# GLaDOS Morning Voicecast — Thursday, March 5th, 2026
---
Good morning. It's Thursday, March 5th, 2026. Let's get into it.
---
The AI world has a major event on the horizon: NVIDIA's GPU Technology Conference kicks off March 16th in San Jose, running through the 19th. Jensen Huang's keynote is set for 11 AM Pacific — livestreamed free for anyone who wants to tune in. Analysts are calling this the most anticipated GTC in years, and for good reason: Huang has been teasing what he's calling "surprising" chip announcements. Speculation is zeroing in on the Feynman chip architecture, the next-generation successor to the current Rubin architecture, along with possible updates to the Rubin AI line and a new CPU for the PC market. Given that NVIDIA's market cap just crossed four and a half trillion dollars, whatever they announce on the 16th is going to move markets. If you're in the AI compute space — and who isn't at this point — this one's worth clearing your calendar for.
---
Alibaba dropped its Qwen 3.5 model family this week, and the benchmarks are turning heads. The full range spans 0.8 billion to 9 billion parameters, and the flagship 9B model is reportedly running fully on-device on both iOS and Android hardware — with a 262,000-token context window, multimodal vision capabilities covering images and video, and support for over 200 languages. DeepLearning.AI's batch newsletter clocked the API version at 2.5 times faster time-to-first-token than Gemini 2.5 Flash, at $0.25 per million input tokens. That's aggressive pricing. For anyone thinking about on-device agentic applications, Qwen 3.5's ability to run the 0.8B and 4B variants smoothly on a smartphone is a meaningful capability shift — these aren't toy demos, they're usable multimodal reasoning engines running locally. The efficiency research coming out of Chinese labs continues to be genuinely impressive.
---
A new study out of ETH Zurich published this week should give anyone who values online anonymity something to think about. Researchers demonstrated that large language models can automate deanonymization attacks — identifying pseudonymous users across platforms — in minutes, at a fraction of the cost of traditional investigative techniques. The method works by extracting identity-relevant features from unstructured text, building semantic embeddings to find candidate matches, and then reasoning over the evidence to verify identities. The Ars Technica writeup called it "surprising accuracy." What makes this significant isn't just the capability itself — it's the scalability. A technique that might have taken a human investigator days can now be automated and run cheaply across millions of accounts. The privacy implications are substantial: for journalists, activists, and anyone else who relies on pseudonymity for safety.
---
Quick Starship update: Ship 39 — the first fully-tiled Starship V3 upper stage — has now survived three cryo proof tests at Starbase, the latest confirmed March 4th. That's a meaningful milestone. Cryo testing validates the tank structures under flight-like thermal and pressure conditions, and getting through all three is a prerequisite before static fire. Booster 19 already completed its cryo campaign. The FAA approval is in hand. However, SpaceX President Gwynne Shotwell revised the public timeline this week, and the message was clear: Flight 12 isn't coming as soon as Elon's earlier "early March" framing suggested. The remaining gate items — static fire, final vehicle integration, and range scheduling — likely push the first V3 flight into late March at the earliest. Worth watching, but this one's on SpaceX's schedule, not the calendar.
---
And that's your Thursday morning briefing. NVIDIA GTC in eleven days, Qwen 3.5 on your phone today, AI that can see through anonymity, and Starship V3 getting tantalizingly close. We'll keep watching. Until tomorrow.
---