Chips to China, Synthetic Research, and AI's Grand Mystery

Chips, China, and controversy. The Department of Commerce is giving Nvidia the green light to export its H200 AI chips to China, but only to vetted customers and with a 25% cut for Uncle Sam. These chips, a step up from their H20 counterparts, have to be at least 18 months old before heading overseas. While Nvidia's thrilled, applauding the move as a win for American jobs, not everyone's on board. Senators Pete Ricketts and Chris Coons are pushing a bill to put the brakes on these exports, citing national security risks. More here.

Aaru's headline valuation magic. Aaru, the AI whiz-kid shaking up customer research, just snagged a Series A round, led by Redpoint Ventures, with a flashy $1B 'headline' valuation. But here's the twist: the real blended valuation is under a billion, thanks to some creative multi-tiered deals. This play gives them bragging rights while keeping investors happy with sweeter terms. Founded in March 2024, Aaru uses AI to simulate human behavior, bypassing old-school surveys for predictive wizardry. More here.

AI's grand mystery show. The NeurIPS conference in San Diego was the place to be for AI aficionados, drawing a record 26,000 attendees this year. This AI extravaganza highlighted the growing mystery around how advanced AI systems actually work. While neural networks have evolved from niche academic projects to global powerhouses, the field of interpretability—understanding these systems—is still in its infancy. Google and OpenAI are taking diverging paths, with Google opting for practical methods and OpenAI doubling down on deep interpretability. More here.

Interpretability gets a prize. Martian launched a $1 million prize to boost interpretability efforts. The field is still asking basic questions like 'What does it mean to have an interpretable AI system?' As one researcher put it: 'We're in the phase of asking: What are electrons? Do electrons exist?'

Evaluation tools lag behind. Stanford's Trustworthy AI Lab noted that current evaluation methods were 'built for a different time.' We don't have the measurement tools to measure more complicated concepts and bigger questions about models' general behavior.