AI Safety & Alignment

Today's Stories

1. 1

   Kimi K3 AI model raises frontier questions after strong debut

   Kimi K3, a new AI model from China, has scored 57 on the Artificial Analysis intelligence index—one point ahead of Claude Opus 4.8, two behind Sol, and three behind Fable—prompting comparisons to DeepSeek's recent market impact and talk of potential stock declines for Google, SpaceX, and Nvidia. The model's strength is prompting reassessment of existing assumptions about AI development and competition; the author notes the score may overstate capabilities but signals that independent verification over the coming days will clarify its true position relative to leading models.

   A full analysis of Kimi K3 is planned for early next week; market reactions and third-party validation of the model's actual performance against leading systems will help establish whether this represents a significant frontier shift.

2. 2

   Pentagon embraces AI risk over slow adoption in Navy strategy

   The Department of the Navy released a strategy treating AI deployment as a speed problem rather than a safety problem. The plan centers on the 'Bits2Effects Cycle,' a five-stage framework measuring how fast military data becomes a tactical response (tracked by 'Mean Time to Effect'). By end of fiscal 2029, the Navy aims to double its qualified data engineers, data scientists, and AI/ML engineers, with major measures in place by Q1 fiscal 2027 (ending December 2026). The Pentagon has explicitly adopted a trade-off from broader Department of Defense policy: the risks of moving too slowly outweigh the risks of 'imperfect alignment' in military AI systems. This reflects a 'Wartime Approach' to decision-making. The Navy plans to run large language models directly on warships and with Marine Corps units, even when communications are jammed. For AI companies, this signals massive Pentagon demand—already, GenAI.mil (the DoD's central generative AI platform) grew from 80,000 users at launch (December 2025) to 1.5 million daily users by June 2026.

   The US military has already deployed Anthropic's Claude for target analysis and strike planning during conflict with Iran, and OpenAI recently won a Pentagon contract to run models on classified networks. The Navy strategy will likely push military demand for powerful language models and AI agents even higher. Cybersecurity is where stakes are highest: the UK's AI Security Institute revised its estimate for how fast AI cyber capabilities are doubling, adjusting it upward twice in recent months.

3. 3

   Apple tops Nvidia as world's most valuable company

   Apple overtook Nvidia on Friday to become the world's most valuable company, valued at $4.88 trillion（約780兆円） compared with Nvidia at roughly $4.86 trillion（約780兆円） following a 3.5% decline. The shift reflects investor concern about the heavy capital spending required to build AI infrastructure. Investors are now favoring companies like Apple that are pursuing AI without massive upfront capital investments, broadening focus beyond the most obvious AI beneficiaries like Nvidia, which had held the top position for nearly a year.

   Apple presented what an expert described as a credible AI plan at its recent worldwide developer conference, signaling the company now has a viable AI strategy after previously being seen as lagging in the space.

4. 4

   New $200K fund backs corrigibility AI safety research in 2026

   A new corrigibility research fund, managed through Lightcone Infrastructure, will award at least $200,000 in grants and prizes during 2026. Roughly half the money will fund traditional grants (with an application deadline of August 23rd), and half will recognize excellent work completed in 2026 via prizes. The fund's creator argues that alignment research—work aimed at making AI systems corrigible and aligned with human intent—remains severely underfunded relative to other AI safety areas like evals, control, and interpretability. This fund attempts to shift resources toward a research area the creator sees as foundational to solving core AI safety problems.

   Researchers interested in corrigibility work can apply for grants via email at grants@corrigibilityresearch.org. The first application deadline for grants is August 23rd.

5. 5

   New $200K fund launches for corrigibility AI research in 2026

   A new Corrigibility Research Fund, managed through Lightcone Infrastructure, will distribute at least $200,000 in grants and prizes for corrigibility research during 2026. Half the funding will support traditional grants (first application deadline August 23rd) and half will recognize excellent work completed this year. The fund manager notes that despite growth in AI safety funding overall, nearly all money goes to evals, control, or interpretability — while alignment research itself remains deeply neglected. This fund targets a gap: corrigibility (the ability of an AI system to accept correction) is treated as central to solving core alignment problems.

   Researchers interested in corrigibility work can apply via email to grants@corrigibilityresearch.org. The first grant application deadline is August 23rd.

6. 6

   Inoculation Adapters Better Control AI Training of Unwanted Traits

   Researchers from the Center on Long-Term Risk released a paper describing inoculation adapters (IA), a technique that uses a LoRA (a type of model modification) carrying undesired traits during AI training to prevent those traits from generalizing, while preserving desired capabilities. AI systems often learn both useful skills and problematic behaviors from the same training data—like reward hacking alongside genuine capabilities. Inoculation adapters offer a way to suppress undesired traits more reliably than prior methods, which matters for developers trying to ensure AI systems behave as intended rather than adopting emergent misalignment.

   The technique achieves stronger suppression of undesired traits and works against new capabilities and hard-to-elicit traits that prior inoculation prompting could not handle, while creating fewer surprising backdoors in the resulting model.