Tavus Griffin: Real-Time Video AI Cuts Conversational Lag
Tavus launched Griffin, a two-way video AI model that replaces cascaded voice pipelines. Early tests show lower latency, but access remains limited.
Models
Model cards, releases, evaluations, and the limits a lab states in public.
Synthesia Sessions launches live interactive avatars
Synthesia launched Sessions, turning AI avatars into real-time video participants. Learn how it works, what it costs, and what questions remain open.
HBM4 on Nvidia Rubin takes 16% of die, SemiAnalysis says
HBM4 interface circuits take up 16% of Nvidia Rubin compute die, according to SemiAnalysis. The bottleneck pushes a custom memory shift on Feynman.
Lambda Boosts GPU Reliability in ClusterMAX 3.0
SemiAnalysis ClusterMAX 3.0 tests show Lambda auto-healing GPU errors in 15 minutes, though legacy CUDA 12.0 and Slurm limits create operational friction.
SemiAnalysis Exposes Real Gaps in AI GPU Clouds
SemiAnalysis fault tests reveal that green dashboards in AI GPU clouds hide crashed nodes, while recovery times range from 43 minutes to complete stalls.
GPT-6.1 Sol sets OpenAI record amid heavy load delays
Sam Altman says GPT-6.1 Sol is OpenAI's fastest-growing model, but acknowledges API slowdowns under heavy load and promises backend fixes.
Podonos Launches Onepin to Audit AI Voice Audio
Podonos released Onepin to audit synthetic voice tracks line by line, catch pronunciation errors, and orchestrate over 100 text-to-speech models.
Perplexity Decisions API arrives with open 27B model
Perplexity launched its Decisions API and open-sourced the 27B decider model under Apache 2.0, targeting automated classification over chat output.
Perplexity Financial Skills Launch for Amex Cards
Perplexity and American Express rolled out 10 financial skills for small businesses on Perplexity Computer, pulling card expenses via Plaid.
NVIDIA Pitches Fungible AI Infrastructure
NVIDIA argues fungible AI infrastructure protects data centers from obsolescence, letting GPUs switch jobs and stay viable longer than custom ASICs.