Nvidia launches a mini AI supercomputer that could change computing forever, Microsoft shows that small AI models can outsmart large models, and HubSpot’s co-founder launches a new platform to build and share AI task agents.
Welcome to Hashtag Trending! I’m your host, Jim Love.
First up – Nvidia’s New Mini AI Supercomputer Could Change Computing Forever.
Nvidia has unveiled Project DIGITS, a mini AI supercomputer that delivers one petaflop of AI performance at a fraction of the cost of traditional systems. Priced at $3,000, the device is powered by Nvidia’s GB10 system-on-chip, offering 1,000 teraflops of performance at FP4 precision.
To put that in perspective, when the Nvidia DGX-1 was launched in 2016, it was the first AI supercomputer purpose-built for deep learning, designed for tasks like image recognition and natural language processing.
Back then, it cost $129,000 and took up the space of a small server rack. Today, Project DIGITS delivers roughly the same AI performance for less than 3% of the cost and would fit on your desk, making high-performance AI computing accessible to a far wider audience.
CEO Jensen Huang introduced a new concept that some are calling “Jensen’s Law,” which suggests that it takes roughly eight years to reduce the price per AI performance metric by a factor of 25.
Project DIGITS exemplifies this trend, bringing AI computing previously reserved for tech giants like Google and Microsoft to individual researchers and businesses.
Nvidia’s strategy with DIGITS is to create a competitive offering by making AI computing more accessible. This approach mirrors how companies like Microsoft and Apple built loyal ecosystems. By making AI more available, Nvidia ensures developers remain reliant on its CUDA platform.
The mini supercomputer’s standout feature is its FP4 processing, optimized for AI inference and delivering five times the performance of FP8. AMD, Nvidia’s closest competitor, hasn’t yet introduced similar technology but we can expect that to happen soon, another reason why Jensen’s company is moving so quickly into this area.
Nvidia’s broader strategy? Disrupt itself before competitors do, keeping its lead in the AI space while redefining what’s possible on desktops worldwide.
Next – Microsoft’s rStar-Math Shows Small AI Models Can Outsmart Larger Ones.
Microsoft has published a paper describing a new technique called rStar-Math, showing that smaller language models can outperform OpenAI’s larger models in math reasoning tasks.
Using a method called Monte Carlo Tree Search, a combination of Chain of Thought Reasoning and the multiple scenarios used in risk management. Leveraging this, these smaller models simulate potential outcomes and make decisions step-by-step, but in the process learn and improve both their accuracy and transparency.
Key innovations include a Code-Augmented Chain of Thought, which allows the model to explain its reasoning using Python code, and a Process Reward Model that improves the accuracy of step-by-step solutions.
Microsoft claims these smaller models achieve state-of-the-art performance without the massive computational resources required by larger models.
Although they sound like big numbers, these small models may have 70 billion parameters but that compares with the 7 trillion parameters of the large models. This can have a huge impact on data, power and processing requirements. So where the concern with larger models is the practical limits of scaling, the smaller models are offering potential new path forward.
And it turns out that these small models are showing the potential of being not just faster and more efficient, but more accurate and smarter than their larger counterparts.
For example, rStar-Math boosted the accuracy of a smaller model from 58.8% to 90% in solving math problems. On the American Invitational Mathematics Examination, the models ranked in the top 20% of high school competitors.
Microsoft’s success challenges the idea that bigger is always better. It could also signal a shift toward more efficient AI models that deliver high performance without the environmental and financial costs of massive systems.
Hugging Face, a popular AI community, plans to release rStar-Math on GitHub, but the code is still under review. One of the researchers behind the project said, “The repository remains private for now. Please stay tuned!”
HubSpot Co-Founder Launches New AI Platform to Build and Share Task Agents.
Dharmesh Shah, co-founder and CTO of HubSpot, has launched Agent.ai, a professional network where users can build, customize, and share AI agents designed to automate tasks. The platform offers a marketplace of agents that can help with tasks like company research, competitor analysis, and website copy editing. Users can also create their own agents using the platform’s Agent Builder tool, currently in beta.
There is also a directory of human “agents” who can assist in providing services for AI implementation.
Shah believes AI agents will become as common as web apps. In a blog post, he predicted, “In a year or two, it’ll be as common to build AI agents to solve customer problems as it is to build web apps today.”
Agent.ai aims to make task automation more accessible to business professionals. Examples include a Company Research Agent that gathers data on competitors and a Web Page Copy Editor to improve readability – and there are many more. By making AI tools approachable, Shah hopes to replicate the success he had with HubSpot, where simplified customer management became a key feature.
The launch comes as businesses increasingly turn to AI for automation. Shah’s vision? A community-driven platform where users share and improve agents, helping businesses keep pace with today’s rapid changes.
That’s our show for today. You can find links in our show notes at technewsday.com or .ca – take your pick. You can reach me with comments, questions, or tips at editorial@technewsday.ca. I’m your host, Jim Love. Have a marvellous Monday!
