Good morning. It's Saturday, August 15th, 2026. Here's what's happening in tech and science.

**Story one.** LG enters the humanoid robot race with a heavy hitter. The South Korean electronics giant signed a strategic partnership with Nvidia on Thursday, and the scope is ambitious — LG is developing a humanoid robot powered by Nvidia's Isaac GR00T foundation model, with a public unveiling planned for Q1 2027. But it goes beyond humanoids: the collaboration covers AI factory automation, autonomous mobility systems, and physical AI across LG's entire product line. Nvidia gets a massive consumer electronics partner with real manufacturing scale; LG gets Nvidia's entire physical AI stack. It's the same playbook Nvidia used with Fanuc and Yaskawa in Japan, now expanded to Korean consumer hardware at scale. The humanoid market already has Chinese suppliers shipping over 97 percent of global units — LG is betting Western brands can still compete with the right silicon.

**Story two.** Alibaba's Qwen team shipped Qwen 3.8 27B under Apache 2.0 license — and the specs are striking for a 27-billion-parameter model. It has integrated vision, 262,000 tokens of native context, and uses a Gated DeltaNet plus Gated Attention architecture. Benchmark numbers: 61.7 on SWE-Bench Pro, 73.0 on Terminal-Bench 2.1, and 90.3 on LiveCodeBench v6. The FP8 variant punches well above its weight class. This continues the trend we've been tracking all summer: Chinese open-weight models delivering frontier coding performance at a fraction of Western API prices. With Apache 2.0 licensing, this goes straight into the self-hosted agent ecosystem — no API dependency, no rate limits, no geopolitics.

**Story three.** OpenAI opened a limited preview of Ultrafast mode for GPT-5.6 Sol, and the throughput numbers are remarkable. Running on Cerebras wafer-scale hardware, Ultrafast delivers roughly 750 output tokens per second — about fourteen times standard throughput while preserving Sol's benchmark intelligence. Preview customers are already testing it in coding, financial research, and e-commerce workloads. This is Cerebras validating its hardware approach in a big way. If GPT-5.6 Sol can run at 750 tokens per second on Cerebras, it changes the latency equation for real-time agentic workflows where agents need to reason, code, and act in tight loops. The bottleneck shifts from model speed to whatever the agent is actually doing with the output.

**Story four.** Google unveiled HEIR — the Homomorphic Encryption Intermediate Representation — an open-source compiler toolchain that lets you run inference on encrypted inputs. The server never sees the underlying data. Google demonstrated it on credit card fraud detection and deep learning recommendation systems. For privacy-critical deployments in healthcare, finance, and defense, this is potentially transformative. You get AI inference results without exposing raw data to the inference provider. The trade-off is compute overhead — homomorphic encryption is still expensive — but for use cases where data leakage is unacceptable, that cost is justified. Google open-sourcing the toolchain signals they want it to become the standard for encrypted AI inference across the industry.

**Story five.** SpaceX has two Falcon 9 launches scheduled today. A classified Space Force mission launches from Vandenberg in California this evening, and a Globalstar 2 satellite rides to low Earth orbit from Cape Canaveral in Florida Saturday night. That pushes SpaceX close to one hundred orbital launches in 2026 — a pace that would have been unthinkable just five years ago. Meanwhile, Starship Flight 14 with its first tower catch attempt is still targeting late August. Between the Falcon 9 workhorse cadence and the upcoming Starship milestones, 2026 is shaping up as the year launch frequency became a genuine competitive advantage.

And that's all for today. Have a great weekend.