Google's December 17 release of Gemini 3 Flash presents frontier-level artificial intelligence (AI) at lower latency and cost for high-frequency work. It is a fitting close to 2025: another substantial model advance whose real significance will be determined by the systems able to use it.
Throughout this year, the center of gravity has moved from answers toward action.
Anthropic's December 2 study of its own workforce describes engineers and researchers completing more work, operating beyond familiar specialties, iterating faster, and addressing tasks that have previously gone undone. It draws on a survey, qualitative interviews, and internal usage data from an organization of unusually early artificial intelligence (AI) adopters.
The productivity story is real. So is a quieter organizational question: what happens to the learning that traditionally occurs inside the work now delegated?
Google DeepMind's November 11 research shows that visual artificial intelligence (AI) models can learn to organize images more like people do. The work uses human “odd-one-out” judgments to reshape the conceptual relationships inside vision models and reports gains in human alignment, few-shot learning, and robustness to distribution shift.
That is meaningful progress. It is also a useful occasion to distinguish human-like perception from human judgment.
Anthropic's September 29 release of Claude Sonnet 4.5 highlights stronger coding, computer use, and sustained work on complex tasks. Alongside the model come checkpoints for Claude Code, a memory tool and context editing for longer agent runs, and an agent software development kit (SDK).
The pairing is instructive. Longer autonomy arrives with better ways to see, constrain, and reverse the work.
Anthropic's September 9 announcement says Claude can create and edit spreadsheets, documents, presentations, and Portable Document Format (PDF) files inside a private computing environment. The release marks another step in artificial intelligence (AI) moving from suggesting content to producing the artifacts through which organizations make decisions.
Once the file arrives polished, the management problem becomes easier to miss.
Google DeepMind and its research partners report on September 4 that an artificial intelligence (AI) controller has undergone testing at the Laser Interferometer Gravitational-Wave Observatory (LIGO) in Livingston, Louisiana. Their Deep Loop Shaping method reduces control noise in a difficult mirror-control loop by a reported factor of 30 to 100.
The achievement is technically specific, and that is precisely why it offers a useful lesson for AI in critical systems.
Students do not wait for an institutional operating model. They find a useful tool and fold it into the work.
Anthropic's April 8 Education Report analyzes roughly one million anonymized conversations associated with higher education. The study identifies four interaction patterns spanning direct and collaborative problem solving and output creation. It also finds that students frequently delegate analytical and creative work—not only routine recall.
Military planning is a coordination technology. It turns incomplete information, command intent, operational constraints, staff expertise, and adversary uncertainty into courses of action that can be compared and executed.
The Defense Innovation Unit's (DIU) March 5 announcement introduces Thunderforge, an initiative to bring artificial intelligence (AI) agents, modeling, simulation, and large language models into operational and theater-level planning. The stated ambition is faster synthesis, course-of-action development, and AI-enabled wargaming.