Archive

The familiar story of internet-scale pretraining followed by human feedback is no longer enough. A frontier model is built through successive learning loops, each compressing different choices into the same weights.

The same model family solved every hidden-test repair through Pi and Grok Build, but one route reported sixteen times the total tokens. Most of the gap was context, not answers.

Summaries fail by keeping the wrong things, not by being too long. The load-bearing part is usually the bit that looks like boilerplate, which is why the cheapest model and the tail of a log are the wrong places to do it.

A PII scrubber in front of an LLM API checks the wrong property. The real options for confidential work are contract, local, or extraction — and extraction is judgment, not tooling.

Whether to use language models for every part of life is the wrong question. Trying one costs nothing and is how you find the tasks that did not exist; building around one is what costs, and the bill arrives as maintenance.

The runtime got the reviews. The transferable part is how a 2,500-pull-request, agent-written repository keeps itself honest, and one of its rules caught a stale instruction in my own setup the same afternoon.

EU GPAI Code adherence supports process reliance. An independent opinion on a specific model release still requires model- and deployment-specific evidence.

AI augments human agency only when people retain practical authority over delegation, scope, intervention, reversibility, contestability, and accountability.

Grand AI forecasts become useful when you stop accepting or rejecting the story as a whole and test the links that make its conclusion possible.

A working thesis, six months after my LLM wiki piece: plain Markdown plus relentless truth-keeping beats sophisticated storage plus neglect.

The sequel to the calibration gap: some governance claims have no number, and that does not excuse them from evidence. Pre-register the rule, invite the refutation, name the residual, report honestly.

AI governance keeps choosing between rigorous-but-slow validation and fast-but-unfalsifiable frameworks. The escape is the experiment itself: for any measurable claim, the measurement is the faster path.

The real test of enterprise AI ownership is whether the organisation can change models without losing the learning accumulated around them.

On the benchmark built to measure AI against real professional work, the best model scores 48% on deliverables that come in a file and 14% on deliverables judged as prose alone. That 34-point gap is the largest single effect in the paper, and it is not in the limitations section.

The most-quoted figure about AI and professional work is accurate, in the paper, and correctly transcribed. It is also a description of a scenario nobody would choose, because what travelled was the number without the condition it was computed under.

When agents generate faster than anyone can read, the standard answer is a second model reviewing the first. Aerospace decomposed what makes a check independent decades ago, and a reviewer model fails the hardest of the three tests.

A green result is a true statement about a stand-in. Whether that stand-in resembles the thing you care about is a separate question, and the check cannot answer it.

Deleting your agent's accumulated instructions is safest for the people who never tested them, and most dangerous for the people who did.

In April my benchmark said coding correctness is model-determined and the harness is a thin proxy. Two newer results show exactly where that claim breaks: the harness matters at the edge of what the model can do.

Cerebras published an unusually honest account of their internal knowledge base. The retrieval engineering is solved. The assurance layer gets one sentence. Every enterprise copying this pattern is importing that asymmetry.

Coordination cost reveals valuable AI workflows, but only cash conversion and competitive durability make them valuable businesses.

Good and redundant are not opposites — they're the most common combination.

For many agent workflows, the right abstraction is not another tool call. It is a bounded program over small primitives.

The useful unit in agent systems is not a better instruction. It is a tested capability package: judgment, code, checks, routing, and boundaries.

Model revenue is not customer return. The economic and risk unit is the harness that turns model output into accountable work.

Once model companies supply the generic agent harness, the valuable work moves into workflow design, human intervention, domain data, and the definition of good work.

Enterprise AI stops being a chatbot when the operating procedure becomes the thing the system can execute, inspect, and improve.

Long-running agents are not defined by the model call. They are defined by the state, rules, tools, failures, and corrections that survive it.

Ken Griffin watched PhD-level finance work compress from months to days. The interesting question is whether bank AI controls are designed for the layer where the work now lives.

If a product wants agents as real users, first-run setup has to be an executable workflow, not a human signup ceremony wrapped in documentation.