Library

Task duration turns autonomy into an interpretable capability measure, but the forecast remains conditional on software-heavy tasks generalising and the measured trend continuing.

Anthropic's own vocabulary rather than settled practice, since the wider conversation organises around the harness and not a loop taxonomy. The demand is still the useful part: make a loop declare what triggers it, what stops it, which primitive carries it and what task it suits. That forces the stop condition into the open, and the stop condition rather than the trigger is what actually caps the cost.

What you withhold is what the experiment measures. Keeping the agent off the internet and away from existing compiler source is clean room; withholding the instruction-set documentation and the published work on register allocation and instruction selection just tests recall of what any engineer would look up first.

Testing shaped around what actually changed rather than a fixed suite, which is the discipline a release gate needs once an agent is writing the diff.

The clearest statement that continuity across sessions is an architecture problem, not a prompt problem.

Pull over push for agent context: give the agent less and let it retrieve, with files as the universal primitive.

Files quietly won as the central abstraction for agent memory, retrieval, and skills, displacing heavier machinery.

When behaviour emerges from code plus model, traces replace source files as the debugging surface.

The 80/20 flip in weeks, and the observation that model failure is now conceptual rather than syntactic: hasty-junior mistakes, not syntax errors.

The line worth keeping between what you delegate to a model and what you must still hold yourself.

What development looks like once the bottleneck is inference time rather than code: parallel projects, output streams monitored rather than read.

My take: The Trust Spectrum

AI work compounds only when artefacts, decisions, and corrections persist as context for future sessions; the essay behind a lot of how this site’s author works.

An enterprise knowledge base that rejects the single source of truth and meets data where it lives; unusually honest engineering, with real thresholds attached.

My take: Retrieval Becomes a Control

Local-first sync as a worked example of an architecture bet paying off in feel, not just benchmarks.

A useful way to see why ninety-something percent can still be nowhere near finished. Average competence and complete delivery are different measurements, and the latter is often the one that matters.

The hard part of professional review is not finding some good answers; it is covering a vast evidence field without silently missing the decisive issue. A benchmark moved into a new runner should be treated as a new test, not the same score with a different command.

The primary account of why rule-heavy scaffolding written for weaker models now hurts — Anthropic cut over 80% of Claude Code’s system prompt for the Claude 5 generation with no measured loss on their coding evals. It rewired my own skill-review gate the day I verified it.

The compression is the point. A few hundred gigabytes holding most of what we collectively know is a medium artists and scientists can work inside, not merely a thing that answers questions.

Reasoning fills the gaps between facts we already hold, which is near-godlike where the solution space is smooth and checkable and no help at all where the system is emergent. The limit was never thinking; it is how fast reality answers, and that clock barely moves.

High-altitude binaries create the appearance of strategy while keeping the room away from the hard customer-specific question.

A job is a bundle of problems someone wants solved, not an object on a board; the best career reframe of the year.

Personal software as software shaped by one person’s judgment rather than a market’s average need.

The disciplined counterweight to AI maximalism: constrain the tool to the box where it demonstrably helps.

AGI defined pragmatically as the ability to figure things out, with the three ingredients already shipped.

Organisational integration beats raw model power; institutional capability is the differentiator, not model choice.

A FINRA-modeled standards body for frontier models, from voluntary pre-release review to a mandatory market-access gate; the second frontier-lab CEO in a month to converge on mandatory third-party testing.

Arrow’s 1962 paradox inverted for AI: the buyer risks giving away knowledge just by using what they bought; the countermeasure is his five-move sequence — Control, Capability, Choice, Cost, Compound.

Complex tasks see larger AI speedups than simple ones, inverting the usual displacement narrative.

Benchmarks measure a model operating inside a frame humans supplied; as framed competence gets cheap, the human work moves to problem-definition, review, and exceptions.

My take: After Automation, Judgment Becomes Infrastructure

A scenario, not a forecast: what happens if AI succeeds and the economy’s redeployment loop breaks anyway.

The organisational thesis: repeated work rebuilt as governed feedback loops, with humans kept at the high-stakes edges.

A first-person field report of company-wide agent adoption; the numbers are self-reported and unaudited, but the shape of the operating model is worth reading anyway.

A five-stage ladder for engineering-org agent adoption where every transition is gated by trust in the verification loop, not agent count; the products column is an Anthropic catalogue, but the guardrail-per-stage pairing is the part worth keeping.