Local AI, Voice Models & Building Reliable Agents

What this episode covers

• xAI Grok Voice 3.0 tops the speech-to-speech index at 86.9 while holding price at $0.08/min.
• NVIDIA Nemotron 3.5 Bolt fits a 74B model into 24GB with open weights, training data and recipes.
• StepFun opens Step 5 preview weights: 600B total, 27B active, up to 1M context.
• DOJ charges an OpenAI engineer over a concealed boardroom recording device; charged, not convicted.
• 406 scientists including Hinton and Bengio call for default airgapping of cyber-capable models.
• Anthropic posts a $16.8B quarter as its implied valuation swings below $2T ahead of a November IPO window.
• The FT reports NVIDIA in talks to acquire or deepen its investment in Reflection AI, the lab behind Beam.
• Meta ships Muse Spark Open — 816B parameters tuned for two Mac Studios, with gated repos and limited commercial terms.
• The local stack question moves from whether models can run to whether they should.

Published record