🎧 Briefs · 🖼️ the infographics →
What a summariser heard, two minutes at a time.
Google’s NotebookLM makes a two-minute brief from whatever sources it is handed. Each brief carries one argument, so a stack of them reads like a census: the lines it got right show what our sources say clearly, and the lines it got wrong show where our sources still compress badly.
29 briefs · 4 days · 713 lines heard · 17% quoted as said ✅ · 12% corrected ❌
The errata, as a map of where we are
- 🔒 55%promises the agent stays inside a line · 16 of 29 briefs
- ❓ 41%a number no repo file carries · 12 of 29 briefs
- 🛑 28%makes the tape a brake: kill switch, failed build, forced boundary · 8 of 29 briefs
- 🎲 28%uses deterministic to describe us · 8 of 29 briefs
- 💸 24%sells rung 2 (99.85 %) as token-bill savings · 7 of 29 briefs
- 🧾 17%claims to prove what the agent did from its own account · 5 of 29 briefs
- 📏 17%reads a real /backup number as what it does not measure · 5 of 29 briefs
- 🐺 7%reads our outreach as pressure on a person · 2 of 29 briefs
- 🔭 7%states a conjecture from our sources as settled physics · 2 of 29 briefs
- 🦹 3%casts the labs as the villain · 1 of 29 briefs
- 🔩 3%grounds the claim in metal or friction instead of a record the actor did not write · 1 of 29 briefs
The errors repeat. A brief that promises a cage read a source that names a boundary and stops before the record. A brief that sells 99.85 % as savings read a rung-2 number printed without its counterfactual. Each recurring class points back at a sentence in our sources to fix, and the next brief made from the fixed sources is the check.
📅 Made 2026-10-13
📝 Why Reality Needs Friction Not Code
/brief/5ca772df · 103 s · ✅ 18% · ✏️ 6% · ❌ 12% · mdx ★★★★★
❌ This is the only one of the three that lands back on the agent: an AI reading its own logs drifts out of touch, the sticky-note point exactly. Then it resolves the drift with “hardware friction” and “only active physical work enforces the boundaries”, where our draft resolves it with a record the agent did not write.
2 errata · 🔩 metal · 🛑 brake
“Take your thermostat at home. Writing 72 degrees on a sticky note isn't going to magically heat up a room, right?” 00:29
⚛️ Why the Universe Runs on Friction
/brief/7a54aa82 · 91 s · ✅ 10% · ✏️ 5% · ❌ 15% · mdx ★★★☆☆
❌ “Math is just the log book” is our line in five words. The brief then spends a third of its length on quantum collapse and tells the listener “we don’t need a conscious observer at all”: the draft’s most tentative conjecture, promoted to its headline.
3 errata · 🔭 conjecture
“Think about it. You can't heat a room by writing 72 degrees on paper.” 00:23
🫠 Why Code Cannot Fight Physical Decay
/brief/9d7669a7 · 105 s · ✅ 13% · ✏️ 4% · ❌ 8% · mdx ★★★★☆
❌ The melting laptop is the sharpest picture any brief has made of a record that cannot see its own failure. Then the brief walks away from the agent entirely and spends its last third on particles “hitting non-negotiable boundaries”, a conjecture in our draft that it states as physics.
2 errata · 🔭 conjecture · 📏 stretched
“Like, a computer simulating a cell will happily report that its digital cell is doing perfectly fine, even as the physical laptop itself melts into a puddle of slag.” 00:30
📅 Made 2026-10-12
🛣️ How Flight Recorders Make AI Insurable
/brief/189466c0 · 109 s · ✅ 25% · ✏️ 0% · ❌ 5% · mdx ★★★★☆
❌ Carrying the lattice and the record our sources state in one breath, it blurred two separate laws — Rice’s undecidability and the data processing inequality — into one 1953 rule. The read is clean exactly where the source pairs its sentence; it slips where the source splits it.
1 errata · 🧾 proof
“That's uninsurable, quote, single-entry hearsay, because it's just the AI grading its own homework.” 00:26
🧮 How Math Makes Autonomous AI Insurable
/brief/27575c0f · 94 s · ✅ 18% · ✏️ 5% · ❌ 9% · mdx ★★★☆☆
❌ The brief reached for an 'inescapable physical toll booth around its hands' and 'unlocking a trillion-dollar market' in the same two minutes it also said 'it doesn't control thoughts, right? It just measures actions.' Our sources state the disclaimer once, cleanly, and let the toll-booth metaphor run a full minute ahead of it without the same sentence attached.
2 errata · 🔒 cage · ❓ unsourced
“It creates an independent hardware-level Richter scale.” 00:52
🔏 How Cryptography Makes AI Insurable
/brief/29f3f2bc · 106 s · ✅ 13% · ✏️ 0% · ❌ 13% · mdx ★★★☆☆
❌ Keeping the receipt’s signature and the honest aside that the savings are secondary, it closed on “mathematically caged” exactly where our sources stop at the declared lane. The boundary tightens in its sentences; ours only gets checked after the run.
4 errata · 🔒 cage · 💸 savings
“Right now, deploying enterprise AI is basically an unpredictable legal nightmare.” 00:04
🧯 How To Insure Autonomous AI
/brief/3ac4c13f · 91 s · ✅ 13% · ✏️ 7% · ❌ 10% · mdx ★★★☆☆
❌ The captain line and the closing 'verifiable shape of its deeds' landed clean. A sentence later, the brief called the boundary “deterministic” and then described it “instantly stopping a digital plumber” — the halt and the chaos-word both sit on the buyer's side of the line in our sources, never the record's side to claim.
3 errata · 🎲 determinism · 🛑 brake · ❓ unsourced
“Right now, safety measures fail because we're basically letting the AI grade its own homework.” 00:16
🧘 How Bounded Sub-Agents Fix Unbounded AI
/brief/4de9f837 · 104 s · ✅ 21% · ✏️ 0% · ❌ 8% · mdx ★★★★☆
❌ It found spec leaves and the haystack, then printed 99.85 % as savings. The rung-2 figure still travels without its counterfactual.
2 errata · 💸 savings · 🔒 cage
“and tells the patient to do box breathing, which House absolutely roasts as Wally doing yoga.” 00:28
🕵️ How Flight Recorders Make AI Insurable
/brief/4e8b8b8b · 99 s · ✅ 14% · ✏️ 4% · ❌ 7% · mdx ★★★★☆
❌ Keeping Rice’s theorem and the drift-not-ghosts argument intact, it reached for “tamper-proof” exactly where our sources stop at re-runnable. Cleanest of the five — the drift sentences land clearer here than the boundary ones do.
2 errata · 🔒 cage · 🎲 determinism
“Relying on an AI's own guardrails is literally just software checking software, which is a total mathematical dead end.” 00:23
🔌 How Hardware Makes Autonomous AI Insurable
/brief/50808659 · 106 s · ✅ 14% · ✏️ 4% · ❌ 11% · mdx ★★★☆☆
❌ Building an entire act out of “physical hardware” and “deterministic data” exactly where our axiom draws the line, it treats a cache miss as the grounding argument instead of a receipt the walk can read, and the chip’s own counter as something already shipped. The hallucinating-brain joke checking itself is the one line it got dead right.
3 errata · 🎲 determinism · 🔒 cage
“a rule called Rice's theorem proves software evaluating other software is fundamentally undecidable.” 00:34
🚑 Why AI Fails in Medical Emergencies
/brief/508bff98 · 108 s · ✅ 22% · ✏️ 0% · ❌ 9% · mdx ★★★★★
❌ The strongest read in the stack: it reached the maximum-likelihood estimator on its own. The two errors are the usual pair — 99.85 % sold as savings, and a limit on the model that we never set.
2 errata · 💸 savings · 🛑 brake
“I mean, imagine calling 911 and getting an automated meditation guide.” 00:33
🚢 How to Insure Autonomous AI Agents
/brief/7fd3ec08 · 113 s · ✅ 23% · ✏️ 5% · ❌ 14% · mdx ★★★★☆
❌ It named Rice and the real customer, then reached for “deterministic” and a merge block. Somewhere in the sources the word still describes us.
3 errata · 🛑 brake · ❓ unsourced
“Because deploying today's enterprise AI is basically like loading cargo onto unrated 18th century merchant ships.” 00:08
🌋 Insuring Autonomous AI With Hardware Telemetry
/brief/931753f3 · 92 s · ✅ 19% · ✏️ 4% · ❌ 0% · mdx ★★★★★
❌ Nothing here promises the agent stays inside a line — the earthquake metaphor and the record-the-AI-didn't-write line both read back exactly as our sources state them. Sentences that keep the disclaimer and the mechanism in the same breath survive a summariser's paraphrase; the other four briefs drifted exactly where that breath got split into two separate sentences.
0 errata ·
“A tech startup called ThetaDriven moves verification totally outside the AI's brain.” 00:39
🧾 Cryptographic Flight Recorders for Autonomous AI
/brief/af0a05cb · 103 s · ✅ 9% · ✏️ 3% · ❌ 15% · mdx ★★★☆☆
❌ Reaching for “mathematical cages” twice, exactly where the tape does its real work, it let 99.85% drift from payload size into a savings figure. Our sources state the lane and the receipt in one breath, but stop before the sentence that says what stays undecided.
5 errata · 🔒 cage · 🧾 proof · 💸 savings · 🎲 determinism
“It's kind of like painting a plimsoll line on an 18th-century merchant ship.” 00:36
✈️ Cryptographic Flight Recorders Make AI Insurable
/brief/ca1c8809 · 95 s · ✅ 11% · ✏️ 4% · ❌ 11% · mdx ★★★☆☆
❌ It heard “flight recorder” and filled in a kill switch. In our sources the recorder still stands next to a stop button, and the summariser joined the two.
3 errata · 🛑 brake · 💸 savings · 🧾 proof
“Honestly, using AI to monitor AI just doesn't work.” 00:58
📼 How Cryptographic Audit Tapes Insure AI
/brief/dbf7efed · 101 s · ✅ 18% · ✏️ 7% · ❌ 11% · mdx ★★★★☆
❌ It answered its own hardest question right — no guarantee — and then called the tape a shield. Our sources state the limit sharply and the remedy loosely.
3 errata · 🔒 cage · ❓ unsourced · 🛑 brake
“Now, wait, you might be wondering, does this guarantee the AI won't make a mistake?” 00:36
⚓ Making Autonomous AI Agents Insurable
/brief/e92eb988 · 92 s · ✅ 27% · ✏️ 4% · ❌ 8% · mdx ★★★☆☆
❌ It quoted the Plimsoll line and the buyer cleanly, then said “mathematical cage” twice. Our sources name the line and the buyer in one breath and stop before saying who reads the record.
2 errata · 🔒 cage · ❓ unsourced
“Think of the Plimsoll line, painted on 18th century merchant ships.” 00:12
📷 Mathematical Boundaries Make AI Insurable
/brief/e95f4ce0 · 105 s · ✅ 19% · ✏️ 4% · ❌ 15% · mdx ★★★☆☆
❌ The brief reused “deterministic boundaries” for the lane and “perfectly measurable and contained” for the mistake, both phrases our sources keep for the agent's chaos, never for the receipt handed to the underwriter. A breath earlier, the camera-versus-vest line got the distinction exactly right — that is the sentence worth keeping word for word.
4 errata · 🎲 determinism · 🔒 cage · 💸 savings · ❓ unsourced
“It's totally like confusing a security camera with a bulletproof vest.” 00:30
🧾 How Receipts Make AI Insurable
/brief/fe0d420d · 113 s · ✅ 14% · ✏️ 4% · ❌ 7% · mdx ★★★☆☆
❌ The surgeon line kept the disclaimer our sources insist on, then a breath later called the boundary adherence evidence “deterministic, hard evidence of boundary adherence.” Our source material separates the receipt from the chaos it watches; this brief closed that gap by reusing the one word kept for the agent, never the record.
2 errata · 🔒 cage · 🎲 determinism
“We're not promising the surgeon performed a perfect surgery.” 01:00
📅 Made 2026-10-05
🌊 Mathematical Boundaries Stop AI Runaway Loops
/brief/02d14790 · 112 s · ✅ 13% · ✏️ 3% · ❌ 13% · mdx ★★★★☆
❌ It took the real /backup figures and turned an observation into a mechanism: 310.4M “truncated” to 15.4M. The curve is printed without the sentence that says what the pair compares.
4 errata · 🔒 cage · ❓ unsourced · 📏 stretched
“because they're treating safety as a moral problem you solve with polite prompts” 00:06
🎳 How ThetaDriven Makes Autonomous AI Insurable
/brief/3128016e · 109 s · ✅ 20% · ✏️ 5% · ❌ 15% · mdx ★★★★☆
❌ Bumpers are a brake; the 2 % line is the keeper. The $11.47 is in no file — the summariser needed a number for the cap and made one up.
3 errata · 🔒 cage · ❓ unsourced
“It's kind of like insuring a bowling alley.” 00:38
💥 The Physics of AI Token Explosions
/brief/460a0377 · 99 s · ✅ 18% · ✏️ 0% · ❌ 18% · mdx ★★★★☆
❌ The metaphors are strong and ours; the overreach is not. “Deception becomes mandatory” is nowhere in the record — the brief upgraded token spend into intent.
4 errata · ❓ unsourced · 📏 stretched · 🔒 cage
“Treating that massive sprawl as a chatbot hallucination is honestly like calling a nuclear meltdown a spelling mistake.” 00:35
🪪 How to Insure Autonomous AI Agents
/brief/7ff668e0 · 103 s · ✅ 17% · ✏️ 3% · ❌ 10% · mdx ★★★☆☆
❌ It caught “labels missing data as unmeasured”, a line almost no reader catches. Then bumpers again: the promise slips back in through the analogy, never through the claim.
3 errata · 🔒 cage · 🧾 proof
“Because of a computer science rule called Rice's theorem,” 00:23
🛫 How to Insure Autonomous AI Agents
/brief/855fd814 · 106 s · ✅ 11% · ✏️ 5% · ❌ 21% · mdx ★★★☆☆
❌ It called rung 2 an in-lane percentage and sold it as 30 % off the bill. The rung-2 label on /backup does not survive two minutes of compression.
4 errata · 🎲 determinism · ❓ unsourced · 💸 savings · 🛑 brake
“Think of it like bowling with bumpers.” 00:42
🚨 Why AI Safety Needs Mathematical Boundaries
/brief/91471313 · 94 s · ✅ 11% · ✏️ 0% · ❌ 18% · mdx ★★☆☆☆
❌ It opened on a breach that never happened. A notebook with a thesis and no scene gets a scene invented for it — the cost of sources that open on the machine.
5 errata · ❓ unsourced · 📏 stretched · 🦹 villain · 🔒 cage
“Asking a multi-million token AI to be polite and honest is like asking water to politely avoid your furniture when a dam breaks.” 00:30
⚡ Why AI Safety Needs Circuit Breakers
/brief/e8f40383 · 89 s · ✅ 21% · ✏️ 0% · ❌ 21% · mdx ★★☆☆☆
❌ Invented incidents again, and the circuit breaker twice. Two briefs from the same day opened on fiction, so the fix is a true scene at the top of the source.
4 errata · ❓ unsourced · 🛑 brake · 📏 stretched
“First, AI safety is a structural problem, not a behavioral one.” 00:12
🏎️ How ThetaDriven Insures Autonomous AI Agents
/brief/ec77bbe9 · 101 s · ✅ 20% · ✏️ 5% · ❌ 15% · mdx ★★★☆☆
❌ The read-only link and the underwriter’s own keys come through clean. “100 % certainty” does not: the placement is certain, the verdict is undecidable, and the brief merged the two.
3 errata · 🔒 cage · 🎲 determinism · 🧾 proof
“Since it's mathematically impossible to guarantee an AI will always be semantically correct, thanks to a little rule called Rice's theorem, this framework doesn't warranty perfect logic.” 00:31
📅 Made 2026-10-04
🧱 The AI Risk Underwriting Playbook
/brief/148be342 · 83 s · ✅ 16% · ✏️ 5% · ❌ 11% · mdx ★★★☆☆
❌ It read our outreach playbook as pressure on a person. The playbook hands over a check and one question; the brief turned the timing into a threat.
2 errata · 🐺 predator
“man, if I don't move now, a carrier sets the standard and I become a price taker.” 00:21
🪑 Pitching AI Risk Standards to Brokers
/brief/4b4128b5 · 91 s · ✅ 24% · ✏️ 0% · ❌ 14% · mdx ★★★☆☆
❌ Same playbook, same misreading: “weaponizing their internal politics”. Two of two briefs read the brick sequence as force, so the sequence needs its why written beside it.
3 errata · 🐺 predator
“It's like high stakes musical chairs.” 00:34
Sufficient for: what Google NotebookLM said in each two-minute brief it made from our sources, verbatim, with its timestamp, the day it was made and the file it came from. Not sufficient for: what our sources say — a brief is a projection of the notebook, never the notebook. Percentages are of the transcript's lines: ✅ quoted as said, ✏️ quoted beside the sentence we would sign, ❌ corrected. Each address is the first eight hex of the audio file's md5. The same summariser's pictures, read the same way, are at /infographics.