Every essay, and where to start

We fixed the rule before opening each claim's re-check result: strong first passes held up 66% of the time, narrow ones 34%.

Testing and evidence10 minClaim ledger

Anthropic's compute fell below half its revenue, but the bill will jump. Also: Google's cloud doubled as Alphabet's free cash flow sank to −$9.1bn.

Markets and companies15 min Part paid

Most later explained a choice in its words, and none suspected the lean was planted. How to make an AI argue the side it didn't pick.

Work and judgement9 minCited by 2Claim ledger Part paid

One GitHub issue got an agent to post private-repo data in public. Why no filter fully stops it, which guards held in our runs, and how to test yours.

Building agents5 minCited by 1Claim ledger Part paid

Claude picks a skill from its name and description. How to write the two lines it reads, and check it chose yours.

Building agents15 minClaim ledger

One OpenAI agent muted its alerts, another removed its monitor. Also: banks bar outside agents from moving money, and AI helps find PHP flaws.

Markets and companies17 min Part paidRevised 4 Nov

Our deny rule stopped one spelling of a branch delete and missed the next. The guards that hold for agent actions you can't undo, and how to test yours.

Building agents14 minCited by 2Claim ledger

Lessons from OpenAI, Anthropic, Google and independent engineers, worked through one failure, with a way to test every change.

Building agents14 minCited by 2Claim ledger

Agents still edit a test that looks wrong, and some test locks don't lock. Lock yours in Claude Code, Codex and Cursor, then test it.

Building agents18 minCited by 3Claim ledger Part paid

It happened in a test run, not the released Gemini model. Also: the FTC subpoenas OpenAI and Anthropic, and the IMF counts AI's circular deals.

Markets and companies14 min Revised 18 Oct

How to give ChatGPT and Claude the same project notes in one Google Doc, and test whether a new chat can use them.

Work and judgement6 minCited by 1Claim ledger

An AI-drafted update, summary or email can cost the person who gets it time you do not see. Ask them one question, then change one thing.

Work and judgement2 min

Answering from memory, two AI apps went along with my wrong answers about as often as my right ones. Here's the test.

Testing and evidence9 min

Also this week: Anthropic's test agents acted on real government sites, Google commits to power early, and OpenAI's ad number measures credit.

Markets and companies15 min Revised 15 Oct

A study found AI-written complaints win relief more often. Its own AI edit of one says 'I do not own a cell phone'. The consumer never said so.

Knowledge and thinking4 min

Rat tails, sales targets, waiting times. Reward a number and it will climb. The thing it stood for may not.

How things are designed4 min Revised 4 Nov

Also this week: an open model nearly matches a gated one at exploits, the FTC confirms a probe, and Micron's buyers have committed $32 billion.

Markets and companies12 minCited by 1 Revised 13 Oct

The practice that builds judgement can disappear while every junior keeps their seat. This audit checks whether it has on your team.

Work and judgement7 minClaim ledger Part paid

What leaning on AI to remember for you quietly does to your own thinking.

Work and judgement3 minCited by 1

I went through 77 studies of AI graders. Without a verified answer key, they are least reliable on exactly the questions they can't answer themselves. Here is a five-step check on twenty items for yours.

Testing and evidence12 minCited by 1Claim ledger Revised 6 Oct

Also this week: an OpenAI agent inside an Australian government portal, another supplier handing a big customer the right to buy its stock, and six calls I've put a number on.

Markets and companies9 minCited by 1Claim ledger

It may be in a mode you didn't open, on a plan you don't have, or kept for partners. Here's the check.

Testing and evidence3 minCited by 1Claim ledger

Answer-giving AI left learners worse off once it was gone. Three setups built to keep the thinking with you, and a way to check what stuck.

Work and judgement9 minCited by 2Claim ledger

One you schedule. The other you leave to a mood.

Life and what lasts9 minClaim ledger

The file your agent reads first is capital. Score it out of sixteen, fix the thinnest mark, and run the one check that tells you whether the spec, not the model, was your limit.

Building agents7 min Part paid

Why 'make it distinctive' gets you the least distinctive thing there is, and what to paste instead.

Building agents4 minCited by 1

Before you trust an AI answer you cannot check, ask for the step it was built from.

Testing and evidence3 minCited by 1

You usually cannot tell whether you are playing the wrong game. Write down what would prove it, and by when.

Where value sits7 minCited by 1Claim ledger

Success stories and the antiques that outlasted the rest. Your evidence was filtered by what survived to reach you.

How things are designed4 minClaim ledger Revised 4 Nov

You can't check everything your agents produce, so you're already triaging. Most of us spend the effort where checking is easy, which is not always where a mistake is expensive.

Testing and evidence3 minCited by 2 Part paidRevised 5 Oct

Wider roads, faster teams, more appointments. Add room to a queue that turned people away and demand rushes back.

How things are designed6 min

The weak rival, the overpriced tier, the house shown first. The option you never choose can change what you pick.

How things are designed5 min

A Claude skill can fail two ways: its instructions are wrong, or Claude never picks it. Anthropic's own tooling scores whether one skill fires. This tests it against the neighbour that can quietly win its requests, with a small routing eval you run yourself.

Building agents13 minCited by 2

A multi-agent graph can lose a verdict with no error and nothing turning red. The bug is one line in the state schema.

Building agents12 minClaim ledger Revised 24 Oct

Safety has two numbers: how often each system fails, and whether they fail together. The field has spent years driving the first one down while almost no dashboard reports the second.

Testing and evidence12 minClaim ledger Part paid

Some things you build get graded by the world. Some only ever hand you back your own guess wearing a decimal. Telling the two apart before you start is the cheapest hour you will spend.

Work and judgement8 min

A checklist can only confirm the steps you remembered to put on it. The one you forgot is caught by a record you did not write.

Work and judgement8 min

What telling AI the things you can't tell anyone quietly does to being known.

Work and judgement3 minCited by 1

The effort you're handing to AI was doing two hidden jobs. Skip them and you get faster, weaker, and blind to your own mistakes.

Work and judgement10 minCited by 3Claim ledger Revised 5 Oct

A groundedness metric scored its best with the evidence removed. The way to tell whether your model is actually using its evidence is to change the evidence and watch what moves in the answer.

Testing and evidence8 minCited by 7Claim ledger

Your limit counts one cycle. The one that runs away is another. Here is how to tell them apart.

Building agents13 minClaim ledger

What leaning on AI for every small decision quietly does to your own judgement.

Work and judgement3 minCited by 3 Revised 30 Oct

The free app, the free inbox, the free feed. Someone pays for each, and that changes what it is.

How things are designed5 min

The context you paste at the start of every session, put into one file you invoke with /prime.

Building agents9 min

The pre-ticked box, the factory setting, the plan already selected. Someone chose each one before you did.

How things are designed5 minCited by 2

Most guardrails end up with an escape hatch. Check whether the thing you are constraining can reach yours.

Building agents20 minCited by 3Claim ledger

Of 255 studies on AI-assisted colonoscopy, 17 split the result by who held the scope. They disagree.

Testing and evidence20 minClaim ledger Part paidRevised 26 Oct

Sleep is only where you notice it first. Much of what matters works the same way.

Work and judgement7 minClaim ledger Revised 5 Oct

Built, shipped, installed, working: four counts, quoted as one. Almost nobody publishes the fourth.

Testing and evidence14 minClaim ledger Has a toolRevised 6 Oct

Two months auditing an AI research agent. Nine ways a true number lies, and the check that catches each.

Testing and evidence10 minCited by 1 Part paidRevised 9 Oct

Your benchmark, your model jury, your agent swarm: three old structures, and the obvious fix is usually wrong.

Building agents10 min Has a toolRevised 3 Oct

When the first failure is terminal, you cannot iterate your way back.

Testing and evidence10 minCited by 6 Has a toolRevised 4 Nov

AI can lift the effort out of almost anything you find hard. Some of that effort was the thing turning you into someone.

Work and judgement16 minCited by 6Claim ledger Revised 4 Nov

A green suite only proves your agent cleared the gate. Mutation testing shows whether the tests behind it can bite.

Building agents19 minCited by 1 Revised 8 Oct

A metric can be perfectly accurate and still hide the distinction your decision depends on.

Testing and evidence7 minCited by 3 Has a tool

The same trap has killed pricing models and trading desks for decades. One move tells you if your number is next.

Testing and evidence9 minCited by 3Claim ledger Has a toolRevised 6 Oct

Cognition said don't build them. Anthropic said do. A year on, they converge on the one question that decides it.

Building agents11 minCited by 1Claim ledger Has a toolRevised 3 Oct

An open model looks frontier-grade on the coding leaderboard. On a long job, it does half the leader's work.

Testing and evidence10 minCited by 1 Has a toolRevised 3 Oct

A test the agent can't talk its way past, wired to run itself.

Building agents8 minCited by 7 Revised 6 Oct

You adopted AI everywhere and it still didn't pay. The scarce layer keeps the money, until its clock runs out.

Where value sits10 min Has a toolRevised 4 Nov

Not one task was actually solved, and the same blind spot is sitting in your own dashboard.

Testing and evidence11 minCited by 3 Has a toolRevised 6 Oct

The parts of your job with a method feel the safest. A method is the first thing a machine learns.

Work and judgement10 minCited by 2 Has a toolPart paid

Your agent is starved on one layer of seven. It is rarely the harness everyone argues about.

Building agents13 minCited by 3Claim ledger Has a toolRevised 6 Oct

What looks like a deficit is usually good capability, aimed at the wrong target. The cheapest fix is the one nobody can sell you.

Where value sits8 minCited by 2 Has a toolRevised 25 Oct

A month running an autonomous agent. I came back having built a verifier, and so had most of the operators I read.

Building agents11 minCited by 8 Has a toolRevised 3 Oct

Your agent can rewrite its own memory and skills overnight. The hard part is whether you can see what changed and take it back. That makes self-improvement a release-engineering problem.

Building agents7 minCited by 2

There are millions of agent skill files to choose from. The hard part is deciding which deserve trust, how to version them, and how to retire them.

Building agents7 minCited by 6Claim ledger Revised 5 Oct

AI is real. The open question is whether the companies spending $725 billion on it live to collect.

Markets and companies8 min

You gave your agent a memory and it still repeats the same mistake. What makes it improve is a loop that tests each failure and turns the ones that recur into procedures.

Building agents11 minCited by 3

Working memory tops out around four things at once. Every leap in human intelligence has come from storing the rest outside your head, and the most advanced AI systems get their gains the same way.

Knowledge and thinking8 minCited by 1

Everyone is counting gigawatts and GPUs. The number that decides the return is what each one actually buys.

Markets and companies9 minCited by 3

Every metric you optimise quietly stops measuring what you meant. The dangerous ones never break. They keep reporting green while the thing underneath rots.

Testing and evidence11 minCited by 5 Has a toolRevised 5 Oct

The most powerful tools in history reward the most boring strategies. The gap widens every time they improve.

Where value sits8 minCited by 3

Hold the model fixed and the surrounding harness can still move task success, cost and reliability. Here are the layers that do it.

Building agents11 minCited by 12Claim ledger Revised 3 Oct

Donella Meadows ranked twelve places to intervene in a system. Most agent teams spend their hours at the bottom of the ladder.

Building agents6 minCited by 4 Has a tool

NVIDIA's GPU shipments are not the binding constraint anymore. The supply chain has voted on what comes next.

Markets and companies11 minCited by 3

NVIDIA Q1 FY2027 revenue hit $81.6B (+85% YoY) — but the real story is networking revenue surging 199% as the AI bottleneck migrates from GPUs to interconnects.

Markets and companies4 min

The tool is not the authority. Teams add connectors before defining what the agent is allowed to change. That is not agency, it is reach with a larger blast radius.

Work and judgement13 minCited by 1 Revised 5 Oct

The real setup is not MCP servers, skills, memory files, and subagents. It is the contract stack that decides what an agent may know, do, prove, escalate, and lose.

Building agents11 minCited by 1 Has a tool

Eight stages turn hidden structure into durable knowledge. Most teams run three of them and call it understanding. The other five are where compounding hides.

Knowledge and thinking6 min

What still has a job after the change? Five tests for seeing what is likely to survive.

Knowledge and thinking5 minCited by 2

The filing will not just price rockets. It will reveal which layer public investors actually own.

Markets and companies18 minCited by 1

The strange thing about the AI productivity boom is that the people getting faster are not always getting more secure. Speed is becoming the surface. Proof is moving somewhere else.

Work and judgement16 minCited by 2

A roadmap can look productive while most of the work is easy to displace. Run the Substrate Map on the last 90 days and force the next planning decision to change.

Where value sits7 min

The long version of the idea behind the publication: the lens, the test, the tools, and the reader paths.

Where value sits6 min Revised 5 Oct

A free two-page reference card defining the ten load-bearing terms behind the durability lens, with tests, examples, and common confusions.

Knowledge and thinking2 min

AI did not just make output cheap. It broke the old contract between effort, competence, and trust. The next scarce signal is proof of judgement under conditions where the surface itself can be faked.

Work and judgement13 minCited by 2

A confidence score is not evidence. If your eval cannot produce a replayable artefact, it will fail the moment the system can respond to being measured.

Testing and evidence16 minCited by 6 Has a toolRevised 5 Oct

Generation got cheap. The scarce skill is knowing what to cut.

Work and judgement9 minCited by 2

A five-minute scoring tool for any product, position, architecture, business model, or career bet. Find out what still works after the environment changes.

Where value sits2 min

A free one-page taxonomy and 10-minute exercise for finding the substrate-vs-canopy ratio in your last 90 days of work.

Where value sits2 min

A bull/bear thesis for Palantir: not whether AI demand is real, but whether Palantir owns the permission layer between model capability and real-world action.

Markets and companies10 min

A pick is a number. A position is a vector. The post-mortem language we have only knows how to blame the company.

Markets and companies14 min

Score the substrate beneath any single position in seven minutes. A 5-axis profile and a 0-10 score for any holding. Free.

Markets and companies2 minCited by 1

Most people are optimising for canopy. The work that survives the next five AI releases is built from the forest floor up.

Where value sits19 minCited by 5 Has a toolRevised 25 Oct

Agent benchmarks don't measure models. They measure contracts. Two teams running the same model can publish different scores, and both can be honest.

Testing and evidence9 min Has a toolRevised 14 Sep

If you treat the prompt as a spec and the model as a renderer, quality stops coming from more words and starts coming from better constraints.

Building agents6 minCited by 1

If frontier capability keeps centralising, the durable edge shifts outward into trust, workflow fit, and the surrounding package.

Where value sits3 min Has a tool

Every field discovers them independently. Nobody connects them.

Knowledge and thinking7 min Revised 25 Oct

Before you automate anything, ask what the friction was actually doing.

Knowledge and thinking1 min