AI Daily Brief — 16 January 2025
Thursday’s anchor came from Hsinchu: TSMC walked the Street through a blockbuster Q4 — revenue up 38.8% YoY, 59% gross margin, record AI-accelerator demand — and set a 2025 capex budget at $38-42 billion, naming AI accelerators as the single largest growth driver for the next five years. Commerce quietly issued a supplemental Diffusion-Rule tweak. NVIDIA shipped NeMo Guardrails NIMs for agentic-AI safety. OpenAI’s ChatGPT Tasks (scheduled actions, agentic preview) rolled out.
Top stories
- TSMC Q4 blowout: revenue NT$868B (+38.8% YoY), net income +57%, 59% gross margin. Diluted EPS of NT$14.45 beat consensus; gross margin expanded 120bps QoQ on better utilization. The call also lifted 2025 capex guidance to $38-42 billion and the five-year revenue CAGR outlook to roughly 20% — with AI accelerators named as the single largest source of incremental growth. via Motley Fool transcript · via SEC 6-K
- BIS publishes supplemental Diffusion-Rule tweak. Three days after the IFR, Commerce modified license exception AIA — restricting it to items designed by “approved” IC designers. A small but meaningful tightening signalling continuing rule-by-correction activity in the administration’s final week. via Covington
- NVIDIA ships NeMo Guardrails NIMs for agentic-AI safety. New NIM (NVIDIA Inference Microservices) variants for guardrails address enterprise concerns about accuracy, security, and control as agent deployments scale — the second NVIDIA enterprise-product drop in two weeks. via NVIDIA
- OpenAI rolls out ChatGPT Tasks to Plus, Team, and Pro. Scheduled one-time and recurring actions (daily weather, news digests, reminders) accessed via “GPT-4o with scheduled tasks” in the model picker. Widely framed as OpenAI’s first agentic step toward Operator. via VentureBeat
- Hachette and Cengage move to join class action vs Google over Gemini training. Two more big book-publisher names broaden the publisher front against frontier labs. via Publishers Weekly
Who shipped
Beyond OpenAI‘s ChatGPT Tasks and NVIDIA‘s NeMo Guardrails, the labs were quiet on the product front. Anthropic, Google DeepMind, Meta, and xAI made no dated launches.
Open-source pulse
MiniMax-01 (Jan 15) was still the day’s loudest open-source signal as benchmark replications and 4M-token needle-in-a-haystack tests circulated on X and Hugging Face.
Research & papers
Two high-signal arXiv drops set the tone for the year’s reasoning literature. “Towards Large Reasoning Models” (2501.09686, Xu et al.) surveyed how RL is being applied to make LLMs master complex reasoning via trial-and-error search — framing the emerging “Large Reasoning Model” class. “Learnings from Scaling Visual Tokenizers” (2501.09755, Hansen-Estruch et al., Meta/FAIR) laid out scaling laws and architectural choices for visual tokenizers in reconstruction and generation. IJCAI 2025 abstract-submission deadline closed.
By the numbers
- +38.8% YoY — TSMC Q4 revenue growth
- $38-42B — TSMC 2025 capex budget
- ~20% — TSMC’s new five-year CAGR outlook (AI-driven)
- Most-mentioned company: TSMC
- Most-mentioned lab: OpenAI (ChatGPT Tasks)
Compiled by AI Feed’s editor from verified web sources for 16 January 2025.