Nvidia releases an open speech model that transcribes 40 languages
THE SO WHAT
Nemotron 3.5 Transcribe plus a 463,062‑hour Granary v2 dataset is Nvidia pushing open, multilingual speech infra deeper into the stack. If you’re paying per‑minute for transcription or running brittle ASR, it’s time to benchmark this — especially for multi‑speaker, non‑English environments.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIAnthropic just launched tools to show companies what their AI spending is buying
Frontier labs are being forced to sell not just tokens and tiers, but marginal ROI per workload. If you own a seven-figure AI budget, expect your board to ask for Anthropic-style transparency from every vendor—and build your own cost-to-outcome dashboards before they do.
Applied AIDOJ Inquires About Nvidia’s Hugging Face Deal, Reuters Says
Regulators probing a $13B model-platform deal is a clear sign that AI distribution and community hubs are now seen as potential choke points. If your roadmap leans heavily on a single model marketplace or ecosystem, map concentration risk and alternative channels this quarter.
Applied AIAmazon warns engineers just using more AI doesn't make them more productive, says they should avoid tokenmaxxing
When Amazon tells engineers to stop “tokenmaxxing,” it’s codifying a shift from AI usage metrics to outcome metrics. If your internal AI program is bragging about tokens, prompts, or seats instead of cycle time and defect rates, expect a hard correction from finance and ops.
Applied AIGitHub launches Copilot Fusion as agents start 40% of pull requests
If agents are already initiating 40% of pull requests and Copilot Fusion is routing tasks across models, the unit of work in software is shifting from “developer writes code” to “system orchestrates changes.” Engineering leaders should be redesigning review, testing, and ownership around machine-initiated diffs, not just human commits.