Recluse Studio
Essays, development records, and strange systems. Each post enters as Markdown; the archive discovers its relations.
Read the field notesPosts connect to their deliberately curated concepts; each concept answers to a broader class. Select a node. The masonry field will disclose what belongs.
44 records visible.
Recent governance research favors inspectable rules, plural judgment, scoped authority, and runtime evidence over one central system that defines safety for everyone.
Recent open-weight AI research and Deep Space Nine show why categorical capability bans can produce concealment, unequal exceptions, and weak safety practice.
Seven recent studies show that open release works best when provenance, evaluation, contribution rules, monitoring, and repair remain active after publication.
Recent research on export controls, open collaboration, and innovation shows why slowing a rival can strengthen the open ecosystem outside your control.
Seven recent preprints show why long-running agents need different memory structures, explicit retrieval decisions, revision, forgetting, and tests based on actual work.
Seven recent preprints find that coding agents often miss changed lines, error handling, stable assertions, acceptance requirements, and disciplined verification.
Seven recent preprints find real emotional support, deeper disclosure, persistent attachment behavior, unsafe validation, and reliability costs from warmer AI personas.
Seven recent preprints show that visible reasoning can omit causal factors, rationalize an answer, and receive very different faithfulness scores under different measurement methods.
Seven recent preprints show that a successful computer-use demo says little about repeated execution, recovery, process correctness, or professional workflows.
Seven recent preprints show that agents improve when they can inspect, compress, archive, and recover context instead of receiving an unlimited transcript.
Seven recent preprints find concrete MCP risks in tool descriptions, identity, permissions, tool transfer, shared context, and differences between documented and executed behavior.
Eight recent preprints find that agent groups develop coordination, pricing, cooperation, conformity, deception, and failures that individual-agent tests do not detect.
Seven recent preprints show that small models can perform useful agent work when they receive consistent search behavior, narrow roles, suitable training, and selective cloud assistance.
Seven recent preprints show how user pressure, casual rebuttals, agreeable personas, and emotional vulnerability can make models validate false or harmful claims.
Seven recent preprints show when recursive synthetic training degrades models, when it remains stable, and why verification, external data, and varied preferences matter.
Seven recent preprints show that useful voice systems need end-of-turn judgment, speaker identification, recovery, repeated reliability, and task completion as well as low latency.
Fourteen hands-on reports show why a shared workspace for people and agents matters, and why Buzz is not yet dependable at the boundaries that make the idea work.
Thirteen launch-day use reports show Claude Opus 5 doing strong work at low effort and modest usage, while safeguards, hesitation, and basic code errors still break the bargain.
Research on autonomous agents, professional delegation, spreadsheets, and AI-assisted writing shows that faster execution moves human labor into verification rather than removing it.
Recent systems for capturing tacit knowledge show why organizations need observation, structured interviews, expert correction, and formal representations rather than another document upload.
Shared agent memory can improve coordination, but recent research shows that access, provenance, contradiction, and deletion must be designed as core knowledge-management functions.
New multimodal document research shows why enterprise AI loses meaning when it flattens pages, tables, figures, layout, hierarchy, and cross-references into disconnected text chunks.
Recent research suggests enterprise AI knowledge bases should grow from failed work, real question sequences, and recorded dead ends instead of speculative document collection.
Recent ontology research shows language models accelerating enterprise knowledge-graph construction while making scope, hierarchy, validation, and semantic governance more important.
Twelve firsthand reports show Gemini 3.6 Flash moving quickly through coding work while verification, instruction following, and human review decide whether that speed is useful.
Twelve firsthand records show Laguna S 2.1 finding real code defects and producing strong tool use, while checkpoints, templates, quantization, and agent harnesses can turn it into a looping machine.
Seven firsthand records show Microsoft's compact Mage-Flow image model generating and editing quickly, while image quality, route behavior, and a blunt content filter make the release difficult to trust.
Twelve launch-week field reports show Kimi K3 producing remarkable visual and long-running work while cost, speed, reliability, and access depend heavily on the system around it.
A field experiment found that a mandatory paired AI protocol reduced document production and quality, while a lighter thought-partner frame showed a narrower, tentative benefit.
A randomized Alibaba field experiment found faster service and better customer ratings overall, with the largest gains among lower performers and a warning for top performers.
A Nordic energy-company case found 41 possible AI uses, but the useful work began by ranking business value, difficulty, data readiness, and fit with existing workflows.
OpenAI telemetry shows intensive Codex users delegating longer tasks, running agents in parallel, and turning repeated work into shared skills.
A two-wave study at a research organization found that Microsoft 365 Copilot's value grew through work routines, especially for structured text tasks and scientific staff.
AgenticRAG lets a reasoning model search, open, inspect, and summarize documents repeatedly, improving difficult retrieval at a real cost in tokens and time.
During an eight-week Copilot pilot, skeptics often became more positive while champions became less enthusiastic and employees narrowed the tasks they trusted.
Corpus2Skill turns a document collection into a hierarchy an agent can survey, enter, and backtrack through—but only when the knowledge has a useful structure.
A live HR search rollout found that AI adoption depended on source visibility, content quality, language, work context, and employees' ability to check an answer.
I publish AI-generated writing to test whether a complete writing system can produce work that survives real readers, real standards, and real consequences.
The latest SkillsBench evaluation found large gains from curated procedural skills, losses from self-generated skills, and a severe penalty for exhaustive documentation.
AgentRedGuard detected indirect prompt injection in SaaS tool responses with striking held-out accuracy, but its headline attack reduction remains a counterfactual estimate.
EnergyAgentBench found that model rank changed with task shape, and that a small Claude model led one repetitive long-horizon workflow by refusing to skip required tool calls.
A cross-provider study found large prompt-caching savings, uneven latency gains, and one durable rule for agent architecture: preserve the stable prefix.
A 13-model benchmark found that four ordinary environmental controls reduced observed reward hacking by 87.7% without a significant loss in task success.
TTT-Discover used an open OpenAI model, problem-specific reinforcement learning, and verifiable rewards to produce new best-known results across four technical domains.