Software

Nvidia's Smaller Model Outperforms Giant Rival in Supply-Chain Task

Nvidia and Palantir demonstrated that a compact 30-billion-parameter model trained on real operational decisions can outperform a vastly larger competitor at optimizing parts distribution.

3 min read

Nvidia and Palantir think small

Nvidia has shown that a more compact model trained on actual supply-chain decisions can surpass a significantly larger counterpart when determining optimal allocation of scarce components.

According to reporting by Paul Sawers, Nvidia's 30-billion-parameter Nemotron 3.5 Lightning achieved 86.7% accuracy on the parts-allocation task, outperforming the 550-billion-parameter Nemotron 3 Ultra, which scored 55.5%. The Lightning model was trained using decisions from Nvidia's own operations team, with Palantir's platform serving as the bridge between those decisions and the underlying supply-chain information. The two companies first disclosed this initiative in June and formally announced it on Thursday.

Both organizations are encouraging other enterprises to adopt this methodology, leveraging proprietary data while retaining ownership of their custom models. Nvidia's findings suggest that selecting a smaller model and fine-tuning it for a particular business use case may deliver superior performance. However, the improvements did not extend uniformly across all tasks. Which supply-chain challenges saw the most improvement through fine-tuning, and which ones continued to resist resolution?

The AI-speed SOC: A peer workshop for security leaders

Alert counts are rising, and AI is enabling attackers to operate faster than most security operations centers can keep pace. On September 15, a limited group of security executives will convene in a private setting to establish what security operations should resemble once AI becomes fundamental to its design.

  • A private working session governed by Chatham House Rules, distinct from a standard webinar format
  • Intimate roundtables with counterparts managing comparable challenges, limited to 25 attendees
  • Restricted to CISOs, SOC Directors, and incident response leaders managing teams of 10 or more

TOP OF THE STACK

Shopify rebuilt its mobile app in 12 weeks after abandoning React Native

Shopify is moving away from React Native toward native mobile applications, deploying AI agents to reconstruct its Shop app over a 12-week period and reimagining how code repositories function with agent-driven development.

In 2020, Shopify made a strategic choice that gained traction throughout the development community by writing mobile code a single time using React Native rather than implementing each feature separately in Swift and Kotlin. On Thursday, the company announced it is reversing that decision.

WHAT ELSE IS NEW?

TDS launched ShipAI: functioning implementations, not promotional presentations

Towards Data Science, our affiliated publication, unveiled ShipAI, a curated collection of genuine AI implementations: screen-recorded demonstrations, editorial review, and preserved as a reference collection rather than a continuous stream.

  • Examine genuine, operational AI implementations rather than written descriptions alone
  • Each submission undergoes editorial assessment prior to inclusion in the collection
  • Highlighted implementations are promoted monthly as essential viewing material

For those developing AI applications and seeking visibility for their work, project submissions are welcome.

FLOW STATE

When state governments supported data center expansion before the extent of power consumption became apparent, and shifted to more cautious positions afterward, what constitutes the significance?

Securing Kubernetes deployments demands organization-wide policies, reinforced infrastructure, and ongoing verification as traditional safeguards and machine learning systems broaden organizational exposure.

Discover how refined approaches to trace sampling can decrease observability expenses, mitigate information saturation, and enable site reliability engineers to diagnose infrastructure problems with greater efficiency.

How did we do today?

A cautiously optimistic summary of developments in technology, commerce, and energy.

Perspectives from leading experts in data science, machine learning, and artificial intelligence.

Collaborative, role-based and competency-focused guides to support your professional development in technology.

The New Stack is a media platform for the people who build and manage software the world relies on.

Source: The New Stack

Source: The New Stack · Reporting supplemented by The Silicon Ledger staff.