The Most Important AI News This Week Is Getting Overlooked

Microsoft’s MAI 2 announcement drew less attention than its potential importance for enterprise AI. Microsoft described the model as a 2.8-trillion-parameter mixture-of-experts system with 80 billion active parameters, trained on 16,000 GB200 GPUs for 38 days. The company said its pre-training efficiency was more than 12 times DeepSeek V4 Pro’s. Mustafa Suleiman told Microsoft staff that MAI 2’s base was stronger than the model powering GPT 5.6 Sole and used one quarter of OpenAI’s training compute, but that comparison depended on proprietary information and was not made public.

MAI 2 has not yet undergone reinforcement learning, so its performance after post-training remains unknown. Microsoft said it beat open models above one trillion parameters, including Kimmy K3, on its pre-training benchmarks. The episode framed the announcement as part of a broader enterprise push for cheaper, tailored models and greater control over the intelligence layer. If post-training brings MAI 2 close to the frontier, the number of US companies competing there could increase to six from two just a few months earlier.

Meta’s Muse Spark Open may have more direct business significance than the company’s higher-profile Muse Wonder and Muse Wonder Full releases. Meta says Spark Open has 816 billion total parameters, 42 billion active parameters, and a one-million-token context window, with a focus on coding and agent work, including long-horizon tasks and multi-step reasoning. Meta claimed open-state-of-the-art results on Jobbench and MCP Atlas, although its coding scores trailed Kimmy K3. Artificial Analysis gave Muse Open a score of 39, tying it with DeepSeq V4-1 Flash for 17th overall and seventh among open models. That was about seven points below the open-model state of the art and 13 points ahead of Nemotron and Inkling.

Nathan Lambert said Muse Open narrowed the gap between US and Chinese open-weight models by at least half, and predicted adoption would “go gangbusters” if its performance matched its scores. The full weights are available through major inference providers and for download. Apple’s MLX toolkit now supports running the model on Apple hardware, but the full version requires a terabyte of memory. The discussion estimated that running it on Mac Studios would require two high-spec M5 Ultra systems costing about $40,000 combined; that configuration had not yet shipped. A quantized version could fit on a Mac Mini, with trade-offs.

Anthropic has embedded Meter’s six-person team as its first third-party safety evaluator, giving the team employee-like access to internal systems and information Anthropic says it cannot publish. Meter can publish findings, subject to narrow, predefined redactions. Its independence was questioned because it shares major backers with Anthropic and has ties to the same effective-altruism ecosystem. Parv Mahajan emphasized that Meter is an assessor working under a voluntary agreement, not a formal auditor or substitute for oversight. Anthropic’s evaluator tiers also give SecureBio less access than Meter, and SecureBio needs prior approval to publish.

The White House AI task force held its first meeting Thursday and has 120 days to produce a safety report covering cybersecurity, critical infrastructure, biosecurity, model control, and accountability. It agreed to form a working group to implement the White House Accord, signed by frontier labs and NVIDIA, which calls for board-level safety oversight and partnerships with external auditors. Enforcement remains uncertain because the Accord is voluntary, although a White House official told Semafor that threatened export controls could be one tool.

Cognition is testing usage-based token limits that estimate the human-equivalent time spent on Devin sessions and raise engineers’ limits when they use budgets efficiently. Devin Coach is intended to catch common mistakes, such as vague requests or choosing an expensive model for routine work. Calvin Giroud argued that limits should reflect efficiency and value; Brian Chong countered that estimated human hours could reward complexity instead of value. Adam Cohen predicted that token allocation would become the most important skill of the next decade.