Published record
Archive
750 Tokens per Second at the Frontier: How Cerebras Wafer-Scale Powers GPT-5.6 Sol Ultrafast
OpenAI and Cerebras have unveiled Ultrafast, a tier running GPT-5.6 Sol at up to 750 tokens per second. Here is how wafer-scale design sidesteps the GPU memory bottleneck — and what the numbers still need to prove.
Enterprises Miss Two-Thirds of Their AI Attack Surface as Sandbox Escapes Trigger Congressional Scrutiny
Security programs see only a third of their organization's real AI attack surface even as full-stack agentic adoption accelerates. A recent sandbox escape has turned that blind spot into an urgent congressional and regulatory issue.
Nemotron 3.5 Lightning and the Routing Layer
A purpose-built MoE executor that activates only 3 billion parameters per token, paired with NeMo Switchyard routing, is redrawing the cost structure of agentic workloads — though the speed claims remain vendor-reported for now.
Same Base, Post-Training Only: A Cyber Jump Z.ai Itself Flagged in GLM-5.3
The 743B GLM-5.3 reuses the identical GLM-5.2 base and credits every gain to scaled post-training — while the vendor itself warns of an emergent jump in cyber exploitation capability.
Qwen3.8 goes open: Alibaba ships 27B and 2.4T weights under a two-tier license
Both members of the Qwen3.8 generation hit open weights 11 days after the August 3 announcement — a 27B multimodal dense model under Apache 2.0 and a 2.4T MoE flagship under a custom license.
LG and NVIDIA turn the physical-AI bet concrete: a 2027 humanoid, AI factories and an AI-defined vehicle platform
An August 13 MOU at Nvidia's Santa Clara headquarters moves robotics, AI-factory and mobility ties from blueprints to execution — a GR00T-based bipedal humanoid targeted for Q1 2027, a Vera Rubin reference site, a planned 80MW Cheonan AI factory and a DRIVE Hyperion vehicle platform, all drawing on 'One LG' capabilities.
Reading Benchmark Contamination off the Residual Stream: Excess Separability
A nuisance-controlled residual-stream probe detects benchmark contamination from internal activations alone — no training corpus, no reference-model retraining — and is reported to survive RL post-training that erases the behavioral signals earlier detectors rely on. Researcher Florian Braun documents both what the protocol measures and where it refuses to claim.