Microsoft MAI-Transcribe-2 leads benchmark
MAI-Transcribe-2 leads real-time speech rankings, but Microsoft claims about speed and cost drew immediate pushback from ElevenLabs.
Models
Model cards, releases, evaluations, and the limits a lab states in public.
Microsoft launches MAI-Transcribe-2-Streaming audio
Microsoft releases MAI-Transcribe-2-Streaming and MAI-Voice-2.1, pairing fast speech recognition with multilingual voice synthesis for voice agents.
Grok in Group Chats Arrives on X for Premium+
X brings Grok into group chats for US Premium+ subscribers, adding multimodal AI and summaries while suspending end-to-end encryption.
Grok 4.7 rolls out to Grok web and mobile apps
SpaceXAI began rolling out Grok 4.7 as the base model in Grok web and mobile apps across Fast, Expert, Build, and Heavy modes. Here is what changed.
Claude artifacts: 50% less session-limit use
From 1 to 15 Oct 2026, Claude artifacts use 50% less of the five-hour session limit. The weekly cap does not change.
Boston Dynamics gives Atlas a 13-DOF four-finger hand
On 1 Oct 2026 Boston Dynamics announced GR3: four-finger Atlas hands, 13 degrees of freedom, sim-to-real training. No pinky. Studio demo, not factory.
AMD shows another Helios after OpenAI lab visit
AMD posted another Helios lab photo on Oct. 1, 2026, tagged OpenAI, and described a tour. The post does not say OpenAI runs that rack.
Harvard physicist on LLM impedance mismatch
Anthropic published Matthew Schwartz's Oct. 1, 2026 guest post on LLM impedance mismatch: models do code and math well, not human lab work.
Character.AI reports 2x AMD inference throughput
DigitalOcean and AMD say Character.AI doubled inference throughput on Instinct GPUs and cut token cost 50% versus a generic setup. No independent audit.
OpenAI DevDay still felt consumer, analyst says
Ray Wang told The Information that OpenAI DevDay still felt like a consumer event. Confirmed quotes, company claims, and what stays opinion.