Key Takeaways
- OpenAI and AMD report up to 14% PPA improvement from model-assisted chip design, with some results in recent tape-outs and Jalapeño silicon.
- Anthropic priced its IPO at $83 per share, with Nasdaq trading expected the next day.
- SpaceX caught a Starship upper stage at the launch tower for the first time and deployed 26 Starlink V3 satellites.
- Google released Genie 4 and Gemini 4 Neon, with early evaluations placing Neon near the top of Vals Index at lower cost.
- Perplexity launched Teammates and Multiagent on mobile and reported three new APIs and nine-figure annualized API revenue.
- NVIDIA released Nemotron-Labs-Sigma open weights and described Factory-based post-training inside an air-gapped network.
1. AI Models and Chip Design
- OpenAI and AMD said models identified up to 14% PPA improvement in AMD flagship GPU IP in under five months, on designs AMD engineers had already optimized for years. The same method was applied to AMD's first custom inference chip, Jalapeño, with gains verified on real silicon; OpenAI also said it is working with Synopsys to bring model capabilities into EDA design and verification tools. 1 2
- Google released Genie 4, an interactive world model combining real-time physics, reasoning and audio, available to eligible Google AI Ultra subscribers in Project Genie, and launched Gemini 4 Neon for Pro and Ultra users. Forwarded evaluations put Neon at 70.9% on Vals Index, second and 0.4 points behind Claude Fable 5.2, at less than half the cost per task, while another forwarded user reported Neon beating Fable 5.2 in that night's tests. 1 2 3 4 5
- NVIDIA released Nemotron-Labs-Sigma open model weights in 14 checkpoints at 3B/8B scale, using block-parallel token generation rather than one-token-at-a-time generation, and said the approach could speed exploration of multiple solutions in math and coding. NVIDIA also said it used Factory, which it described as the only commercially supported harness deployed in its environment, to post-train Nemotron entirely inside its isolated network. 1 2
2. Business and Infrastructure
- Anthropic priced its IPO at $83 per share, with trading expected to begin on Nasdaq the next day; it said the registration statement tied to its Class A common stock was declared effective by the SEC on November 18, 2026, and that the offering is being made only by prospectus. 1
- SpaceX caught a Starship upper stage at the launch tower for the first time, which Elon Musk called the last major piece of the puzzle; he said Ship 42 needs only minor repairs before reflight but will retire for public display. The mission also deployed 26 Starlink V3 satellites, all in operation, adding 26 Tbps downlink and 4.2 Tbps uplink capacity, which forwarded material compared with about 10 Falcon 9 launches. 1 2 3
- Perplexity launched Teammates and Multiagent on iOS and Android, with Teammates supporting voice control for Max subscribers, and said Multiagent usage doubled in a day. It also introduced three APIs—Monitor, Orchestrator and Context—and said API revenue grew 4.3x year over year to a nine-figure annualized run rate. 1 2 3
3. Safety, Evaluation and Research
- OpenAI described four layers of protection across its model lifecycle—model alignment, safety and sandbox isolation, blocking safeguards, and monitoring and response—and said advancing to the next stage requires higher safety and alignment standards as model capability rises, while calling this a baseline operating discipline rather than a final solution to alignment. OpenAI also said researchers' AI use is rising exponentially and called for labs to track and publicly report recursive self-improvement metrics. 1 2
- A forwarded evaluation analysis said the harness is a key factor in model performance: native harnesses usually raised both solve rates and cost versus a Bash-only environment, but results varied—Codex models 6.1 Terra and 6 Luna gained about 10 percentage points, newer 6.2 changed little, Grok and DeepSeek benefited clearly, Muse was flat, and Fable 5.2 was significantly cheaper under its native harness. 1
- Yann LeCun said superintelligence will not arrive next year or in two years, arguing current AI is strong on symbolic tasks such as math and code but understands the physical world less well than a cat, one reason home robots remain uncommon. He also advised students to choose learning how to learn over narrowly useful technical skills, citing a World Economic Forum projection that 39% of workers' core skills will change by 2030. 1 2
- Greg Brockman said OpenAI's models found a pre-authentication RCE in NGINX 1.31.8, the current mainline version, called NGINX2Shell, in work done with close cooperation from discovery through responsible disclosure; he also said a separate effort independently found a vulnerability that could form a generic pre-authentication RCE in Ruby on Rails. 1 2
