Meta’s AI Model Helps Solve Five Unanswered Math Problems

Morning, {{first name | folks}}! AI is showing up in some unexpected places today. Meta is using models to tackle open math problems, the U.S. is forming a new AI task force, and smaller models are getting easier to run locally. Meanwhile, data center expansion is facing new pressure from local communities. Let’s get into it.

Meta AI says researchers used its Muse Spark model while working on six mathematical papers, five of which contain answers to previously open questions. The problems spanned probability, differential equations, group theory, optimization, and number theory. In one case, the model wrote a search program that found a counterexample to a mathematical conjecture; in another, it helped connect ideas across number theory and p-adic string theory.

The researchers directed the work and checked the resulting proofs and arguments. Meta says the model was used through its standard Meta AI interface rather than a custom research system. The results add to a growing list of cases where AI systems are being used to investigate problems where researchers don't already know the answer.

Donald Trump has created a new “Super Intelligence Force” and given it 120 days to assess the risks and opportunities around AI. National intelligence director Jay Clayton will lead the group, with FTC Chair Andrew Ferguson, Pentagon CTO Emil Michael, and OPM Director Scott Kupor serving as vice chairs.

The group will also draw on government agencies and outside companies as it prepares its report. Trump’s order calls for plans to respond to AI-related threats while avoiding regulations that could slow development or competition. Clayton has also warned that the U.S. risks falling behind China if it loses the race to build more capable AI systems.

ByteShape has released new GGUF versions of Qwen3.8-27B that shrink the model down to 8.8GB at the smallest setting, with other versions ranging from 9.9GB to 13.1GB. The release uses ByteShape’s ShapeLearn system, which picks different data types for individual parts of the model instead of compressing everything the same way.

That makes the 27B model much easier to run on consumer hardware. The files can be loaded through llama.cpp, Ollama, LM Studio, and other local tools, with the release tested across coding, math, general knowledge and agent tasks against the full-precision model.

Amazon says it has stopped using nondisclosure agreements with local government officials involved in its data center projects after growing criticism over how much of these developments happen behind closed doors. NDAs have sometimes prevented officials from sharing basic details about projects, including who is behind them and how much water and electricity they could use.

Amazon is also putting $1 billion into communities hosting its data centers over the next five years, funding community colleges, job training, and energy-efficiency projects. The move comes as more than 100 US communities consider moratoriums on new data centers, putting pressure on the companies racing to build the infrastructure needed for AI.

OpenBMB has released a 323M-parameter model designed to speed up its 2B-parameter MiniCPM5. The smaller model works as a “draft” model, suggesting several tokens at a time while the larger model checks them instead of making the larger model generate everything itself.

The approach is called speculative decoding, and OpenBMB says the draft model can get an average of 5.5 tokens accepted per verification step. The company is releasing it alongside MiniCPM5-2B, a model built for local use with a 131K-token context window. The interesting part is the setup: a much smaller model doing the early work so the larger model has less to do.