Meta Brings On Alexandr Wang — Zuckerberg Himself Is Set to “Sell” the AI

According to CNBC, Meta has brought on Scale AI founder Alexandr Wang as it develops new AI models. Notably, the company is building a structure in which Mark Zuckerberg himself will directly take on the role of pitching the model both inside and outside the company.

Meta has so far entered the AI race through massive investments, and the Wang hire appears aimed at strengthening its technical leadership. Scale AI is an industry leader in data labeling and AI training data, and Wang's expertise could accelerate Meta's model development. Zuckerberg taking on the sales role himself is a sign of just how much the company is betting its future on bringing AI products to market.

China Cuts 12,000 University Degree Programs — A Massive Pivot Toward the AI Era

According to the South China Morning Post, China is cutting 12,000 “obsolete” degree programs from its universities as it rushes to shift to an education system fit for the AI era.

This is an extraordinarily ambitious reform in scale. The figure of 12,000 programs signals that China is getting serious about moving talent development away from traditional academic disciplines and toward new fields such as AI and data science. China's posture of treating AI talent cultivation as a top national strategic priority is now taking concrete shape through a fundamental overhaul of its education system. As the capacity to supply talent becomes a key factor in the coming AI race, the impact of this policy is expected to unfold over the long term.

EAGLE Inference Acceleration Merged into llama.cpp — Faster Reasoning for Local LLMs

EAGLE, an inference acceleration technique, has been merged into llama.cpp, the inference engine widely used across the local LLM community. Reported in Reddit's LocalLLaMA community, this news is a significant update for developers running LLMs on local machines.

EAGLE substantially improves inference speed while preserving output quality, using speculative decoding — a technique that predicts and verifies “draft tokens” during inference. With its integration into llama.cpp, the most widely adopted local inference engine, a large number of users will benefit from the speedup with no additional configuration required.

Are MIT-Licensed Open Weights Losing Ground? — A Poll from z.ai

A poll run on X by Li Zixuan of z.ai (Zhipu AI) has been making the rounds, with results suggesting that MIT-licensed open-weight models are losing support. The topic was also picked up in Reddit's LocalLLaMA community.

More than 1,800 votes were cast, and the results suggest that models under more restrictive licenses (such as Gemma) tend to be more popular in the community than fully open, MIT-licensed models. It's a fascinating signal that the very concept of “open source AI” may be changing. As more companies brand their models “open weights” while imposing restrictions on commercial use, how to preserve truly open source models has become a key point of debate.

Controlling Super-Intelligent AI Is Mathematically Impossible — What the Computation Shows

According to research reported by Science Alert, computational analysis has led to the conclusion that fully controlling “super-intelligent AI” is impossible.

The study analyzes the AI control problem through the lens of computational complexity theory, suggesting that once an AI system possesses intelligence beyond a certain threshold, no algorithm can exist that verifies and controls its behavior in advance. The conclusion is based purely on theoretical computation, but it raises the possibility that “complete control” is an unrealistic goal in AI safety debates, and it could influence the direction of alignment (AI value tuning) research.

---