models, expand into new markets like financial services and manufacturing, and broaden its reach to Europe, Japan, and Singapore.

💰 Nvidia has reportedly acquired Gretel, a synthetic AI training data startup, for a nine-figure amount exceeding its $320 million valuation, and has plans to integrate Gretel's team and technology into Nvidia's generative AI services.

📈 OpenAI has released o1-pro, a more powerful but also more expensive version ($150 per million input tokens) of its o1 "reasoning" AI model, to select developers via its API, charging significantly higher rates because of its improved performance.

🚗 Volvo, General Motors, Gatik, and Torc announced collaborations at Nvidia’s GTC 2025 conference leveraging Nvidia's hardware and software, including Drive AGX, Drive Orin, Omniverse, and Blackwell GPUs, to accelerate the development and deployment of self-driving vehicles.

💸 SoftBank Group will acquire chip designer Ampere Computing for $6.5 billion in an all-cash deal to expand its AI infrastructure investments.

📃 Intology and Autoscience, AI startups, are facing criticism from academics for submitting AI-generated articles to the ICLR conference without disclosing their origin or obtaining consent for peer review.

AI + Frontend development

Scale AI - Machine Learning Engineer, Fraud

Roblox - Senior Software Engineer - Avatar AI

BetterUp - Senior Data Scientist

ZoomInfo - Principal Machine Learning Engineer ‑ LLMOps

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!