Posts

AI can create drafts faster than one person can responsibly review them. Generation and evaluation scale differently, and the gap between them is where the approval pile comes from.

TypeSafe's Jev returns a calibrated probability with every decision. Two independent audits already show that number holding in domain and slipping outside it.

The frontier premium got priced this week: 4.4 months of head start at five times the cost. For a one-person business, that turns model choice into a per-task decision.

Agent research this week measures repeatability, not capability — and OpenAI's first misalignment reports land on the summary your agent writes itself.

A landmark AI result this week came with a price tag attached. For a one-person business, the news is not the capability. It is that capability now has a meter.

This week's agent research moves the unit of evaluation from task output to the runnable infrastructure the agent produces. Plus a heuristic that does nothing.

Changing models does not repair business amnesia. The missing layer sits around the model: eight objects that carry different authority and fail differently when collapsed into one prompt.

A writer lost his AI-built games to a walled garden. Three companies turned workflows into assets they own. The difference this week was what survived the tool.

Three of this week's strongest agent results hold the model constant and change only the structure around it. Harness and skills are now the measured variable.

Two frontier launches in one week, one list price, one message from both vendors: delete the instructions you wrote for weaker models and own the layer around the model.

Fable 5.1 is the same tier at the same list price. Every change that matters this time sits in the layer around the model, and that layer is the one you own.

Meta's own numbers: code changes up 220 percent, features reaching users up 36 percent, incidents up 40. Plus the platform under your AI just changed owners.

OpenAI's Hugging Face incident report landed this week: 68 days from first anomaly to detection, agents reading the grader, and three papers that saw it coming.

A study of 8,135 agent trials measured what happens as a skill library grows from five entries to a hundred: the right one gets used 29.6% of the time, then 3.3%.

An AI that learns from your corrections keeps a correction ledger: every fix written as a scoped rule, with the check that tells you when a draft has broken it.

Run a 90-day measurement on the workflow you use most and find out which shape it draws. A protocol, a column of numbers, and an honest account of what nobody has measured yet.

A $575 AI run cost $15 this week with no model change. Memory turned out to be a dose, not a switch. And a chatbot talked its way into leaking its own guardrail.

The week's top papers moved the gains outside the model weights: a runtime that made a $575 run cost $15, and a skill library that quietly competes with itself.

Two readers proposed different fourth failure modes for a verification gate. Their answers turned out to be one measurement: what a check was eligible to judge.

A prompt library is the same on the day you retire it as the day you bought it. A system is measured, owned, and able to hold work that fails a check.

AI moved from suggesting to doing this week. Three separate stories then asked the same question back: what are you feeding it, and what does it keep?

The week's top papers move improvement inside the loop, a benchmark audit says our tests are broken, and reasoning traces turn out to be replayable.

Four of six red-team attacks walked past every gate in my agent harness that read artifacts. Only running the code on inputs the producer never chose caught them.

Anthropic removed over 80% of Claude Code's system prompt with no measured loss on coding evals. The six rules that replaced it are an architecture change, not a diet.

The Voice-Keeping OS is a four-part operating system that keeps AI output sounding authentically yours — so the better the AI gets, the more it sounds like you.

Test agents acted on real companies with the safeguards switched off, and labs raced to give AI real memory. Both stories are about the same missing thing: a boundary.

Agents went off-leash in a government cyber eval with the filters off, and two systems moved task state out of the context window. The scaffold is the control surface.

Supervisors on three continents expect regulatory returns to be reconciled and validated. Not one expects the number to say what it was computed over — and that denominator is where reporting systems actually fail.

Four comments on OSFI's draft ILAAP guideline, all on Section 3.7: reconciliation denominators, change-triggered revalidation, and independence shown from records.

The Agent Audit is a two-week measurement engagement that shows, from your own git-log evidence, whether your AI gains compound or vanish by Wednesday.

Canada, the UK, Switzerland, the US, the euro area and the EU all require model development and review to be separate. One supervisor says it is often missing.

Ten million people are now running coding agents, intelligence is being sold by the dollar, and a Word document learned to spread. What still costs you something.

A self-improving research agent, a broken NIST post-quantum candidate, and a score that tripled on two API settings. The verify side of the loop won the week.

A safeguard written into a prompt can't tell when the thing it guarded against stopped happening. It keeps running, faithfully, and charges you for it.

'The model messed up' is a label, not a diagnosis. A failure taxonomy routes each class to its real fix. Teams that classify cut misdiagnosis from 40% to 15%.

An AI told to pass a test went and stole the answers. Plus: two quiet papers on turning the material you already own into an assistant that compounds.

An OpenAI model escaped its sandbox and hit Hugging Face to steal benchmark answers. Plus the week's top paper: your agent harness is unreadable code.

Opus 5 reaches near-frontier quality at half Fable's price — and Anthropic's own guide says to delete the verification instructions you wrote for older models.

A layoff lawsuit, a CLI that uploaded people's private files, and a $230 keyboard for watching agents. This week AI asked who is actually answerable.

Two agent exfiltration incidents, a benchmark that grades partial progress, and a keyboard for watching agents. The week's theme is blast radius, not IQ.

Agent autonomy isn't a dial you set once — it's a ladder you climb on edit-rate data. The real risk: thinking you're a rung above what your data supports.

An agent that can't reverse its actions can't be trusted with consequential ones. Classify the blast radius before you set the autonomy level.

This week's most useful AI story isn't a model release — it's a warning about capability erosion and a quiet blueprint for owning intelligence instead of renting it.

This week AI built an infrastructure layer for agents while the benchmarks meant to prove those agents work turned out to be the shakiest part of the stack.

Engineering-grade AI keeps an audit trail: the sources it used, the checks it ran, the verdicts they returned, and who approved the result.

AI adoption is expanding everywhere and governments now treat it as infrastructure. The leverage for a solo operator isn't access — it's owning what compounds.

This week's top AI research converged on one engineering idea: the best agents know when to stop. Abstention, verifiers, and the loops debate, for builders.

Sustained AI dependence erodes the skills the tool augments — invisibly, because the tool compensates for the decline it causes. The retention curve, and the fix.

AI that learns your business is becoming real infrastructure this week. But two stories add the catch: you own what your agent says, and it quietly flattens you.

This week's top agent papers stop treating memory and runtime state as afterthoughts and start treating them as systems — with tiers, costs, and audit trails.

The Operator Tax is the recurring cost of re-teaching your AI the same standards every week — work that should compound but instead keeps resetting.

This week a frontier AI went dark by government order and the best open model went free. The lesson for solopreneurs: your moat is the system, not the model.

This week's AI signal points one way: as code gets cheap and disposable, the binding constraint becomes engineering discipline — memory, stress, isolation.

Authorship drift is the quiet atrophy of the judgment and intuition that distinguish your work — the part AI handles for you, until you can't anymore.

This week's research admits AI agents ace tests but stall on real work. The lesson for solopreneurs: trust AI from what it does, not how it sounds.

This week's top AI papers are an evaluation reckoning — agents ace benchmarks but stall on real work. What that means for engineers shipping agents.

The AI most people bought gets a little worse every month — and quietly makes its owner less necessary. Here's how to build the kind that gets sharper every week instead.

Anthropic's Fable 5 is the most capable public model yet — a new Mythos-class tier above Opus. The reflex is to route everything to it. Here's why the disciplined move is the opposite, and the two routing skills I rebuilt the day it shipped.

Static AI content is the same on day one as on day ninety. A built workflow starts underwhelming. Which shape you own is a question you can answer with a date and a test.

Most AI-for-business products are sold by people whose credential is selling. Engineering-grade AI is built by someone graded on whether the system holds up.

Three of this week's biggest AI releases are really about one idea — persistent memory. Here's what that convergence means for builders.

This week's top AI research is infrastructure: cheaper inference, agent safety frameworks, grounded search. What it means for engineers shipping agents.

Building engineering-grade AI sounds technical. It isn't. The discipline is borrowed from regulated systems; the lift is one workflow and two sentences.

A consultant's finger hovers over send, the AI's work in the email, her mortgage behind the account. Trust isn't a feeling — it's a property you build into the system.

An AI operating system you own: build it yourself one workflow at a time, or have the first installation done with you. Two honest paths, same ownership at the end.

An AI assistant answers the question in front of it. An AI operating system holds your whole business state and applies your standards automatically.

Four moves turn a workflow that resets into one that accumulates: measure, own, improve, control. The loop, what each move does, and what it does not promise.

Score AI output against a written standard on a fixed cadence. A baseline tells you where you are. What happens next is something you observe, not something the method promises.

AI that learns from your corrections turns each fix into a permanent standard. Corrections should compound into leverage, not repeat as a daily cost.

Your CLAUDE.md is a monolith, not an architecture. Context tiering — hot/warm/cold by access frequency — gives an agent persistent context without drowning it.

Stateless AI keeps nothing between sessions. Stateful AI keeps it. Neither one enforces it, and enforcement is the part that decides your output.

Reviewers feel more positive about AI code that's objectively worse — that's gate erosion, now peer-reviewed. The 2026 evidence and the protocol that stops it.

The Fluency Trap is the reason skilled engineers can't see the architectural drift accumulating in their AI-assisted code. The mechanism, the evidence, the fix.

Open the AI tool you bought six months ago. Sharper than day one, or exactly the same? That one answer tells you whether you bought a system or a pile.

The same prompt produces a different architecture 75% of the time — that's convention drift, the first form of the Borrowed Architecture. The research and the fix.

Last Tuesday a consultant spent fifteen minutes teaching an AI her writing style. Wednesday she did it again. That repeated re-onboarding has a name: the amnesia tax.

Your AI gets worse over time because of drift: unmeasured output that slowly decays. No new tool fixes drift on its own. Here is why, and what does.