Daily AI Briefing — January 16, 2026

Cerebras announced a $10 billion deal with OpenAI, positioning its wafer-scale engine as an alternative to Nvidia's dominance in AI chips. The agreement provides Cerebras opportunity to prove performance at scale.

President Trump announced 25% tariffs on Nvidia and AMD AI chip sales to China, implementing a novel arrangement where the government takes a cut of sales after reversing export prohibitions on H200 chips in December. The scheme creates an unusual revenue-sharing model.

Sam Altman's new brain-computer interface startup Merge Labs emerged from stealth with $252 million in funding from OpenAI and others. The company aims to use ultrasound technology to read from and write to the brain.

Comprehensive safety evaluation of 7 frontier models: GPT-5.2, Gemini 3 Pro, Qwen3-VL, Doubao 1.8, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5 across language, vision-language, and image generation using unified protocol.

Large-scale empirical analysis of 100+ trillion tokens of real-world LLM usage through OpenRouter platform. Studies usage patterns across tasks, geographies, time, and the shift to reasoning models following o1's release.

Releases Molmo2, a state-of-the-art open-source VLM family with video understanding and point-driven grounding capabilities. Provides complete open weights and training data, addressing lack of transparency in the field.

Studies neural scaling laws using transformers trained on random walks on graphs, demonstrating scaling laws emerge even without power-law structure in data. Provides new theoretical perspective on scaling law origins.

Studies how AI discourse in pretraining corpora causally influences alignment outcomes. Training on misalignment discourse increases misaligned behavior while aligned discourse reduces it.

Anthropic announces 4th Economic Index report introducing 'economic primitives' - foundational metrics on AI usage including task complexity, education level, purpose, autonomy, and success rates.

John Carmack's #PaperADay series reviewing 'LOCAL FEATURE SWAPPING FOR GENERALIZATION IN REINFORCEMENT LEARNING'. Discusses CLOP (Channel-consistent local permutations) technique for reducing overfitting in RL by swapping neighboring positions in tensors while maintaining channel consistency. Notes data augmentation in latent space is more efficient than input space.

Fei-Fei Li shares new research on Interactive 3D world models as intuitive representations for learning robotics actions in dynamic environments.

Anthropic reports Claude achieves 50% success on 3.5-hour tasks via API, with high reliability on longer tasks - longer than METR benchmarks but in user-iterable contexts.

Anthropic shares results from AI for Science program - highlighting 3 research labs where Claude is enabling novel scientific insights and discoveries.