The Most Important AI News This Week Is Getting Overlooked

Anthropic has begun embedding outside safety evaluators inside the company after CEO Dario Amodei previously promised the move as part of efforts to make frontier development more accountable. The first group is a six-person team from METR, which Anthropic says has “employee-like access,” including company computers, office space, and internal tools. Anthropic is also working with SecureBio and Accenture and is discussing possible cooperation with other organizations, including Hugging Face. Anthropic’s argument is that it cannot publicly disclose all safety information because some details could help malicious actors or expose commercially sensitive training information. Instead, evaluators receive deep access and publish evidence and conclusions about risks, narrowing the gap between what an outside reviewer knows and what the company knows. The arrangement has drawn criticism over METR’s independence: METR and Anthropic share major backers and roots in the effective-altruism community, even though Anthropic says it does not fund METR. Critics compare this to a bank selecting its preferred auditors. Publication control is another issue. Anthropic says evaluators cannot be vetoed from publishing, but the company can require redactions for narrowly defined categories such as customer data, legally privileged information, and model-architecture details. METR is in the top access tier; SecureBio begins in the lowest tier reviewing previously disclosed chemical and biological risks, with a possible promotion after its first report; and Accenture is expected to work on red teaming, guardrail testing, and alignment assessments. METR’s Parv Mahajan called the arrangement a prototype, not formal oversight, because there are no agreed safety standards. The host concludes that genuinely independent auditing will probably require a broader system, potentially involving major accounting firms, although qualified personnel are scarce. Elon Musk said SpaceX would probably adopt a similar approach.

Cognition is introducing a system that dynamically adjusts how much engineers can use its coding agent, Devon. The company’s VP of Customer Engineering, Calvin Jiroud, argues that static limits are wrong: a user’s allowance should depend on the efficiency of their usage and the value they receive. Cognition says an AI can examine a Devon session and estimate how many human work hours would have been required to achieve the same result. When an engineer reaches a token limit, the system grades how efficiently the budget was used. A satisfactory score automatically raises the user’s limit. Jiroud compared the mechanism to a credit card: users receive higher limits as they demonstrate that they are responsible and repay on time. Cognition says this approach is preferable to the earlier “token maxing” era of unlimited budgets and the later “token austerity” era of organizational caps. Its customers include large public companies that must explain software efficiency to shareholders, so Cognition views unlimited token spending as unsustainable. The company is also launching Devin Coach, which gives engineers feedback on common mistakes, such as making ambiguous requests or choosing an expensive model for a routine task. Some developers reacted angrily. Jesse, a software engineer, said software engineering was becoming a factory role; Amisha called the system “very Soviet”; and Brian Chong argued that measuring equivalent human hours might reward unnecessarily complicated work rather than valuable work. Others, including founder Adam Cohen, said the specific method may be imperfect but that token allocation will become an important skill. The broader conclusion is that organizations will increasingly manage AI usage through efficiency metrics rather than simply encouraging maximum consumption.

The first meeting of the White House AI Task Force established several responsibilities. The President’s Council of Advisors on Science and Technology must produce an AI safety report within 120 days, covering AI risk, cybersecurity and critical infrastructure, biosecurity and dual-use applications, and model control and accountability. Another working group will implement the White House Accord signed by frontier labs and Nvidia, which commits companies to safety controls with board-level oversight and partnerships with external auditors. Defense Undersecretary Emil Michael will lead efforts to create reporting protocols for security incidents. FTC Chair Andrew Ferguson will coordinate a law-enforcement and consumer-protection committee across the Justice Department and FTC. Office of Personnel Management Director Scott Kapoor will work with public and private organizations on hardening critical systems and will coordinate a Labor Department study of AI’s effects on the workforce and economy. Enforcement is unresolved because the accord is voluntary, though a White House official told Semafor that export controls could serve as leverage; the administration has already used that tool against Anthropic. The episode then turns to speculation that Tesla and SpaceX could merge before Tesla’s December 9 shareholder meeting. Barclays analyst Dan Levy said recent management signals suggest benefits from combining the companies, especially a joint AI-chip plant, although neither company has hired bankers, formed a negotiating committee, announced a deal, or filed with the SEC. Levy suggested an all-stock exchange of approximately 0.7 Tesla shares for each SpaceX share, which would leave Musk with about 37% of the combined company and roughly 56% voting power. Prediction markets put the chance of a merger before year-end at only 24%, despite Musk saying that close collaboration could lead to such a move.

YouTuber Max Blade and fellow streamers created Kong Royale, which Blade calls the world’s first AAA “vibe-coded” battle royale. The game features a large island, vehicles, and a giant gorilla as the player character. Blade had explored the idea during the early days of vibe coding but said the available models were not capable enough at the time. Using Opus 5.5 and several Claude Max accounts, he produced the game after a few weeks of work. Blade called the process “hive coding.” During his livestream, human developers and audience members proposed ideas while multiple AI agents participated in the chat. Overnight, the agents reviewed the discussion, selected promising changes, and shipped them into a build ready the next morning. The result was effectively a crowdsourced development process in which the streamer’s community supplied ideas and AI agents implemented and iterated on them. The host places Kong Royale about a year and a half after Peter Levels’ plane game helped popularize the idea of vibe-coding entire games. He says the example shows what a solo developer can now accomplish with persistence and an AI subscription. Riley Brown argued that social creation is the natural next step for games: streamers and communities will build games together and then play them. Alex Hansen imagined “bros’ nights” in which friends bring pizza, Mountain Dew, and a whiteboard, then use an advanced model to develop or modify a game overnight. The episode presents this as evidence that improving models may make community-built games increasingly professional.

The host’s central argument is that Microsoft’s and Meta’s latest model announcements matter more for business adoption than for benchmark spectacle. Neither release produced the kind of result that would dominate AI leaderboards, but both address practical questions about cost, customization, data control, and vendor dependence. Enterprises are discovering that routing every task through the most capable proprietary model is not cost-efficient, especially as agentic software expands AI usage and token bills. Companies are consequently adopting a more “model-omnivorous” approach: they want to switch among models depending on whether a task requires high capability, low cost, privacy, or specialized behavior. Microsoft’s earlier Frontier Tuning initiative was designed for this shift by letting customers fine-tune models on top of Microsoft’s base models. CEO Satya Nadella has argued that companies should own the intelligence layer around models rather than handing that layer to a single provider. That layer may include model routers, open models, and locally hosted systems. The host says enterprises with technical resources will increasingly build architectures that mix frontier models for difficult orchestration and reasoning with cheaper, tailored models for repeatable workflows. Open models offer another route to sovereignty and cost control, but Microsoft’s existing position in business software may make a Microsoft-based approach more comfortable for many companies than building infrastructure around an on-premises model from China. The two releases therefore expand the practical choices available to organizations and could increase total AI workloads by making it cheaper and more flexible to deploy models across the stack.

Microsoft AI CEO Mustafa Suleyman announced MAI Base 2, a new base model with 2.8 trillion total parameters and 80 billion active parameters in a mixture-of-experts architecture. The model is comparable in size to Kimi K3, slightly larger than some models mentioned in the episode, and was trained from scratch without distillation from other models. It has not yet undergone reinforcement learning or other post-training, but Microsoft says its pre-training results are unusually strong. In an internal town hall, Suleyman told employees that MAI Base 2 was stronger at the pre-training stage than the model powering GPT-5.6, while using one-quarter of OpenAI’s training compute. Microsoft did not publish that proprietary comparison, but it said MAI Base 2 outperformed every open model above one trillion parameters on its public pre-training benchmarks, including Kimi K3, sometimes by a wide margin. The run used 16,000 GB200 Blackwell GPUs for 38 days, and Microsoft claimed more than 12 times DeepSeek V4 Pro’s pre-training efficiency. The release matters because Microsoft has had an awkward dependence on OpenAI. It invested $13 billion in OpenAI, benefited enormously from ChatGPT’s early momentum, and then experienced uncertainty during Sam Altman’s 2023 removal and return. Microsoft hired Suleyman in March 2024 to lead a new internal AI division, but its first MAI-1 models were merely adequate commodity systems. Those models were mainly important because they supported Frontier Tuning, through which customers could customize Microsoft’s base models. MAI Base 2 could make that strategy substantially more compelling: if post-training brings it near the frontier, Microsoft becomes an independent high-end competitor while offering enterprises a way to customize and retain more control over their AI stack.

Meta’s Muse Spark Open is the other major release the host believes businesses should watch. It has 816 billion total parameters, 42 billion active parameters, and a one-million-token context window. Meta AI CEO Alexander Wang said it is designed for coding, long-horizon agentic tasks, difficult software engineering, and multi-step reasoning. On benchmarks, it sits between leading U.S. open models such as Inky and Nemotron 3 Ultra and Chinese open models such as GLM 5.3 and Kimi K3. Its coding scores trail Kimi K3, but Meta claims open state-of-the-art results on agentic tests including JobBench and MCP Atlas. Artificial Analysis gave Muse Spark Open a score of 39, placing it 17th overall and seventh among open models, approximately seven points below the open-model state of the art but 13 points above Nemotron and Inky. Unlike some competing systems, its full weights are available for download and the model is offered through major inference providers. Researcher Nathan Lambert called its return to near-frontier U.S. open models a major moment, saying it cuts the gap with Chinese models by at least half and could see rapid adoption. Meta’s move is also a reversal from the disappointment of Llama 4, whose planned two-trillion-parameter Behemoth version was never released. The company subsequently rebuilt its AI division and appeared less committed to open weights. Muse Spark Open now returns Meta to the top of the U.S. open-model ranking. Apple has updated its MLX toolkit so the model runs efficiently on Apple hardware, and Wang joked that “mommy and daddy” were playing nicely again after years of Meta–Apple conflict. Running the full model requires roughly one terabyte of memory—possibly two high-end M5 Ultra Mac Studios costing around $40,000—but quantized versions can run on cheaper systems, including a Mac mini, with trade-offs. The host concludes that Microsoft and Meta have given enterprises more credible options for private, customized, and mixed-model AI deployments.