started · updated
OpenAI and Cognition advance AI hardware and software capabilities
OpenAI has accelerated the development of its first proprietary AI chip, named Jalapeño, moving from concept to silicon in less than 20 months. Partnering with Broadcom, which handled the physical design, the project utilized a small team of fewer than 100 people. OpenAI leveraged its own AI models to speed up the coding phase, a process described by hardware VP Richard Ho as giving engineers ‘superpowers’. The chip reportedly achieves 13.4 petaflops in 4-bit and provides up to 3.6 times less latency than Nvidia’s GB300.
In the software sector, Cognition has introduced two new AI-driven development tools: SWE-2 and Fusion. SWE-2 is a specialized programming model built using reinforcement training on the 2.8 trillion parameter Kimi K3 model. The system is designed to optimize computational efficiency by penalizing unnecessary execution steps, aiming to reduce costs. According to company data, SWE-2 achieved a 50.0% score on the FrontierCode 1.1 Main benchmark while maintaining a 64% lower cost compared to previous iterations.
Entities
Broadcom · Cognition · Nvidia · OpenAI · Sam Altman