Weekly 3x3: Chips act is dead. Vance on AI opportunity. Hardware-aligned models.
My top reads in markets, tech, and AI research this week.
MARKET MOVES
Fed says stagflation can't be ruled out | February 20th 2025
The U.S. economy is facing the risk of so-called stagflation, when the labor market softens as inflation heats up, St. Louis Federal Reserve Bank President Alberto Musalem said on Thursday. Market Watch
The Chips Act may be dead. Another hurdle for semiconductor stocks | February 20th 2025
Moves expected by the Trump administration to gut a key government scientific agency could bring another challenge for semiconductor and chip-equipment companies — and their stocks. Market Watch
Tech's need for US manufacturing | February 20th 2025
Alger says tech companies need to "start the engine" again on domestic manufacturing in the US. Bloomberg
TECH TALK
JD Vance warns of overregulation | February 11th 2025
US Vice President JD Vance delivered a keynote speech during the final day of the Paris AI Summit where he warned global leaders and tech CEOs that excessive regulation would kill the rapidly growing AI industry. YouTube
Open AI designing chips in-house | February 10th 2025
OpenAI is designing its first in-house AI chip. The chip design is expected to be finalized in the coming months, with TSMC handling manufacturing. Reuters
Salesforce launches AI energy usage benchmark | February 10th 2025
To address AI's environmental impact, Salesforce, partnering with Hugging Face, Cohere, and Carnegie Mellon University, has launched the AI Energy Score. The tool aims to standardize energy efficiency reporting for AI models, providing a consistent way to measure and compare their footprint. AI Magazine
RESEARCH RADAR
Native sparse attention: Hardware-aligned and natively trainable sparse attention | February 16th 2025
NSA, a natively trainable sparse attention mechanism, achieves efficient long-context modeling by combining a dynamic hierarchical sparse strategy with hardware optimizations, delivering substantial speedups and comparable or superior performance to full attention models. ArXiv
How do LLMs acquire new knowledge? A knowledge circuits perspective on continual pre-training | February 16th 2025
By studying how LLMs build their internal "knowledge networks" while learning, researchers discovered that new information is easier to grasp if it connects to existing knowledge, that this learning process happens in distinct phases (building then refining the network), and that LLMs learn from the ground up, which can help us design more effective training strategies. ArXiv
CRANE: Reasoning with constrained LLM generation | February 13th 2025
Constraining LLM outputs for syntactic and semantic correctness can hurt their reasoning ability, but CRANE, a reasoning-augmented constrained decoding algorithm, addresses this by carefully expanding the output grammar, preserving reasoning capabilities while ensuring correct outputs, and achieving significant performance gains on challenging symbolic reasoning tasks. ArXiv