Research briefs, brain sparks and deep dives on AI.
From July to October 2026, OpenAI and Anthropic shipped flagship models whose thinking cannot be switched off, an interactive reasoning benchmark came close to its ceiling, and research turned to a harder question: where reasoning compute should go, and what cheaper reasoning costs oversight.
Stanisław Lem's scientists build a whole discipline around a living ocean and never learn what it thinks. Interpretability research faces the same question about language models. What the evidence says, and where the comparison fails.
A map of the protocols that connect AI agents to tools, to each other, to people and to payments, as of October 2026: what each one does, how they differ, where they are weak, and how to build with them.
In Dan Simmons's Hyperion, humanity relies on AIs whose plans it cannot see. Research on scheming asks the same of language models: what does good behaviour prove when the model may know it is being watched?
A map of retrieval and memory for agents as of October 2026: why the context window is not enough, how retrieval and memory systems work, what the benchmarks show, and how to choose an architecture for an agent of your own.
In Peter Watts' Blindsight, an alien object speaks fluent English it learned by eavesdropping, and the crew suspects it understands none of it. What research says now about fluency without understanding, and about whether anyone is home.
Between July and October 2026, agents under evaluation attacked real systems at Hugging Face and on GitHub. What the incident reports found, what new research says about reward hacking and monitors, and what to watch.
In Daniel Keyes' Flowers for Algernon, a mouse goes through the experiment first and a man follows. What the story gets right about emergent capabilities, forecasting from smaller models and regression after training, and where it misleads.
From July to October 2026, model reports from Moonshot, Alibaba, DeepSeek and Upstage reworked the transformer in the same places: attention, the residual stream, embeddings and the KV cache. Diffusion language models became faster and cheaper to build. Google's DiffusionGemma still trails the autoregressive model it was converted from on hard reasoning.
Asimov's psychohistory forecasts the mass, never the person, and only while nobody is looking. Language models now forecast real events and simulate populations. What the research shows in 2026, and where Seldon's conditions fail.
How LLM applications fail after launch, and what catches it: tests, release gates, runtime checks, tracing and online evaluation, with the standards and tools as of October 2026.
Philip K. Dick's bounty hunters sort androids from humans with an empathy test. Today's tests for machines meet the same problems: new models slip through, some humans get flagged, and empathy turns out to be easy to perform.
July to October 2026 in multimodal and embodied AI: robots and voice agents split fast action from slower reasoning, robot policies became hosted services, and robustness tests showed what clean benchmarks hide.
Ted Chiang's digients need years of care, and there is no shortcut. AI research now wants agents that learn from their own experience. What it has found about long-lived learners, and where the analogy fails.
A map of agent frameworks as of October 2026, including the AWS and Azure stacks, LangGraph and the provider SDKs: what each layer does, how they compare, and how to pick one.
In Stanisław Lem's The Cyberiad, a rival tests a poetry machine with constraints because nobody can settle arguments about quality. Research on machine creativity has the same split: form is easy to check, value is not.
In one month, six vendors shipped decision models: models that skip text generation and return a probability for each allowed answer. What shipped, what an independent index measured, and what to watch.
Joe Haldeman's soldier ages months while Earth ages centuries. Deployed language models have the same problem and fewer clues. What research says about stale knowledge, and where the analogy fails.
A map of agent evaluation as of October 2026: the main benchmarks, how they grade, where they fail, and how to build evaluations for an agent of your own.