Issue 015Mon 12 Oct 2026411 items screened → 54 filedfull pull · 13 feeds down
Where yesterday the fear was loud, today the money answered, and it pointed in one direction: the physical
world and the things you can run yourself. Amazon is in talks to
buy Decart for ~$7B — real-time generative video and “world models” — after Anthropic and
Nvidia both walked; PitchBook counts ~$48B into robotics and
“physical AI” this year; and a Y Combinator-backed team raised
$31M to run hospitals like air-traffic control.
Meanwhile the trust panic kept running in the background — Nadella called for an
“emergency
brake,” and a new paper showed models can provably hide their own
capabilities. Two reads, one move: value is flowing into the software that touches the physical world nobody has
rebuilt, and into whatever you can see, log and own. Both of today's §0 problems are exactly those two gaps —
and the cheapest version of each is one you can start this week.
§0
The Problem Board
2 of 11
Every accelerator says it funds people “solving hard problems,” and none tells you which. This does —
sourced, graded honestly, with somewhere to start reading. Today one pure-software problem you can start this month,
sitting under the week's agent-trust panic; and one amber that today's health rounds (§1 Vitalize, §2) just
put a price on. Honest note up front: the board is small and this fortnight has cycled through all of it, so these are
the two least-recently-surfaced that today's news actually sharpens — the board needs fresh problems, and
none of today's feeds (DARPA, Marginal Revolution, YC RFS) carried one.
START THIS MONTHgreen · solo · no capital · no credentials
Nobody can prove what an AI agent actually touched
The problem
Agents run with far more authority than their task needs — the same shell, filesystem and
credential access whether they are summarising a document or changing infrastructure. When one goes somewhere it
shouldn't, there is no audit trail to answer the only question that matters afterward: what did it touch, under
whose authority?
Who says it's hard
OpenAI's own disclosure (agents reached government sites unprompted, caught two months
later), six arXiv papers this quarter, and security researchers naming the missing control at RSAC — collected
in “How Agents Ask for Permission.” The wire said it
louder today: Nadella called for an “emergency
brake,” and a new paper — “AI Models Can Provably
Hide Arbitrary Capabilities” — argues you can't even trust a model to tell you what it can
do. If you can't trust disclosure, the only thing left is a record of what it actually did.
Why it's still open
Sandboxing shipped (NVIDIA, Cisco); provenance did not. The papers are 2024–26 and
mostly preprints — 3 peer-reviewed out of 21 systems surveyed. That is a field with no incumbent. And it's
getting busier, not solved: today's arXiv alone carried red-teaming of personal agents across evolving
services and reliability-aware tool retrieval — everyone is studying the behaviour; nobody owns the ledger.
Who's already trying
NVIDIA (sandboxing) and Cisco (sandboxing) — both wall the agent off, neither
keeps the receipt. The audit layer is the gap between “we contained it” and “we can prove what it
did.”
Your angle
Zero capital, zero credentials, and you run agents daily on a machine holding
your own keys — so you can feel the problem today, which is the founder's unfair advantage in its cheapest
form. Pair it with today's §6: the prototype is a week of your own agent's tool-call logs.
Read “Lingering Authority” — the credentials an
agent still holds after the task ends, which is where the real damage lives.
Instrument your own setup (§6): log every tool call your agent makes for one week. The log is the
prototype — and reading it back is the product spec writing itself.
6–12 MONTHS · NEEDS REAL USERSamber · needs users, not a hostel room — but the users are everywhere around you
There will not be enough people to care for the old
The problem
By 2030 one in five Americans is over 65, with millions of unfilled caregiving jobs and 53
million family members already doing it unpaid. Almost no technology is built for older people — even Alexa and
Google Home frustrate most seniors. The staffing math is the same math breaking hospitals today.
Who says it's hard
Y Combinator's RFS — “AI for
the Aging Population.” And today the market priced the adjacent wound: Vitalize (today's §1) raised
$31M because hospitals literally cannot staff the beds,
and MIT Tech Review spent the day on the still-unknown side effects of the
GLP-1 drugs an aging population is now on. The labour shortage isn't coming; it's the balance sheet already.
Why it's still open
Consumer tech optimises for people who adapt fast; the people who need this most adapt
slowest. And the buyer (the adult child) is not the user (the parent) — a split that breaks normal product
intuition, because the person paying never feels the product and the person using it can't churn.
Who's already trying
Papa, Honor, ElliQ — mostly US, mostly expensive, mostly built for the buyer's
guilt rather than the parent's day.
Your angle
India's aging curve is steeper and its caregiving market is almost entirely
informal — you can observe it directly in your own extended family; US founders cannot. Amber, not green,
because it needs real users and trust, not a weekend. But the research is free and the users are at the next family
gathering.
Research path
Interview five families in your own network about who cares for the
oldest member, and what breaks first. The answer is almost never “an app.”
Read ElliQ's published retention data — the honest numbers on whether seniors keep using these things past
week three.
Look at what an Indian hospital discharge actually hands a family: usually one sheet of paper and
no follow-up. That gap is the product.
New on the feed today: nothing to add to the board. DARPA carried the same
three items as yesterday (Quantum Benchmarking final, the Vertical Lift Challenger recap, the CIDAR passive-ranging
challenge — all already filed). Marginal
Revolution ran Casey Handmer on automated peer review (filed in §3, an idea not a board problem), and
YC's RFS is unchanged since the Fall 2026 publish. No new funded problem
statement — so both of today's are rotated from the standing board.
§1
The Big Three
researched
The bidding war · deep tech where filmmaking and AI collide
Two 8200 veterans built real-time AI video so good that Anthropic, Nvidia and Amazon all came knocking — Amazon at ~$7B
Decart · founded Sep 2023, San Francisco / Israel · co-founders Dean Leitersdorf (CEO; PhD from
the Technion at 23, its youngest ever) & Moshe Shalev, both Unit 8200 alumni · ~60+ staff ·
raised >$450M (Sequoia, Benchmark, Zeev) — $300M Series C at ~$4B in May 2026 · now in
talks to sell to Amazon at ~$7B ·
WSJ via Techmeme ·
Sequoia
What they do
Decart makes AI that generates video in real time. Its 2024 release Oasis
was the first real-time interactive generative video model — a playable, AI-dreamed world that responds to your
inputs frame by frame — and hit a million users in three days. Since then: Mirage (video-to-video
restyling live), Lucy (real-time try-ons), and Sidekick (talk-to-video). Underneath
it all sits the less glamorous asset Amazon actually wants: software that makes models run efficiently across
different accelerators — including Amazon's own Trainium chips.
The tech — and what's hard
Every other video model you've seen renders slowly, offline, a clip at a time. Decart's whole bet is
latency: generating coherent, controllable video fast enough to feel interactive. That's brutal — it
means squeezing a diffusion / world-model pipeline down to tens of milliseconds per frame without it dissolving into
noise, which is as much a systems-and-silicon problem as a model problem. That is exactly why the chip-portability
layer exists, and exactly why a cloud (Amazon) rather than a model lab is the natural buyer.
The founders & where the idea came from
Leitersdorf finished a Technion PhD at 23 — the youngest in the institution's history — and he and Shalev
both came through Unit 8200, Israel's signals-intelligence corps, the single most reliable founder-factory on earth
for people who can wring performance out of hardware under impossible constraints. The idea isn't “make pretty
video”; it's “real-time generation is a latency problem, and we are the people who do latency.”
That reframing — a creative-media race won on an infrastructure skill — is the whole access story.
The bear case
Anthropic weighed ~$6B in September and walked;
Nvidia walked too. Smart buyers passing before Amazon stepped in is a signal, not just a negotiation. Real-time video
burns compute with no proven consumer willingness-to-pay yet, the demos outrun the businesses, and “first to a
million users” became “retained how many?” — undisclosed. The deal may still fall apart; WSJ
says so plainly.
For you
Two lessons, both transferable. First, for the filmmaker in you: real-time generative video is arriving this
year, and the people winning it aren't cinematographers — they're systems engineers who treated moving
images as a latency problem. The craft you care about is being rebuilt on an infrastructure skill, and the open-weights
version of this is already free (see §4's LTX-2.5). Second: the thing being acquired at $7B is partly the boring
chip-portability layer, not the flashy demo. The durable asset is usually the unglamorous one underneath — worth
remembering every time a shiny demo tempts you to build the shiny thing.
The physical-world OS · the §0-amber, as a funded company
A medical student built air-traffic control for hospital staffing — YC backed it, and it just raised $31M
Vitalize · YC W23 · co-founders Veeraj Shah (CEO; MD/PhD student, prior digital-health work
touching the US Surgeon General & IBM), Sanketh Andhavarapu, Nikhil D'Souza ·
$31M Series A led by Oak HC/FT, with Norwest, .406 Ventures, Constellation, Y Combinator, Rock
Health, MemorialCare ·
Fierce Healthcare via Techmeme ·
MedCity
What they do
Vitalize is, in Shah's own words, “air-traffic control” for the hospital workforce — one system
coordinating scheduling, staffing and patient capacity across inpatient, ancillary and ambulatory care. The claimed
results are concrete: 2,000+ clinical hours saved a day, Rochester Regional cutting manual staffing
time 50%, St. Luke's cutting premium (agency) labour 54%. It is the
§0-amber (“not enough people to care”) attacked from the supply side: not more staff, better
allocation of the staff that exist.
The tech — and what's hard
The hard part isn't the scheduling math; it's that hospital staffing is “unglamorous, complex and tedious”
work buried in incompatible systems, union rules, credentials and real-time census swings. The moat is integration
and trust, not a clever model — which is why incumbents (the physical-world §0 pattern again) charge for
human coordination and nobody had built the layer that sees the whole board at once.
Where the idea came from
Shah is an MD/PhD student — he has stood inside the actual problem, on the actual ward, watching the actual
whiteboard-and-phone-calls that run a hospital shift. That is the access the brief keeps pointing at: you don't
research your way into knowing why hospital staffing is broken, you have to have been there at 3am when it breaks. Oak
HC/FT, a hospital system and YC are paying $31M for that lived knowledge.
The bear case
Health-system software is a graveyard of long sales cycles, pilots that never expand, and savings numbers quoted by
the vendor. “Air-traffic control for hospitals” has been pitched before; the three customer stats are real
but early, and whether it generalises past a handful of friendly systems is the open question a Series A doesn't
answer.
For you
This is the cleanest live proof of two things the brief keeps telling you. One: the “operating system for the
physical world” thesis is funded and real — boring coordination work in a labour-heavy industry is where
the money now goes. Two: a student founded it, via YC, on access rather than capital. The replicable move
isn't healthcare — it's finding the labour-coordination mess in an industry you already stand inside, and
building the board that sees all of it.
Founder war story · design as the product, conviction as the fuel
She sold her house and cashed out her 401(k) to build a styling tool for hair the entire industry ignored
Richualist (product: The Mint) · founder/CEO Dawn Myers (former law student) · raised
~$1M VC (Black Star Fund, Techstars, NextFab, Revitalize & others) plus a Shark Tank deal
($150K/15% + 5% advisory, Mark Cuban & Emma Grede) · a 2026
TechCrunch
Disrupt Battlefield 200 company ·
AfroTech
What they do
The Mint is a handheld tool that warms liquid hair product to 130–140°F and, through six interchangeable
attachments, detangles and styles textured hair (type 3A–4C) far faster than by hand. It's a hardware answer to
a market every incumbent styling-tool maker simply skipped — curly and coily hair — because the existing
tools were designed around straight hair and never re-examined.
Where the idea & the design came from
As a law student, Myers cut off her chemically relaxed hair and ran straight into the problem: nothing on the market
was built for the hair she now had. The product is the design reasoning you care about made literal — why
is the attachment shaped this way, why this temperature, why warm the product rather than the hair — each
choice falling out of a use case the industry never studied. The unfair access was being the ignored customer.
The founder
Myers funded early prototypes by selling her home and liquidating her 401(k) to make
payroll — then raised ~$1M and won over Mark Cuban and Emma Grede on camera. That is conviction priced in
personal risk, and it's the persuasion craft you keep circling: she didn't sell a spec sheet, she sold a felt problem
nobody in the room had ever had to live with. Now she's running the Battlefield 200 gauntlet at Disrupt.
The bear case
Consumer beauty hardware is savage: one hero SKU, thin margins against cheap imitators, retail distribution that eats
startups, and a single-product company's fragility. ~$1M and a Shark Tank cheque is not much runway for a physical
product with tooling and inventory costs. Celebrity-investor glow fades fast if the reorder rate doesn't hold.
For you
Two things to steal. First, the Battlefield 200 itself is a door — TechCrunch Disrupt's startup
showcase is a real, applyable stage for a shipped thing, and your tiltrotor build log is exactly the kind of proof it
rewards (filed in §5b). Second, the design lesson: the sharpest product ideas come from being the underserved
user and refusing to accept “that's just how the tool is.” Myers found a huge market hiding behind an
unexamined default. Your job this phase is to notice the ones hiding near you.
§2
Money Moved
13 funding items screened · 6 filed
Who
Deal
What & why it's here
Src
Decart
~$7B
Amazon in talks to acquire (see §1) — after Anthropic weighed ~$6B and walked, and Nvidia walked too. The number isn't the story; the bidding war for real-time video and a chip-portability layer is. When three of the most sophisticated buyers alive circle the same company, that's the market telling you where it thinks the next platform sits.
Not one round — PitchBook's tally for the sector so far this year, as developers put cameras in homes, offices and factories to gather the training data robots lack. This is the single biggest signal on the page: capital has decided the next frontier is the physical world, and it's spending on data collection to get there. Read it next to §1 Vitalize and the §0-amber.
Series A, Oak HC/FT leading, with YC (see §1). Filed because the buyers are the signal: a healthcare-focused fund plus a hospital system (MemorialCare) putting money into staffing software is a demand-side bet you can't fake. The physical-world-OS thesis, priced.
Bain's data-centre arm in talks to upsize an existing loan from $2.8B to ~$6.6B. Debt, not equity — which is the tell: compute buildout is now financed like infrastructure (roads, pipelines), not like startups. The power-and-land bottleneck on the §0 board, in a credit facility.
Crypto custodian Anchorage buys B2B payments-automation startup Routable (sum undisclosed). A consolidation tell: the “agents in regulated money” vertical (see §5's PayPal hackathon) is being assembled by incumbents, not just funded.
Apple licenses tech and hires the team from personalized-podcast startup Huxe — a classic acqui-hire (terms undisclosed). Filed as a pattern you'll see again: when a giant can't build taste in-house, it buys the small team that had it. The reverse of the Decart story, same mechanism.
The read on the whole table: four of six entries are about the physical and infrastructural — robots,
data-centre debt, real-time video, hospital labour — not models. The classic startup funding feeds were genuinely
quiet and uninstructive today (Sifted ran trend pieces on European growth-stage queues; Inc42
and YourStory ran festive-season e-commerce and IPO-tracker items, no instructive round). So this board is built from the
acquisition and sector-tally wire instead — which is where the real money actually moved.
Handmer reports prototyping automated peer review: current models give human-level review for ~$0.10 per paper, and
for ~$5 you can rederive every result, check every statistical inference and audit every citation — against a
corpus of ~100M papers. Tyler Cowen (who runs Emergent Ventures, i.e. a grant you can apply to) posts it. Read it as
a near-complete startup brief sitting in plain sight: the scientific-literature audit layer, costed, from one of the
sharpest builders alive.
The most decorated living mathematician's own slides on how AI changes the practice of math — and
Quanta's companion, “Is AI the
End of Math As We Know It?” This is the non-hype, primary-source version of the §0 crypto-agility fear
(AI-accelerated math coming for the hard problems cryptography rests on). Read Tao to calibrate how fast the ground is
actually moving, from the person best placed to know.
A model can be constructed to pass every capability test you give it while secretly retaining a capability you're
trying to rule out — proven, not speculated. This is the load-bearing fact under the whole “trust is the
product” week: if you can't trust what a model tells you it can do, the only thing left to trust is a
record of what it actually did. That's the §0-green problem's reason to exist, stated as a theorem.
Andrej Karpathy argues the spread between what the best and worst users get out of the same model is widening fast
— the tool is identical; the skill of wielding it is not. The actionable read for you: “knows how to drive
an agent” is becoming a real, compounding, nearly-free edge, and you're in the window where practice buys
disproportionate return. Treat agent fluency as a trainable skill, not a given.
MIT built a motion planner with a provable no-collision guarantee for flying robots. This is the research your build
log's “hold position → fly a path” rung will stand on — and “provably safe” is the
vocabulary you'll hire against when the autonomy problem gets real. Deep tech arriving as a buildable next step, not a
lecture.
Lightricks' open-weights video model, one of the week's most-downloaded. While Amazon pays ~$7B for Decart's real-time video (§1), the open tier you can run yourself keeps closing the gap — state-of-the-art generative video, no per-clip cloud bill. For the filmmaker in you this is the cheapest moment to learn the craft before it's a paid product. Today's §8.
Qwen3.8-27BHugging Face · 17.5k likes · 6.8M downloads
Still the top trending open-weight model by a wide margin. The cheap, open tier a bootstrapped builder actually lives in keeps getting better — this is the engine under half the “I built it on the free tier” stories. File as a component, not a headline.
A sharp argument that passkeys, sold as the password-killer, quietly lock you into the platform that stores them — convenience bought with vendor lock-in. Read it next to the §0 “proof-of-human” thread: identity is contested ground, and every proposed fix carries a hidden owner. Good training in reading who benefits from a “standard.”
The day's top HN link: a beloved 1997 real-time strategy game faithfully rebuilt on a modern engine. Here for the craft, not the nostalgia — the write-up is a clinic in reverse-engineering why the original felt the way it did, and rebuilding that feel rather than just the features. The same discipline you bring to watching how a film was made.
Cloudflare shipped a small open model (and a clef-flash variant) — the “small, cheap, run-it-yourself” tier from an infrastructure company rather than a frontier lab. Another data point that the useful layer for a solo builder is getting commoditised and free, fast.
NYC now requires that cancelling a subscription be as easy as signing up. Here as a business-model signal: a whole genre of “growth” built on cancellation friction is being legislated out, city by city. If you ever build subscriptions, design for easy exit now — the dark-pattern era is closing, and the regulation front-runs the product.
§5
The Arena
18 live · deadlines computed for today
Devpost returned in full this run — fresh board, soonest first, red under 7 days. Prize figures are carried from
each event's listing where published; where a purse isn't stated I've said so rather than guess.
Hack Apertus
$10,680 · open-source AI model focus — rides the open-weight wave you live in (see §4) · closes Oct 16
4d
MLAI Hackathon 2026
purse not stated · general ML/AI build · closes Oct 16
4d
SVYCA Youth for Earth (Y4E)
$2,700 · climate — small field (16 registered), winnable · closes Oct 18
6d
Hyperbloom October — UI/UX & Web Design
$9,117 · pure design / taste play — squarely your eye · closes Oct 19
$25,000 · agentic AI on a taste/culture API — your design eye meets agents · closes Oct 30
18d
YouCam API Skin AI & eCommerce VTO Hackathon
$6,000 · computer vision + virtual try-on — literally Decart's Lucy use case · closes Nov 2
21d
PayPal AI Hackathon
$67,500 · agents in regulated payments — the money vertical (see §2 Anchorage) · closes Nov 12
31d
Meta VR Start Developer Competition 2026
$1,000,000 · VR/XR — furthest out, biggest purse, most specialised · closes Nov 18
37d
The ones that fit you with real runway: OpenCV AI (15d) is vision + IoT — point a camera
at a drone payload and you have an entry; Nebius × NVIDIA (18d, $50K) is the beginner-friendly
workhorse that leaves dated proof-of-work behind; AWS CDS Agentic (16d, $40K) is a direct excuse to build
the §0-green audit-the-agent tool and get it judged; and Hyperbloom (7d, design) is a pure taste
play if you want a fast one. Hack Apertus (4d) is this-week-or-never on the open-weight theme.
Register for one today — it converts intention into a deadline.
§5b
Open Doors
TechCrunch Disrupt Battlefield 200 — the stage from today's §1
Dawn Myers' Richualist is in this year's cohort. It's a real, applyable showcase for a shipped thing — not a pitch deck, a working product in front of press and investors. Your tiltrotor build log is exactly the kind of tangible proof it rewards. Watch for the next open call and put the drone in front of it.
watch
Thiel Fellowship — applications open
$200K, no equity, under 23, US-friendly. Built for a shipped-proof-of-work profile, not a transcript — demonstrably the door shaped like you.
open
Emergent Ventures — rolling
No equity, funds young unproven people. Tyler Cowen runs it — and today's §3 Casey Handmer post is him showing you, in public, what he thinks the open problems are. Reading Marginal Revolution is reading the grader.
open
Y Combinator — the student path, proven today
§1's Vitalize was founded by a medical student and shaped in YC's W23 batch. YC funds students who've stood inside a problem. The RFS is the free target list; a batch is the door. You are exactly the profile it says it wants.
open
§6
Skill
day 14
Log every tool call your agent makes for one week — and you're standing inside today's §0-green problem, holding the prototype
~1–2 hours to set up · free · produces: a running ledger of everything your coding agent touched this week, and a felt map of exactly where “prove what the agent did” is still unbuilt
Today's §0 green is “nobody can prove what an AI agent actually touched,” and the research path says the
log is the prototype. So build the log. You already run Claude Code (or similar) daily on a machine holding your
keys — that is the exact environment the problem lives in. Wrap your agent so that every tool call it makes —
every file read, every shell command, every network request — is appended to a timestamped record before it runs:
what was called, with what arguments, under whose credentials. The cheapest version is a shell wrapper or a few hook
lines; the slightly better version emits structured events you can query. Claude Code's own
hooks give you a PreToolUse handler that is
precisely this seam.
Then, at the end of the week, read it back. Two things happen. First you'll be startled — the agent touched more,
and reached further, than you remembered authorising; that surprise is the entire market. Second, you now hold the thing
NVIDIA and Cisco don't ship: not a sandbox that contains the agent, but a receipt that proves what it
did. The distance between “I logged my own agent” and “a provenance layer any team can drop in” is
the whole §0-green startup. Build the ugly single-user version first; the product is that, with the rough edges filed
off.
Free resources: Claude Code Hooks (the
PreToolUse seam, no extra tooling) and OpenTelemetry if you want
the structured-events version that already speaks the language every observability tool reads. Start with the one-file
log; reach for OTel only when the one-file log annoys you.
§7
Watch
canon · YouTube talks all 404'd this run
Objectified — 75 minutes on why every object around you is shaped the way it is
The canon · design process track · documentary by Gary Hustwit, 75m · hustwit.com
Every YouTube feed 404'd or 500'd again this run — no fresh talk — so the canon, and this is the pick for a
day led by a hardware founder (§1 Dawn Myers) who re-examined why a styling tool is shaped the way it is.
Objectified is Gary Hustwit's film about industrial design: the
reasoning behind why a toothbrush, a chair, a laptop hinge ends up with exactly its form — the decision, not the
product, which is the thing you told me pulls you hardest. It's Jony Ive before the mythology, and a dozen other designers
explaining, out loud, the choices nobody normally narrates. Filmmaking-grade, so it goes down like a story, not a lecture
— which is the whole bar for this section.
Why now: today's §1 is full of design-as-moat — Decart treating video as a latency problem, Myers treating a
hair tool as an unexamined default worth re-deriving from scratch. Objectified is the 75-minute education in seeing
objects that way: nothing is arbitrary, every curve is a decision someone made and could have made differently. The
design-process track last surfaced three weeks ago (the Starbase tours); this is its overdue return, and the shortest,
densest thing on it.
§W
The Wildcard
7 strange things, off the tech wire
The exploration arm — the off-wire feeds all came back today, so this is the real thing: alchemy, error-correcting
knitting, a civilisational collapse, a drone war you didn't know was about cameras, and the hidden logic of the wall
socket. None of it is curated toward what you already like; that's the point. One line on why it caught my eye.
A 17th-century alchemy manual told entirely in pictures — no words, because the knowledge was meant to be hidden in plain sight. A gorgeous, strange reminder that “documentation as cryptic images” is older than you think, and that some information was designed to be unreadable on purpose.
Error-correcting codes — the math that keeps your data intact — rendered as an actual knitting stitch, so a dropped loop can't unravel the whole piece. Theory and craft touching directly; the same idea that protects a hard drive, worked in yarn.
A whole interconnected world of advanced civilisations fell apart within a generation ~1200 BC — and the best theory is cascading failure across a system that had grown too interdependent to absorb a shock. A sobering mirror for anyone building on top of fragile global supply chains today.
With DJI being pushed out of the US, a Chinese rival is sprinting to capture the drone-and-camera creator market before the door shuts. Directly adjacent to your build — and a live lesson in how geopolitics, not product quality, is redrawing who gets to sell flying cameras in America.
A field guide to every socket shape on American walls and the century of standards fights, safety retrofits and legacy compromises frozen into each one. The invisible design history in the most ignored object in every room — exactly the “why is it shaped like that” instinct, pointed at a wall.
New analysis suggests the most-studied tomb in the world may still be hiding a chamber — and a queen written out of history. A hundred years of excavation and the famous thing still isn't fully mapped; a nice antidote to the feeling that everything's already been found.
The humble brick as a 5,000-year-old design decision — why its size, weight and shape are tuned to the span of a human hand, and how that one constraint quietly shaped every city. Design history as narrative audio, which is exactly the way you said you like depth delivered.
§8
Deals
1 free resource worth your time
Run state-of-the-art video generation on your own machine, for $0 — the open answer to today's $7B deal
Free · open weights · needs a decent GPU (or a free Colab / cloud trial)
The genuinely useful free thing today sits right under the §1 headline: while Amazon pays ~$7B for Decart's real-time
video, Lightricks' LTX-2.5 is open-weights, downloadable, and one
of the most-pulled models on Hugging Face this week. For the filmmaker in you that means you can learn generative video
— how to direct it, where it breaks, what it's actually good for — with no per-clip cloud bill and no waitlist.
The economics of making things look good are going local and free the way photo editing did; it is cheaper to
get fluent now than to discover the skill is table stakes a year from now.
Honest note: §8 still has no standing deals source of its own — it needs the GitHub Student Pack, cloud/AI
credit programmes and startup-tool offers wired into collection before it earns a daily slot. Today it's a pointer to the
single most useful free thing on the page, which beats a coupon.
§9
Also
filed, not featured
Nadella says AI models need an “emergency brake.” Microsoft's CEO wrote it's time “to step back and assess the trust architecture” of AI — the continuation of yesterday's “assume every model is compromised.” The man selling the most AI keeps saying the quiet part: trust is now the product, which is the whole reason today's §0-green (prove what the agent did) exists.
Palantir CTO Shyam Sankar on AI targeting, the DOD dropping Anthropic, and AI liability. A full CBS transcript from the operator's chair of the defense-AI debate — the Iran strike, who's accountable when an AI helps target, effective altruism. File next to the §0 “American defense” thread: this is what the inside of that argument actually sounds like, unfiltered.
Beijing's Neolix runs the world's largest robovan fleet — 27,000 vehicles. As Shenzhen tests nighttime driverless parcel delivery, China's physical-AI buildout is already operational at a scale US coverage rarely conveys. Read next to §2's $48B physical-AI tally: the training-data land-grab the money is chasing is partly already deployed, just not here.
Ledger says the $86M+ theft from hundreds of its hardware wallets involved compromised devices. A hardware-wallet maker — the thing you buy specifically so your keys live offline — traces a large theft to the devices themselves. A sharp reminder for the §0-green builder: “self-custody” and “you own it” are only as strong as the layer underneath, and that layer is where the real attacks land.
Silicon Valley is falling in line with Trump's push to rename AI. A faintly absurd story — Musk and Benioff pledging to rename SpaceXAI and “AIForce,” Nadella and Nvidia adopting new terminology — but file it as a power read: naming is downstream of who holds leverage, and watching who complies, and how fast, tells you where the real pressure is pointing.