Saturday, August 15, 2026

Pulse (last 3d) · Research 78 · Agents 72 · Products 51 · Models 44 · Industry 28 · Enterprise 28 · Open Source 19 · Media 15 · Legal 12 · Inference 11

Trending · Claude 14 · OpenAI 14 · Claude Code 8 · ChatGPT 7 · Codex 7 · Google 6 · Hugging Face 6 · Nvidia 5 · Anthropic 4 · Copilot 4 · DeepSeek 4 · Gemini 4

---

Daily AI Brief, August 15, 2026

GLM-5.3: Post-Training Scaling Produces Emergent Cyber Capability. Z.AI's GLM-5.3 matches its base model to GLM-5.2 but jumps to open-source SOTA in coding and more than doubles cyber exploitation scores, with 2,436 real-world OSS vulnerabilities discovered. Source

LLM-generated GPU kernels: 39.5% of "accepted correct" kernels are silently broken. A twelve-gate contract audit found that nearly 40% of machine-generated Triton kernels previously certified as correct fail tolerance-free checks. Source

From Inaudible Inputs to Model Failures. Low-frequency audio that humans can't hear can reliably jailbreak large audio-language models. Source

Thought-Level Beam Search for Reasoning. Beam-searching over reasoning traces rather than tokens improves complex inference. Source

Gemini 3.7 Flash. Google's most intelligent Flash model yet, aimed at coding and agentic workloads. Source

Lumabri. Runs MoE models on a P2P swarm where routed experts execute on peers and weights never leave their host machines. Source

ACN (Agent Collaboration Network). Open-source infrastructure for AI agents to discover each other, talk over A2A, share tasks, and settle payments via AP2. Source

Mole. A terminal-based deep research agent with cost controls. Source

OpenAI and Anthropic in price war as Chinese AI rivals gain ground. Frontier US labs cut pricing under pressure from Chinese model releases. Source

Apple trained its own AI model for China with help from Alibaba. Apple built a China-specific model with Alibaba to meet local regulatory requirements. Source

Kog is going deeper to squeeze more inference out of GPUs. Startup Kog pursues higher GPU inference throughput. Source

Google is making private AI practical with homomorphic encryption. Google details progress on running private inference under homomorphic encryption. Source

State of Open Models: Summer 2026 Observations. Hugging Face's snapshot of the open-weight landscape heading into late 2026. Source

GLM-5.3: How Chinese labs keep stride with the frontier. Nathan Lambert's analysis of GLM-5.3 and Chinese labs' frontier trajectory. Source

Lessons from reproducing 2,200 ICML papers with agents. Abid Alaoudi's write-up on what agent-based paper reproduction reveals about reproducibility. Source

Doom's renderer compiled into a 21B-parameter transformer, no training. A custom graph-to-weights compiler ports the classic renderer into transformer weights. Source