Issue 012Fri 9 Oct 2026399 items screened → 53 filedfull pull · 0 feeds down
OpenAI dumped 722
machine-written math manuscripts on the world this week — then quietly withdrew three of them for a sign error,
with roughly half the rest still unverified. The lesson isn't that the machine is a genius or a fraud; it's that
volume is now free and verification is the bottleneck.
Watch where the money went while the proofs piled up: not into the abstract middle but into things you can check and
touch — $5B of debt into Waymo's robotaxis, $60M into the company that
records humans so robots can copy them, a speech model shrunk to 16.9MB so it runs on a
microcontroller, and $10M to a 19-year-old building for the one user he understands completely:
himself. When the abstract layer floods, value runs to what's verifiable, physical, and close to the person using it.
§0
The Problem Board
2 of 11
Every accelerator says it funds people “solving hard problems”, and not one of them tells you which. This
does — sourced, graded honestly, with somewhere to start reading. Today one pure-software problem you can
prototype in a weekend, now that agents write to your repo — and one you cannot start at all, but which this
week's 722-proof flood just made sharper.
START THIS MONTHgreen · solo · no capital · no credentials
API providers announce their breaking changes instead of just applying them
The problem
When an API changes, every customer hand-edits their own codebase to catch up — the
same migration, done thousands of times, badly, in parallel. The provider knows exactly what changed and how to
fix it, but ships a changelog and a deprecation warning instead of the fix itself. The missing layer is the one
that lets a provider push a code change into its customers' repos, not just a notice.
Who says it's hard
Y Combinator's Requests for Startups,
from a founder who worked with 50+ API vendors and watched every one of them announce changes they could have
propagated.
Why it's still open
Two years ago nobody would grant an outside tool write access to their repository.
Claude Code, Devin and Greptile normalised exactly that — agents proposing commits against a stranger's
codebase is now ordinary. The permission barrier dropped recently enough that the obvious product on top of it
hasn't been built.
Who's already trying
Greptile and Ellipsis (code review), Renovate (dependency bumps, not API
semantics). Nobody connects a provider's breaking change to an automatic, reviewable PR in each customer's repo.
And the write-access plumbing keeps spreading — Docker's
agent tooling and the week's small agent models (§4) are all bricks in the same wall.
Your angle
Pure software, zero capital, and you already understand agentic code editing
from the inside — you are a Claude Code user who's felt a dependency break your build. Narrow enough
to prototype in a weekend against a single noisy API. That's exactly this issue's §6 skill.
Research path
Pick one API with frequent breaking changes (Stripe, Twilio) and
read two years of its changelog — feel how much of it is mechanical.
Write the migration codemod for
a single breaking change by hand (today's §6). You'll understand the real shape of the work in an
hour.
Read Renovate's architecture docs on how an automated PR against someone else's repo actually earns
enough trust to get merged — that trust, not the diff, is the hard part.
NEEDS A LAB, STANDARDS AND A DECADEred · filed so you recognise it, not so you start it
AI-accelerated math is coming for the encryption the internet runs on
The problem
The public-key cryptography protecting banking, messaging and identity rests on a bet that
certain math problems stay hard. This week made the bet look less safe: OpenAI released
722
manuscripts of machine-generated proofs across geometry, algebra and computer science. The deeper problem
underneath is that almost no system in production can swap its cryptographic primitives quickly
(“crypto-agility”), so a break wouldn't be a patch — it would be a multi-year migration nobody
has rehearsed.
It needs frontier cryptography, standards bodies and coordinated global migration at
once. Post-quantum standards exist on paper; getting the world to adopt crypto-agility before a break
rather than after is a decade-long coordination problem, not a product.
Who's already trying
NIST (post-quantum standards), Cloudflare, Google, Apple, PQShield. All large, all
slow, all working on the “answer” while adoption lags years behind.
Your angle
Red, and honestly so — a standards-and-PhD problem, not a weekend build.
Filed so you recognise it, not so you start it. The one cheap, boring slice that's just software: tooling that
inventories which cryptographic primitives a codebase actually depends on, the unglamorous first step of
any migration. Learn the vocabulary now; don't found here.
Research path
Read Aaronson's “The
Mathocalypse” for the honest state of the threat, hype stripped out.
Skim NIST's post-quantum
standards to see what “solved” already looks like on paper.
Notice the cheap adjacent slice:
crypto-agility tooling — “what am I even using, and where” — is software even when the
cryptography isn't.
New on the feed today: YC's RFS is unchanged since the Fall 2026
publish. DARPA posted three items — four
more teams into the Quantum Benchmarking final, a CIDAR
passive-ranging challenge, and the Vertical
Lift Challenger recap — all of which sit under the board's existing american-defense and compute
entries rather than opening a new one. Marginal Revolution ran essays, not problems. No genuinely new hard problem to
add today; the crypto-agility entry filed yesterday just got a louder week.
§1
The Big Three
researched
Founder war story · a 19-year-old who already did it once
Cal AI's teenage founder sold a $30M-a-year app, then raised $10M to collapse your phone into one wrist-worn assistant
Persona · founded 2026 by Zach Yadegari (19) and co-founders · $10M led by Vine
Ventures (w/ Z Fellows' Cory Levy, Collective Global) · valuation undisclosed · $179
wearable band ships December; free iMessage beta live; “a few thousand” beta users, five-figure preorders ·
TechCrunch ·
Runtime
What they do
Persona is a personal AI assistant with a hardware shell: a $179 band you tap (or trigger with a
wrist-flick) to speak to an agent that books rides, orders food, handles email and plans travel. Deliberately
not ambient — it isn't always listening, which is the pointed design choice against the always-on
Humane/Bee generation. The thesis, in Yadegari's words: “all of the utility apps you use day-to-day will
collapse into a single super app powered by an AI assistant.”
The founder
This is the part worth your attention. Yadegari built his first business in ninth grade and sold a gaming
startup for $100,000 at 16. Then he co-founded Cal AI, a calorie-tracking app that hit
>$30M in annual revenue in under two years and got acquired by MyFitnessPal in March 2026. He
did the acqui-hire, stayed through June, then left to start Persona. He is 19 and on his third company.
Where the idea came from
Same move as Cal AI: build for a user he is. Cal AI worked because he was a young person who wanted to
track food without friction; Persona is a young person who wants his phone's busywork to disappear. The unfair
advantage isn't a technology — it's that he ships consumer apps fast for a demographic he belongs to and
reads instinctively, then lets the app-store ranking tell him if he's right. He's done that twice.
The bear case
Hardware is a graveyard — Humane and the first wave of AI pins died, and a few thousand beta users plus
five-figure preorders is a rounding error. Meta (Muse), Amazon (which bought Bee) and Instinct have distribution
he can't match. The “super app” thesis has been attempted and failed in the West for a decade, and a
button-triggered band is a harder sell than “it's just on your phone.” Software, not the band, is
probably the real product.
For you
Forget the wristband; steal the pattern. A 19-year-old with no network and no capital built a $30M business by
shipping fast for himself and reading the app-store scoreboard honestly — the single most replicable founder
story in this brief, and close to your own position. Note the exact mechanic: pick a user you already are,
ship something ugly, let the ranking tell you the truth. His gaming-startup exit at 16 is the version of
“proof of work” that no transcript can give you — and it's the thing that made VCs return his
call at 19.
Buildable · AI small enough to live on your drone
A speech-recognition model in 16.9 megabytes — smaller than a photo, running on a microcontroller with no internet
Whistle, by Cactus Compute (YC S25) · San Francisco, 8 people · founders Henry Ndubuaku
(turned down an Nvidia research role) & Roman Shemet (ex-quant) · open weights on Hugging Face,
code on GitHub ·
Cactus blog ·
HN (625 pts) ·
GitHub
What they do
Whistle is a speech-to-text model that fits in a single 16.9MB file and runs on a CPU with no
dependencies — phones, wearables, cars, smart-home gadgets, and microcontrollers. Seven languages with
auto-detection, word-level timestamps, all on-device. No server, no API bill, no network round-trip.
The tech — and what's hard
The hard part is doing it small without going deaf. Whistle's encoder is eight “Simple Attention”
blocks with a Monarch–Hadamard MLP; the whole thing hits 11.1ms to first token on an Apple
M4 Pro CPU and decodes 1,319 tokens/second. The headline claim: it beats OpenAI's Whisper Base
(145.3MB — 8.6× larger) on word-error rate for LibriSpeech, SPGISpeech and
Earnings-22, and loses to it only on a couple of noisier conversational sets. Roughly a tenth the size, better on
clean speech.
The founders
Cactus is eight people out of YC's Summer 2025 batch. Ndubuaku walked away from an Nvidia research position to
build an on-device inference engine; Shemet is a former quant and economist. They met through YC co-founder
matching in London. The whole company is a bet that the interesting frontier isn't bigger models in data centres
— it's squeezing useful ones onto the chips already in your pocket.
The bear case
Tiny models have a ceiling — strong on clean, read speech, weaker on cross-talk and accents, which is most of
the real world. Apple and Google ship their own on-device speech for free inside the OS, and open weights are hard
to turn into a business. “Better than a two-year-old Whisper Base” is a real but modest bar.
For you
This is directly load-bearing for the tiltrotor. On-board speech or audio inference with no radio link,
in 17MB, on the kind of microcontroller a drone already carries, is a capability you can actually wire in this
month — and it's open, so you can. Beyond the drone, study the design move you keep saying you love:
every architectural choice here exists to hit a size budget. That's the “why is this part shaped this way”
instinct applied to a neural net. And Cactus is the shape of company you could be in two years — eight people,
one sharp constraint, shipping.
The robotics moat · four founders with no robotics background
Sequoia put $60M into a company that pays strangers to film themselves making coffee — the training data robots don't have
Mecka AI · founded 2024 by Josh Gao, Jason Chong, Mogen Cheng & Duy Nguyen
(none from robotics) · $60M Series B led by Sequoia, w/ Nvidia and Microsoft's M12 ·
$500M valuation · 2 years old ·
TechCrunch ·
TechCrunch (Sept)
What they do
Mecka (from “mecha”) pays ordinary people to record themselves doing everyday physical tasks —
making coffee, fixing a car — wearing body sensors and holding a phone. That motion data is the fuel for
training humanoid and other robots. The explicit pitch: be Scale AI for robotics — the
human-labelled-data layer, but for bodies in space instead of text.
Why it works (if it works)
Language models had the internet to train on; robots have almost nothing — there is no web-scale corpus of
“how a human hand actually opens a jar.” Whoever manufactures that corpus owns an input every robotics
lab needs and none can scrape. Sequoia, Nvidia and Microsoft all writing cheques says the data bottleneck is
believed to be real and worth owning upstream, the way Scale sat upstream of the whole LLM wave.
The founders
The tell: none of the four come from robotics. Gao and Cheng (both Canadian) built a restaurant-fintech
startup; Chong sold a crypto exchange to Coinbase and worked there; Nguyen runs operations. They didn't win on
robotics expertise — they won by recognising that the valuable, unglamorous job in any AI boom is building
the data pipeline, and moving on it first. That's a pattern, not a fluke (it's literally the Scale AI story).
The bear case
Data-labelling is a margin-thin, commoditising business — Scale's own moat is under pressure. Competitors
are already crowding in: XDOF is reportedly raising at a $1.2B valuation, Micro1 at $500M, and
Scale itself is expanding into robotics. It's unproven that human motion capture transfers cleanly to a robot
with different joints and limits. And big robotics labs may simply collect their own.
For you
Two things to pocket. First, a door: this is a real way for ordinary people to earn by recording data — worth
understanding the economics of, because the “humans generate the training data” market is going to be
large and is hiring. Second, the mental model, which is the valuable half: in every AI wave, the people who
owned the data pipeline made more durable money than the people who built the models — and they usually
weren't domain experts, just first. Ask what the equivalent un-scraped dataset is in a corner you can actually see.
§2
Money Moved
41 funding items screened · 6 filed
Who
Deal
What & why it's here
Src
Waymo
$5B
A loan from Blackstone and PIMCO — Alphabet's robotaxi arm's first-ever debt financing, for fleet expansion. The signal isn't the size; it's the form. You raise equity for a bet and debt for a machine that reliably prints cash. Lenders pricing Waymo as infrastructure means autonomy crossed from story to utility.
The company behind the Manus agent raised its first round since splitting with Meta (Boyu, IDG; Tencent and HSG back in). The agent wave is funded at nine figures — which is exactly why the unowned control layer under it matters more, not less.
The crowd-vote AI leaderboard nearly doubled to a $3.1B valuation in 10 months (Lightspeed, Khosla) — and is now scoring models on lying and alignment, not just capability. Whoever defines the benchmark quietly steers the whole field; worth watching what they choose to measure.
Sequoia (w/ Nvidia, M12) for human-motion data to train robots, at a $500M valuation (see §1). The “Scale AI for robotics” bet: own the data pipeline, not the model.
Conversion Capital led a round for software that puts swarms of agents into mortgage origination. The pattern to file: agents are landing in slow, paperwork-heavy, regulated verticals first — the §0 physical/compliance world, where the incumbents are weakest.
Vine Ventures for Cal AI founder Zach Yadegari's wrist-worn personal assistant (see §1). A 19-year-old's third company — the raise is a note; the founder is the story.
Read the cheques against the backdrop: OpenAI's annualised revenue is reportedly
$20B
lower than previously projected — a reminder that even the center of the AI story is being re-priced. Money
this week concentrated into things with a physical or verifiable backstop (robotaxis, robot data, on-device models,
agents aimed at regulated work) and away from undifferentiated model hype. Also noted, not rowed: China's
Biren and others keep raising for domestic compute regardless of price — the constraint there is
strategic, not financial.
The story moved this week from “can AI do math” to “AI can do a lot of math, and nobody
can check it fast enough.” OpenAI released 722 manuscripts across 372 problems, roughly
half unverified, then withdrew three for a sign error. Mathematicians called it a “slop drop”;
Terence Tao is asking what to
even teach students now. Why it changes what you read next: volume is free, verification is the
bottleneck — the defining shape of every field AI touches, starting with the cryptography in today's
§0 red. This is the six-month weather report.
A rumour that an AI had cracked the Unique Games Conjecture sent three computer scientists sprinting to publish
their own result first. It's the most human version of the math story — what it actually feels like to work
next to a machine that might finish your life's problem before you do, and why some of them chose to race rather
than wait. File it as the emotional ground truth under the hype.
MIT's “SANDO” gives a flying robot a provable guarantee it won't hit anything while navigating an
unknown space — built for disaster-response drones. This is your tiltrotor's exact next problem (autonomy
past “hold position”), explained by the people who formalised it. Read it to get literate in how
guaranteed-safe path planning is actually posed, because that vocabulary is what you'll hire against.
Unchanged since Fall 2026 and still the highest-authority idea source there is. Today's §0 green
(“API change propagation”) lives on it word for word. Free, and a map of where the smart money
wants founders to go — including the defense category a cabinet secretary signed.
An argument that a cheap, fast Chinese open model quietly matched the frontier on the things most apps actually need — and the West shrugged. Read it for the cost lesson that governs you: the floor price of “good enough” intelligence keeps dropping, which is what makes the agent tooling in today's §0 and §6 cheap to run at all.
A hand-drawn illustration of a house, wired up as a live Home Assistant control surface — lights and sensors mapped onto art instead of a grid of toggles. Here for the taste: the best interface wasn't a UI kit, it was a commissioned drawing. The “why is this the wrong shape” instinct, applied to your own home.
DuckDB's take on an open “lakehouse” table format — serious data infrastructure you can run on a laptop. Boring-but-load-bearing: the kind of free, local building block that lets a solo founder handle real data without an AWS bill. File it next to the $5 VPS thread as part of the cheap-infra toolkit.
Mitchell Hashimoto (Ghostty, HashiCorp) proposes a tiny standard so any CLI can report progress/status to any terminal. Worth reading not for the spec but for the craft: watch a serious engineer propose a shared convention — the social work of getting an ecosystem to agree is the same skill as the §0-green “how does an automated PR earn trust.”
Another strong open model with a million-token context window shows up on OpenRouter. Pair it with the DeepSeek piece above: the open, cheap, long-context tier is getting genuinely good, which is the tier a bootstrapped builder actually lives in. File as a component, not a headline.
An open photo editor that runs its AI entirely on your machine — no subscription, no cloud. For the filmmaker in you: the editing stack is quietly going local and free, which changes the economics of making things look good. Worth a look even if you keep Lightroom.
§5
The Arena
25 live · deadlines pulled from Devpost today
Day-counts are each event's live submission deadline, soonest first; red under 7 days. The six in red only make sense
if you can ship this weekend — but two of them (cybersecurity+design, founder-framed) fit you unusually well.
UnivaBio
$40,000 in prizes · beginner-friendly, health/bio — closes Oct 13
3d
Multimodal AI Hackathon 2026
$1,176 · ML/AI, multimodal — closes Oct 14
4d
WarriorHacks 2.0
beginner-friendly, open-ended AI — closes Oct 14
4d
INDUX 5.0-2026
$2,647 · cybersecurity + design — the §0 trust theme as a build brief, and your design eye · closes Oct 15
5d
Next Founders Hackathon
$275 · founder-framed — build the thing you'd actually start · closes Oct 15
5d
Hack Apertus
$10,680 · open-source AI model focus — rides the open-weight wave (see §4) · closes Oct 16
6d
Hyperbloom October — UI/UX & Web Design
$9,117 · design — a pure taste play · closes Oct 19
9d
Build, Ship, Shape: Amazon Developer Hackathon
$138,000 · ML/AI, productivity — biggest general-purpose purse on the board · closes Oct 23
13d
OpenCV AI Competition 2026 (AWS)
$20,250 · IoT + vision — the closest thing on the board to your drone · closes Oct 27
17d
AWS CDS Agentic AI Partner Hackathon
$40,000 · agentic AI — build the §0 API-migration idea and get judged on it · closes Oct 28
18d
HackFW: MADE Challenge
$1,900 · IoT + robotics — small purse, dead-on your hardware · closes Oct 30
The ones that fit you with real runway: OpenCV AI (17d) is IoT + vision — point a camera
at a drone payload and you have an entry; HackFW MADE (20d) is IoT + robotics, small money but exactly
your hardware; AWS Agentic (18d) lets you build the §0 API-migration idea and be judged on it; and
Nebius × NVIDIA (20d, $50K) is the beginner-friendly workhorse that leaves dated proof-of-work
behind. If you want one this week, INDUX (5d) is cybersecurity + design — squarely your eye.
Register for one today — it converts intention into a deadline.
§5b
Open Doors
US internships — and your edge just got sharper this week
The US moved to suspend green-card applications from Microsoft, Infosys, TCS and other large employers over process abuse. That's your peers' path getting harder — and a reminder that you are US-eligible without sponsorship, which sidesteps the entire fight. Early-to-mid October is peak cycle; a defense-tech, robotics or AI-infra role doubles as a §0 on-ramp.
now
Thiel Fellowship — applications open
$200K, no equity, under 23, US-friendly. Zach Yadegari (§1) is exactly the profile it funds — a teenage repeat founder with shipped revenue, not a transcript. Demonstrably the door built for your shape.
open
Emergent Ventures — rolling
No equity, funds young unproven people. Tyler Cowen runs it — reading Marginal Revolution is reading the grader. Lowest-friction real door on the list.
open
YC Requests for Startups — the free map
Not an application, a target list. Today's §0 green is lifted straight from it. Build one of its named problems and you've pre-aligned with the highest-signal funder there is.
open
§6
Skill
day 11
Write a codemod — automate one breaking API change by hand, and you've built the §0-green prototype
~1–2 hours · free · produces: a script that rewrites old API calls to new across a whole codebase, and a felt understanding of the problem nobody's productised
Today's §0 green is “API providers announce breaking changes instead of applying them.” The cheapest
way to understand it is to do one migration the hard way. Pick an API you know has changed — Stripe and
Twilio both have long, breaking changelogs — and choose a single change: a renamed field, a moved parameter, a
deprecated method. Now write a codemod: a small program that parses source code into a syntax tree,
finds every place the old pattern appears, and rewrites it to the new one automatically. ast-grep
is the gentlest way in — you write a pattern that looks almost like the code itself, and it does the
find-and-rewrite across every file. (The heavier, JavaScript-native tool is jscodeshift.)
Two things happen. First, you feel the shape of the problem: most breaking changes are mechanical, the
provider already knows the exact transform, and yet every customer rediscovers it by hand. Second, you're holding the
primitive. A codemod that reliably upgrades a caller from v1 to v2 is literally the first feature of the product the
YC RFS is asking for — the thing Renovate does for dependency numbers but nobody does for API
semantics. Write it once for yourself; the gap between “my script” and “a provider ships
this to all its customers” is the whole startup.
Free resource: ast-grep — its playground lets you test a rewrite
pattern in the browser before you touch your repo, no install. Then read
Renovate's docs on how an automated PR against someone's codebase actually
earns a merge — because trust, not the diff, is the part that's hard.
§7
Watch
fresh · alternating off yesterday's canon
Adam Nash: “Making Great Products” — an operator who shipped at LinkedIn, Dropbox and Wealthfront on why most products are mediocre
Stanford eCorner · full talk (~50 min) · a builder's talk, not a lecture · YouTube
Yesterday ran canon (Barbarians at the Gate), so today a fresh talk. Adam Nash has shipped consumer products
at scale for twenty years — LinkedIn, Dropbox, Wealthfront, now Daffy — and this is his distilled argument
for why great products are rare and what separates them from the merely functional. It lands directly on the two
things this issue keeps circling: Persona (§1) winning because a founder built for a user he is, and the
“why is this the exact right shape” instinct you told me you care about more than almost anything. Nash is
unusually concrete — prioritisation, what “quality” actually means, why most teams optimise the wrong
thing — which is why it's a talk worth your time and not a lecture hall.
If you've only got two minutes, a16z's “What is the story of
your company?” is the same lesson compressed: a product that can't be explained as a story doesn't get
bought. Watch Nash for the craft; watch the short for the pitch — both are the persuasion skill you said you want
to train.
§W
The Wildcard
8 strange things, off the tech wire
The exploration arm. None of this is curated toward what you already like — that's the point. One line on why it
caught my eye, no justification of usefulness.
A months-old subreddit about what men ate for dinner quietly became a confessional for four million lonely people. A strange, specific window into male loneliness and why it found that door and not another.
An endangered moth that navigates by the stars reappeared in enormous numbers, hinting the species is clawing back from the brink. Insects that use the night sky as a compass, nearly lost, maybe returning.
A “whale necropolis” on the seafloor in the Diamantina Zone — the bones of giants piled where they fell, millions of years ago. The deep ocean keeps records we're only now learning to read.
Sixteenth-century Venice wanted to expel its Jews but couldn't run its economy without them — so it invented the world's first ghetto as a compromise. The word, and the idea, start on one island in 1516.
A paper arguing a lot of adult ADHD is really a broken body clock — and that treating the sleep timing, not just the attention, helps. A reframe that quietly rearranges a diagnosis millions carry.
A tiny website that picks a random average person from somewhere in human history and tells you what their life was actually like. A machine for puncturing the illusion that history is only kings and dates.
A close look at the messy, beautiful, wildly varied brick hiding on buildings all over Chicago. Design appreciation for the single most overlooked object you walk past a hundred times a day.
The domain of Google's dead social network — huge in Brazil and India, gone since 2014 — is live again with a note from its founder. Internet archaeology, and a small meditation on what platforms owe the communities they abandon.
§8
Deals
1 free resource worth your time
8,000+ free university courses, and the trick to auditing the paid ones for nothing
Free · Class Central
No coupon noise — just free education, curated. The one worth bookmarking is the
8,000+ OpenCourseWare courses from MIT, Stanford and
Yale that are simply free, paired with the mechanic underneath it:
how to audit Coursera courses without paying.
If you want something aimed at the next step after building, Class Central's
product-management roundup is the
closest free analogue to the “how do I turn a build into a product” question §7's Adam Nash talk is
about. For someone whose whole strategy is off-campus, branch-agnostic proof of work, pick the one course that
unblocks your next build — not a shelf of them.
Honest note: this section 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
free-course pointer, which beats a coupon.
§9
Also
filed, not featured
OpenAI's revenue is reportedly $20B below earlier projections. The number everyone extrapolated from turns out to have been optimistic. Filed as a corrective to read alongside every “$X billion AI market” claim — including the ones in §2. Even the center of the story is being re-priced, which is good news for anyone selling a real product cheaply rather than a projection.
CoreWeave is entering India, 240MW with AdaniConneX. The compute land-grab reached your backyard. Indian data-centre buildout is where the global AI-power story becomes local — local permitting, local power, local jobs — and it's adjacent to your degree's subject matter. Worth watching as the energy/subsurface thread's nearest live instance.
Not just AI — China is building tech across the whole stack. From batteries and solar to drones and satellites, a reminder that the US–China contest is far wider than models. A useful map if you're deciding which deep-tech corners are about to get geopolitically expensive — and which stay open to a small US-eligible builder.
Theranos.world — sit at Elizabeth Holmes' desk. An engineer turned 1,000+ trial exhibits — emails, slides, texts — into a website that lets you rifle through the Theranos founder's desk. Half art project, half the best free case study in how narrative outran reality. The inoculation against your own future pitch deck, playable in a browser.
Software developers are not okay. A widely-shared essay on the morale hit of the AI-coding transition — what it does to people whose craft is being automated and reshaped under them. Filed because the human side of the shift is a real thing to understand before you hire into it or build for it.