Issue 014Sun 11 Oct 2026401 items screened → 52 filedfull pull · 17 feeds down
The message from the top got blunt this week. Microsoft's Satya Nadella said
we should
now assume every AI model is “compromised” and called for an
“emergency
brake”; days after Anthropic admitted it can't reliably control its own agents and
cut
its evals off the live internet rather than keep watching.
The builders and the money answered the same way all week — move down, into the layers you can own, host and
check. A solo developer shipped Talorys, a personal AI agent that runs entirely inside your own
Cloudflare account; $2B went to Axiom for cloud-and-power infrastructure you can point
at; and fuel-cell and fusion startups raised to make the electricity the whole thing burns. When you can't trust the
black box, value flows to whatever you can hold in your hands — and the cheapest version of that move is one you
can make this weekend.
§0
The Problem Board
2 of 11
Every accelerator says it funds people “solving hard problems”, and not one tells you which. This does
— sourced, graded honestly, with somewhere to start reading. Today one pure-software problem you can start this
month, sitting right under today's §1 Talorys; and one DARPA just put money behind, graded red and filed
honestly — but with an on-ramp that is the exact math your drone needs the day GPS drops or lies.
START THIS MONTHgreen · solo · no capital · no credentials
Deploying personal software is still hard, and agents made it urgent
The problem
Agents made it trivial to build software that will only ever have one user or a
handful — a tool for your team, a tracker for your own workflow. Deploying and sharing it is as hard as it was
in 2015. AWS and Azure were designed for software that scales to millions, and that complexity is now the
bottleneck between “I built a thing” and “other people can use my thing.”
Who says it's hard
Y Combinator's Requests for Startups —
“A Cloud for Small Software.” And today, the market said it louder than the RFS could: the
top of Hacker News was a solo-built personal agent
(today's §1) whose entire selling point is “one command and it's deployed to your own account”
— 270 points and a 124-comment argument about what “self-hosted” even means anymore.
Why it's still open
The hard parts aren't the hosting. They're auth and permissions for arbitrary
user-written code, per-person environment customisation, and letting non-technical people share executable code
without it becoming a security incident. The Talorys thread shows the trap vividly: the “free tier”
quietly produced surprise bills for some users —
the economics of running one person's software are still nobody's solved product.
Who's already trying
Vercel, Replit, Val Town, Cloudflare Workers, Deno Deploy — and note that
Cloudflare just absorbed Deno and is shutting Deno Deploy, so the
field of independent runtimes is shrinking exactly as demand grows. That's a gap opening, not closing.
Your angle
You already live in this problem: you build with Claude Code and then have
nowhere good to put the result. That is the founder's unfair advantage in its cheapest form — you are the
user. Pair it with today's §6: the prototype is “one command that takes a Claude Code project and gives me
a URL with a password.”
Research path
Build the ugliest possible version for yourself first — take one
Claude Code project and get it to a live URL with a login this weekend (§6 walks the exact path).
Read the Talorys
HN thread end to end for the real objections: lock-in, surprise billing, and “is this even
self-hosted” — each complaint is a feature spec.
NEEDS A LAB · CLEARANCE · A DECADEred · the company is red; the on-ramp is laptop-sized
Finding and ranging a target without emitting a thing
The problem
Radar and lidar work by shouting — they emit a signal and time the echo, and that
emission gives your position away and can be jammed. The hard problem DARPA just put money behind is doing it
silently: locating and measuring the distance to objects using only what the world already radiates, no active
emission at all. It's the sensing analogue of GPS-denial — operate when you cannot, or dare not, broadcast.
Passive-only ranging is an underdetermined physics problem: without a known emitted
pulse you're inferring depth and distance from faint, ambiguous cues, which needs sensor fusion and estimation most
teams can't do well. The math is genuinely unsolved, not merely unbuilt.
Who's already trying
Defense primes and DARPA performers, and — from the satellite side — the
LEO-PNT players (Xona, TrustPoint) attacking the same GPS-denial problem by broadcasting a better signal instead of
needing none. Two opposite answers to one problem.
Your angle
Red as a company: defense, hardware, clearance, capital. Said honestly so you
file it, not start it. But the on-ramp is the exact question your tiltrotor faces the day GPS drops or lies:
emitter-less positioning and monocular / passive depth estimation, provable in simulation on a laptop. Own the
vocabulary; the company is not yours to start yet.
Research path
Read the
DARPA CIDAR page —
what exactly they're scoring, and why “passive-only” is the constraint that makes it hard.
Study monocular depth estimation as a signals problem — the laptop-sized version of the same question,
and a real rung on your drone's autonomy ladder.
Connect it to GPS-denied navigation: the positioning
problem has an active-satellite answer (Xona) and a passive-sensing answer (CIDAR), and knowing both is the
literacy that makes you useful to a defense-autonomy team.
New on the feed today: the problem-feed sources all returned this run. YC's
RFS is unchanged since the Fall 2026 publish — no new funded problem statement to add.
Marginal Revolution
ran Patrick Collison on the economics of agents (filed in §3, not a board item). The one genuinely new DARPA item
— four teams advancing in the Quantum
Benchmarking Initiative — is a program update, not a fresh “here is a hard problem and here is
money” solicitation, so it's noted here rather than added as a board entry. Both of today's problems are from the
standing board, rotated.
§1
The Big Three
researched
Build this weekend · the §0-green problem, as a working repo
A solo developer shipped a personal AI agent that lives entirely in your own cloud account — and topped Hacker News with it
Talorys · by rociiu (GitHub) · open-source (MIT) · one-command deploy via
npx create-talorys · runs on Cloudflare's free tier ·
GitHub ·
HN (270 pts, 124 comments)
What they do
Talorys is a personal AI assistant — chat, persistent memory, tasks, notes, projects and scheduled reminders
— that you deploy into your own Cloudflare account with a single command. No signup, no accounts, no
multi-tenant anything: it's “single-user by design.” You set one owner password at install; the data
lives in your account, not the author's. It is the §0-green problem answered in the most literal way: build a
tool for exactly one user, and make deploying it one command.
The tech — and what's hard
TypeScript throughout: React + Vite front end, a Hono Worker as the private API, and
— the interesting part — a Cloudflare Durable Object with SQLite holding conversations,
memories, tasks and the scheduler (recurring jobs run off Durable Object alarms). Chat runs on Workers AI
(glm-4.7-flash). It deliberately uses only free Cloudflare services — no R2, D1, KV or
Vectorize — which is the whole trick and the whole fragility. The hard parts are exactly the §0 ones:
permissions (it wants scoped OAuth tokens, not your global key) and cost (the free AI tier is ~10,000 “neurons”
a day, roughly 20–30 minutes of back-and-forth before it throttles).
Where the idea came from
It reads as scratch-your-own-itch: one person wanting a private assistant they fully control, in a week when the news
was all about agents you can't control (the thesis above). The design answer — put the whole thing
inside infrastructure you already trust and can see — is the same instinct driving Oxide and Xona at far larger
scale. Ownership as a feature.
The bear case
Hacker News took it apart, and the objections are the product spec. It's not really “self-hosted”
— it's self-managed on someone else's platform, with real Cloudflare lock-in (Durable Objects don't port). The
“free” tier bites: commenters reported surprise Workers AI charges and, in extreme Durable-Object cases
elsewhere, bills reaching $34K — set spend limits. And the free model is slow. This is a demo
of a pattern, not a finished product.
For you
This is the most directly actionable item in the issue. It's the §0-green problem with a working repo attached:
clone it, deploy it, read where it's thin, and you're standing inside the problem YC wants funded. The transferable
move isn't the code — it's that one person, over a weekend, took the week's biggest anxiety (uncontrollable
agents) and shipped the smallest ownable answer, and it went to the top of HN. That's the whole exploration-phase
playbook: build the small ownable version of the thing everyone is worried about.
The energy wedge · deep tech where your degree is the bottleneck
Petra Power is betting ceramic fuel cells can power both data centers and military vehicles — without burning anything
Petra Power · founded 2017, Warren, Ohio · CEO/co-founder Aaron Goodman (Princeton PhD,
Mechanical & Aerospace Engineering, plasma physics lab), CTO/co-founder Phillip Clift · ~15 people
· ~$9M in U.S. Defense Department contracts · venture funding not disclosed ·
Techstars Alabama EnergyTech ·
TechCrunch
What they do
Petra builds solid oxide fuel cells (SOFCs) — ceramic devices that turn a fuel like natural
gas directly into electricity without combustion. Goodman's framing: a normal generator goes fuel →
explosion → heat → motion → electricity, losing energy at every step; an SOFC “takes the
electrons directly off of the fuel — one step, no loss.” Two target customers, oddly symmetric: data
centers desperate for on-site power, and military vehicles that need quiet, efficient auxiliary power in the field.
The tech — and what's hard
The physics is real and old; the hard part is manufacturing ceramic stacks that survive high temperatures, thermal
cycling and years of runtime at a cost that beats just burning the gas. SOFCs have promised efficiency for decades
and mostly died on durability and price — which is why Petra's timeline is honest and slow:
first infrastructure deployment in 2028, full-scale production in 2029. This is not a this-year
story; it's a watch-the-vocabulary story.
The founders
Goodman's doctorate is in mechanical/aerospace engineering out of Princeton's plasma physics lab — i.e. energy
conversion at the level of electrons and materials, not slideware. The unfair advantage is domain depth in exactly
the layer that's hard: the cell chemistry and stack engineering. The company is nine years old and tiny, which tells
you how long this kind of deep-tech grind takes before the market shows up.
The bear case
No disclosed venture round and only ~$9M in DoD contracts for a nine-year-old hardware company is thin —
deployments are 2028+, hyperscaler deals are “ideal targets” with none signed, and defense use is still
in testing with no live vehicle. SOFC is a graveyard of good physics killed by manufacturing cost (Bloom Energy took
two decades and billions to reach scale). Capital-intensive, slow, and competing against the thing that already
works: just burning the gas.
For you
File this against your open thread: AI's real bottleneck is power, and your field of study is adjacent to the energy
layer. Petra is a live, named datapoint in that wedge — and the defense-vehicle angle means it sits one domain
from your §0-amber too. Don't start a fuel-cell company. Do learn to read these: the money is quietly
flowing to on-site generation (see §2's fusion and grid-flexibility rounds), and the person who can talk fluently
about where the electrons come from will be in the room when the autonomy and compute people realise they can't get
any.
Founder war story · access is the only real moat
The PillPack founders sold to Amazon for $753M, left, and are back — a16z just put $120M behind their second act
General Medicine · founded 2023 · CEO TJ Parker & president Elliot Cohen (co-founders
of PillPack, sold to Amazon for $753M in 2018), with ex-Amazon exec Ashwin Muralidharan (now CPO)
· $120M Series B led by a16z, with Eli Lilly, Mercy Health, Matrix, BoxGroup
·
Fierce
Healthcare ·
a16z ·
Endpoints
What they do
General Medicine is building a “general store” for healthcare — a single storefront for telemedicine,
prescriptions, imaging, labs and specialists, with transparent pricing and your history carried across all of it. It
runs its own medical group plus a marketplace of outside clinicians. In the brief's language, it's a
§0-style “operating system for the physical world,” pointed at the single most coordination-starved
industry there is.
The founders
TJ Parker and Elliot Cohen built PillPack in 2013 — the mail-order pharmacy that pre-sorted
your pills by dose — and sold it to Amazon in 2018 for ~$753M, where it became Amazon Pharmacy. They ran
Amazon's pharmacy and clinic efforts, learned the whole machine from inside the most operationally ruthless company
on earth, then left
in 2022. Parker only stepped in as CEO this month. This is the second run at the same market, with the scar
tissue.
Where the idea came from
The access is the whole story. Almost nobody alive has both built a healthcare-logistics company from zero
and watched Amazon try to scale it — they know exactly which walls the $500B prescription market puts
up, because they spent four years hitting them inside Amazon. That's not a market you can research your way into; you
have to have lived in it. a16z, Eli Lilly and a hospital system are paying $120M for that lived knowledge, not a deck.
The bear case
Healthcare “super-apps” are a boneyard — the coordination is hard precisely because incentives,
regulation and incumbents all fight it, and a famous founding team doesn't repeal any of that. “One storefront
for everything” is what Amazon itself, with infinitely more resources, could not make stick in health. The
founders' edge is real; whether anyone can win this category is the open question.
For you
The transferable lesson is the one the brief keeps hammering: the durable moat is unusual access to a problem,
and the surest way to get it is to go work inside the problem before you try to own it. Parker and Cohen's “unfair
advantage” is four years inside Amazon's pharmacy, not a patent. Your version isn't healthcare — it's the
§5b advice made concrete: a US internship inside a hard problem (defense autonomy, robotics, energy) buys you the
same kind of access, and it's a door you can actually walk through.
§2
Money Moved
18 funding items screened · 6 filed
Who
Deal
What & why it's here
Src
Axiom Solutions
$2B
Raised at a $37.5B valuation (General Catalyst, Koch Equity) — a cloud-and-power infrastructure spinout from Flex. The biggest round of the week, and it's not a model company: it's the picks-and-shovels of compute and the electricity to run it, bundled. That pairing — servers plus power — is the whole thesis of today's §1 Petra and §9 energy items in one cheque.
For fault-tolerant quantum computers (Arch, Spark, Khosla). File next to DARPA's Quantum Benchmarking final in §0 and §9: capital and the Pentagon are both betting the error-correction problem is finally cracking. “Fault-tolerant” is the word that separates a lab toy from a machine — and ties straight to the §0 crypto-agility shadow.
“AI-native digital manufacturing systems.” Alongside Vinci ($250M, $1.5B val, an AI platform for hardware engineers), the tell is clear: money is pouring into software that designs and makes physical things. That's the design-process layer you care about — the reasoning about why a part is shaped the way it is, turned into a fundable product.
Distributed-energy grid-flexibility platform (Generation IM, Vitol). The software layer on top of the power problem — squeezing capacity out of a grid that can't build fast enough. The energy wedge keeps showing up as software, which is the part you could actually touch.
Fusion (Breakthrough Energy Ventures). Still a decade out, still red on the §0 board — but a $200M round is the market pricing “we will need fundamentally more power” as fact, not hope. Watch, learn the vocabulary, don't start.
Series B led by a16z, with Eli Lilly and Mercy Health (see §1). The PillPack founders' second act — filed here because the investors are the signal: a drugmaker and a hospital system putting money in is a demand-side bet you can't fake.
The week's top-10 was “AI and energy,” in Crunchbase's own framing, and that's the whole read: half the
biggest rounds were about making and moving electricity, not models. Meanwhile your backyard cooled —
Indian startups raised about $123M
across the week, dominated by a single D2C Series C (Nat
Habit, ₹142 Cr). One more reason the brief keeps pointing you at US doors you can walk through without a visa.
Collison (Stripe) lists what he's mulling about agents in the economy, and Tyler Cowen — who runs Emergent
Ventures, i.e. a grant you can apply to — posts it. The sharpest framing: “companies are in some sense the
original superintelligence,” using amortized cognitive surplus against inattentive consumers. Read it as a list
of unclaimed problem areas from two of the highest-signal people alive, and remember who's holding the chequebook.
A study finds the efficiency gains from AI coding get “absorbed” by the human review bottleneck
— more code produced, not more shipped software. This is the mental model under today's §0-green and
yesterday's vibe-code thread: the scarce resource moved from writing code to trusting it. Whoever
shrinks the review bottleneck for non-experts owns the next layer.
Human mathematicians sprinting to finish a proof before an AI did — the concrete, non-hype version of the
§0 crypto-agility fear (AI-accelerated math coming for the hard problems cryptography rests on). Pair with
Quanta's companion piece, “Is AI
the End of Math As We Know It?” Not a lecture — a race, with stakes.
Teaching a physical system to act aggressively and stay provably safe by separating the safety filter from
the task objective, instead of blending them into one reward that fights itself. Your tiltrotor's exact next problem
after “hold position,” and it sits in a cluster of safe-control papers this week
(contextual safety filtering for motion). Read one to get literate in
how “guaranteed-safe control” is actually posed — the vocabulary you'll hire against.
MIT built a motion planner with a provable no-collision guarantee for flying robots. Between this (don't hit things)
and the FAITH paper above (stay safe while acting hard), you have both halves of the control problem your build log
is marching toward. File as the research your “hold position → fly a path” rung will stand on.
Unchanged since Fall 2026 and still the highest-authority idea source there is. Today's §0-green (“A Cloud
for Small Software”) is lifted from it word for word — and §1's Talorys is someone building exactly
that. Free, and a map of where the smart money wants founders to go.
Today's §1 build piece and the day's most actionable repo. Read the 124-comment thread as much as the code: it's a live argument about what “self-hosted,” “free” and “yours” actually mean once your software runs on someone else's platform — the exact §0-green fault lines.
A hands-on walkthrough of building a small, non-LLM “decision model” — the same category Jev/TypeSafe raised $870M for and Cloudflare shipped (clef) this week. Here's the from-scratch version so you understand what the billion-dollar round is actually selling: not everything needs a giant language model.
Another strong open-weight model tops the trending charts, a comfortable margin ahead of everything else. The open, cheap tier a bootstrapped builder actually lives in keeps getting better — and Talorys above runs on a free model precisely because this tier exists. File as a component, not a headline.
Lightricks' open-weights video model, one of the week's most-downloaded. For the filmmaker in you: state-of-the-art generative video you can run yourself, no per-clip cloud bill. The economics of making things look good are going local and free the way photo editing did — cheaper to learn now than to discover in a year.
A paywall-skipping tool died, so someone rebuilt it over a weekend and shipped it. Here as a pattern, not an endorsement: the half-life of “a useful small tool vanished” to “someone replaced it” is now days. The §0-green problem's demand side, in miniature.
A game whose premise is that the city actively resists being built — here purely for the design idea. Taking the core verb of a familiar genre and inverting it is the cheapest, sharpest form of originality there is, and it transfers straight to product: what's the thing everyone assumes your category should do, that you could invert?
§5
The Arena
13 live · deadlines computed for today
Devpost returned this run, so this is a 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 (see §4) · closes Oct 16
5d
Hyperbloom October — UI/UX & Web Design
$9,117 · design — a pure taste play, squarely your eye · closes Oct 19
8d
EurekaDev 2026
purse not stated · general software build · closes Oct 20
9d
Build, Ship, Shape: Amazon Developer Hackathon
$138,000 · ML/AI, productivity — biggest purse on the board · closes Oct 23
12d
Life After Code
purse not stated · AI / future-of-building theme — founder-framed · closes Oct 27
16d
OpenCV AI Competition 2026 (AWS)
$20,250 · IoT + vision — the closest thing on the board to your drone · closes Oct 27
16d
AWS CDS Agentic AI Partner Hackathon
$40,000 · agentic AI — build a §0-style agent (think Talorys) and get judged on it · closes Oct 28
purse not stated · agentic AI on a taste/culture API · closes Oct 30
19d
Galuxium Nexus V2
purse not stated · general AI build · closes Oct 31
20d
YouCam API Skin AI & eCommerce VTO Hackathon
purse not stated · computer vision + e-commerce · closes Nov 2
22d
PayPal AI Hackathon
purse not stated · agents in payments — the regulated-money vertical · closes Nov 12
32d
Meta VR Start Developer Competition 2026
purse not stated · VR/XR — furthest out, most specialised · closes Nov 18
38d
The ones that fit you with real runway: OpenCV AI (16d) is IoT + vision — point a camera
at a drone payload and you have an entry; Nebius × NVIDIA (19d, $50K) is the beginner-friendly
workhorse that leaves dated proof-of-work behind; and AWS CDS Agentic (17d, $40K) is a direct excuse to
build the §0-green Talorys-style agent and get it judged. If you want one this week,
Hack Apertus (5d) rides the open-weight wave you already live in. Register for one today — it
converts intention into a deadline.
§5b
Open Doors
US internships — the access play from today's §1
General Medicine's founders got their moat by working inside Amazon's pharmacy for four years. Your version: a US defense-autonomy, robotics or energy internship buys the same lived access — and you're US-eligible without sponsorship, the single barrier that blocks most of your peers. Apply to one this cycle; the role doubles as a §0-amber on-ramp.
now
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 Collison post is him showing you, in public, what he thinks the open problems are. Reading Marginal Revolution is reading the grader.
open
YC Requests for Startups — the free map
Not an application, a target list. Today's §0-green is lifted straight from it, and §1's Talorys is someone building it. Build a named problem and you've pre-aligned with the highest-signal funder there is.
open
§6
Skill
day 13
Deploy one thing to a real URL this weekend — and you've stood inside the §0-green problem, not read about it
~2–3 hours · free · produces: one Claude Code project live on the internet behind a password, and a felt map of exactly where “deploy my small software” is still broken
Today's §0 green is “deploying personal software is still hard,” and §1's Talorys is one person's
answer. The cheapest way to understand the problem is to feel it: take one thing you built with Claude Code and
get it to a live URL with a login, today. Use Cloudflare
Workers (or Pages for a static front end) — the free tier, one wrangler deploy command. Add the
simplest possible auth: a single password check, exactly like Talorys does. Then read the
Talorys repo as a worked example of the same path done well —
Durable Objects for state, scoped tokens for permissions, a scheduler built on alarms.
Two things happen. First you hit the real walls: auth is fiddly, “free” has sharp edges (watch the billing
— set a spend limit before you start), and sharing the URL with one non-technical friend surfaces every gap at
once. Second, you're now holding the primitive. The distance between “I deployed my own tool with a password”
and “one command that does this for anyone with a Claude Code project” is the entire §0-green startup.
Build the ugly version for yourself first; the product is the ugly version with the rough edges filed off.
Free resources: Cloudflare Workers Get Started
(deploy in minutes, no card for the free tier) and the Talorys source as
a reference implementation. Bring one real project you made; a hello-world teaches nothing here.
§7
Watch
canon · YouTube talks all 404'd this run
Never Split the Difference — FBI hostage negotiation, turned into the dealmaking you can use this week
The canon · money, power & dealmaking track · book by Chris Voss, ~7h audio / 8h read · publisher
Every YouTube feed 404'd or 500'd this run — no fresh talk to run — so the canon, and this is the pick for a
week about who you can and can't trust. Never Split
the Difference is Chris Voss's field manual from running FBI hostage negotiations, rewritten for ordinary
high-stakes talks: a fundraise, a cofounder split, a first customer, a salary. It's the non-fiction version of the thing
you already love watching — Ari Gold working a room — except it hands you the actual moves: tactical empathy,
calibrated questions, the “no”-oriented open, mirroring. The reason it's in the canon and not a dozen other
negotiation books: it is immediately usable, this week, in a conversation you'll actually have.
Why now: you're in an exploration phase that will, soon, run on conversations you don't yet know how to win —
convincing a stranger to intern you, a teammate to join, a judge to pick your hackathon entry. Most startups die of human
problems, not technical ones; this is the underpriced skill, and it reads like a thriller, not a textbook. The
dealmaking track last surfaced with Michael Ovitz; this is the usable-this-week companion to 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: history, biology, the
cosmos, global reporting, pure whimsy. None of it is curated toward what you already like; that's the point. One line on
why it caught my eye.
The most distant radio signal yet pulled from the early cosmos — light that left before the Earth existed, decoded now. A clean reminder of the scale everything else in this brief is happening inside.
An argument that we could simply eliminate the mosquito species that kill hundreds of thousands a year — and mostly choose not to. A jolt of agency: some vast old problems are open not because they're unsolved but because nobody decided to close them.
On the writer who made deep time readable, and how he structures a story so that rock becomes narrative. For someone who watches films to see how they were made: this is the prose equivalent, the craft of construction laid bare.
A startling thesis: the same armies fight badly across decades not from kit or courage but from how their societies handle information and initiative. A case study in how culture quietly sets the ceiling on performance — a company lesson wearing a military uniform.
The internet is physical — a few subsea cables through one contested strait — and Google and Meta are literally rerouting the planet's data around a war. The cloud you think of as weightless runs through a handful of vulnerable pipes on a seabed.
A “book rave”: nightclub, but the activity is silent communal reading on the floor. A tiny, delightful piece of evidence that there's still unclaimed space between familiar formats — nobody said a club had to be loud.
Someone recreated the iconic 1976 supercomputer — its curved bench and all — out of discarded Mac Minis, which each dwarf it in power. Equal parts hardware archaeology, sculpture, and a quiet joke about how fast “the fastest machine on Earth” becomes a shelf of e-waste.
§8
Deals
1 free resource worth your time
Cloudflare's free tier is enough to run today's whole §6 — a real backend for $0
Free · no card required for the free plan
The genuinely useful free thing today is the one that makes the §6 skill cost nothing:
Cloudflare's free Workers tier gives you
100,000 requests a day, Durable Objects for stateful apps, and a daily allowance of Workers AI inference — enough to
deploy a real backend, a personal agent, or a small tool with auth, without a credit card or a cloud bill. It's exactly
what Talorys stands on. One honest caveat, straight from the §1 HN thread: the free tier has sharp billing edges once
you cross it, so set a spend limit before you
deploy anything you'll leave running.
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
one free resource this issue is built around, which beats a coupon.
§9
Also
filed, not featured
Satya Nadella says we should assume every AI model is “compromised.” Microsoft's CEO argues we can no longer treat AI as a “set of nested black boxes” whose actions we simply accept, and (separately) calls for an “emergency brake.” The thesis of this whole issue, said by the man who sells the most AI: trust is now the product, which is why value is flowing to layers you can audit.
Nvidia is in talks to acquire or invest in Reflection AI, a US open-weights startup. The FT reports it may be structured as an acqui-hire to dodge antitrust, with the Trump administration hoping it becomes a US rival to cheap Chinese models like DeepSeek. Open weights have become a geopolitical instrument — file next to §4's Qwen: the cheap, open tier you build on is now contested between states, not just labs.
Ukrainian drones knocked out an AI data center belonging to “Russia's Google.” Cheap, attritable drones destroying a multi-hundred-million-dollar compute facility is the §0-amber (“American defense”) and the energy/compute threads colliding in one event: the expensive, centralised layer is exactly what the cheap autonomous layer is built to kill.
Four more teams advance in DARPA's Quantum Benchmarking Initiative. Read next to §2's $475M for fault-tolerant quantum (Oratomic): both the Pentagon and the venture market are betting the error-correction problem is finally moving — the quiet precondition for the §0 crypto-agility threat becoming real.