OpenAI debuts ‘Ultrafast’ for GPT-5.6 Sol — up to 14x faster: A Cerebras-powered mode pushing up to 750 output tokens per second, now a preview for a small group of users. TechCrunch · Cerebras · OpenAI
Anthropic in talks to buy world-model startup Decart for $6B: Decart builds ‘world models’ that simulate the physical world plus chip-efficiency software. A deal would be Anthropic’s largest acquisition yet, ahead of its anticipated IPO. Fortune · Reuters
Anthropic reportedly plans a $2T October IPO — the largest ever: The listing would eclipse SpaceX and become the biggest IPO on record, per sources. Fortune
Jeff Dean’s new AI startup in talks at a $10B valuation: The former Google chief scientist is reportedly raising at a $10B valuation for his new company. Business Insider
The frontier battle is shifting from ‘smarter’ to ‘faster’: Ultrafast shows model competition now spans latency and throughput, not just reasoning quality — and inference-specialist chips like Cerebras are becoming strategic assets. TechCrunch · Cerebras
Frontier labs are bringing world-model capability in-house: Anthropic’s Decart move signals a push to own physical-world simulation, with implications for agents, robotics, and real-time generation. Fortune
Text watermarks are fundamentally defeatable: Amid Anthropic’s Claude watermarking plans, a technical rebuttal argues text watermarks are always easy to strip, making watermark-only provenance strategies fragile. Sean Goedecke
Prototype low-latency interactive agents: At up to 750 tokens/sec, latency-bound products like real-time voice, live coding, and conversational editing become practical to prototype. TechCrunch · Cerebras
Design around the ‘recall bottleneck’: Google Research shows models know more than they can recall — teams can win by pairing models with retrieval and structured context rather than relying on parametric memory alone. Google Research