Skip to content

INDEPENDENT RESEARCH / PRACTICE NOTES

Ideas in motion, not ideas behind glass.

This is my working notebook on AI engineering, strategy, knowledge infrastructure, organizational transformation, and the human systems that determine whether innovation becomes real capability.

The notes range from emerging research questions to practical operating models. Some will become papers, tools, talks, or products. Others are here because thinking improves when it is made visible.

Evidence before theater Systems over slogans Useful, accountable AI

Subscribe via RSS Follow Field Notes in any feed reader

Latest writing

Human-like perception is not human judgment

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.

Enterprise AI at 350,000 seats is an organizational redesign

Anthropic and Cognizant's November 4 announcement makes Claude available to as many as 350,000 Cognizant employees. The plan spans corporate functions, engineering, delivery, legacy modernization, agentic systems, and client work.

At that scale, artificial intelligence (AI) adoption is no longer a tool rollout. It is an organizational redesign conducted while the organization keeps operating.

A safety classifier is policy made executable

OpenAI's October 29 release introduces gpt-oss-safeguard, a pair of open-weight reasoning models that classify content against policies a developer supplies. Instead of fixing every moderation category during training, the system can interpret an organization's written policy at inference time.

That flexibility exposes a governance truth: a safety classifier is policy made executable.

Shared AI workspaces need shared accountability

OpenAI's October 23 update introduces a company-knowledge capability that can assemble context from connected workplace systems. The company also announces the acquisition of the maker of Sky, a desktop interface that can understand what is on a user's screen and act through applications.

Artificial intelligence (AI) is moving closer to where organizational work actually lives: across documents, messages, tickets, code, and the desktop. That proximity makes context more useful. It also makes accountability harder to fake.

AI infrastructure strategy is becoming full-stack strategy

Two October announcements, three days apart, reveal the widening scope of artificial intelligence (AI) strategy. OpenAI and Broadcom outline plans for custom accelerators and network systems at enormous scale. Google DeepMind and Commonwealth Fusion Systems describe a partnership using simulation, optimization, and reinforcement learning in the development of fusion energy.

One reaches down from models into chips. The other reaches out from models into a physical energy system. Both make the same point: AI strategy now spans the stack.

Visual agent builders do not remove engineering

OpenAI's October 6 introduction of AgentKit includes a visual canvas for composing and versioning multi-agent workflows, a connector registry, interface components, and expanded evaluation tools. The release makes a useful class of artificial intelligence (AI) systems easier to see and assemble.

Ease of assembly should not be confused with absence of engineering.

Long-horizon agents need short feedback loops

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.

At the edge, every joule is a model requirement

The Defense Advanced Research Projects Agency's (DARPA) September 24 update describes its Mapping Machine Learning to Physics (ML2P) program. The program aims to relate machine-learning performance to physical electrical characteristics so designers can optimize not only for accuracy, but also for the useful performance returned by each joule.

At the tactical edge, that is not an efficiency exercise. It is mission engineering.

READER-NEUTRAL SUBSCRIPTION

Follow Field Notes via RSS.

Copy this address into the RSS reader you already use. New notes will appear there automatically—no account, email address, or tracking required.

Open the feed