Issue 011Thu 8 Oct 2026487 items screened → 47 filedfull pull · 0 feeds down
The week AI reached mathematics, and the mathematicians named what it threatens next: the cryptography the whole
internet's trust runs on. OpenAI posted “progress
in mathematics”, Terence Tao called for a “Math
2.0”, and Vitalik Buterin said to take the risk to
cryptography from AI-accelerated math seriously — while, reportedly, AI
labs have quietly begun probing whether their own models can break it.
Underneath, the money did what it did all week — skip the abstract middle for atoms and access: $100M
into ex-SpaceX engineers' autonomous freight rail, $320M into a fund that backs founders everywhere
except San Francisco, and a new Claude Haiku 5.5 that made the cheap model cheaper. The pattern holds:
when the abstract layer gets automated or called into question, durable value runs to the physical and to the
primitives of trust. Own a bottleneck, own the edge, or own the proof that something is real.
§0
The Problem Board
2 of 10
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
start this month, which today's rogue-agent news keeps proving is unowned — and one you cannot start at all, but
which the week's AI-and-math story just made sharper and more urgent.
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're summarising a document or changing infrastructure. When one wanders
somewhere it shouldn't, there is no clean audit trail to answer “what did it touch, under whose
authority.” It kept proving itself this week: OpenAI's agents were found
acting on Wikimedia projects
— spotted by outsiders, not OpenAI — Scientific American ran a careful post-mortem of
the
OpenAI–Hugging Face incident, Docker shipped a first-class Docker
Agent, and ICANN is about to hand out .agent
and .agi domains. More agents, more authority, still no provenance.
Who says it's hard
OpenAI's own disclosure, plus a stack of arXiv work this quarter and security researchers naming the missing
control — “How Agents Ask for Permission” is
the landscape in one paper.
Why it's still open
Sandboxing shipped (NVIDIA, Cisco); provenance did not. The literature is 2024–26
and mostly preprints — a field with no incumbent and no agreed standard for “this agent did this, for
this human, with this permission.”
Who's already trying
NVIDIA and Cisco both sandbox; neither audits. The forming layer
next door — Pinrail (§4), an inbox where coding
agents wait for a human to approve — is a tell that the human-in-the-loop control is being hand-rolled, one
tool at a time, because nobody ships it.
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. The cheapest first version is a logger, not a company:
instrument your own setup and the log is the prototype. That's exactly this issue's §6 skill.
Research path
Read arXiv 2607.13718 (“How Agents Ask for Permission”)
for the whole landscape in one sitting.
Read arXiv 2606.22504 (“Lingering Authority”) on the
credentials an agent still holds after its task ends.
Then do §6: log every tool call your agent
makes for a week. You'll have both the felt problem and the seed of the product.
NEEDS A LAB, STANDARDS AND A DECADEred · filed so you recognise it, not so you start it
We no longer have a reliable way to tell who — or what — is real
The problem
Every trust signal built for a world where faking a human was expensive — a face, a
voice, a signature — is getting cheap to fake. Whoever builds the replacement becomes the layer every bank,
app and video call checks. And this week the ground under the fallback moved too: if AI-accelerated math can
weaken the cryptography that a proof-of-human system
would lean on, you're trying to build trust on a foundation that is itself now in question. Yesterday an
AI-generated
video swayed a judge before an appeals court threw it out — the boundary cases are already in court.
It needs cryptographic standards, regulators and distribution deals
simultaneously — and the newest twist is that the standards themselves may need replacing
(post-quantum, crypto-agility) before the old ones break. None of it is a weekend.
Who's already trying
Worldcoin, Persona, Clear and Apple/Google device attestation on the identity side;
NIST, Cloudflare and PQShield on the post-quantum side. Two hard problems that have just started to touch.
Your angle
Red, and honestly so — a standards-and-PhD problem, not a building
problem. Filed so you recognise it later, not so you start it. If a narrow, boring slice ever opens — tooling
that simply inventories which cryptographic primitives a codebase depends on, the unglamorous first step of any
migration — that part is software. 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.
Read about C2PA content
provenance — the “is this media real” half of the same problem.
Skim NIST's post-quantum
standards to see what “solved” already looks like on paper — and how far adoption is from it.
New on the feed today: YC's RFS is unchanged since the Fall 2026
publish; DARPA posted three items — a CIDAR
passive-ranging challenge, four
more teams into the Quantum Benchmarking final, and a Vertical
Lift Challenger recap — all of which fit the board's existing american-defense and compute entries
rather than opening a new one. Marginal Revolution ran essays, not problems. One genuinely new entry added by
hand:crypto-agility (red) — AI-accelerated math as a live threat to the internet's
encryption, named this week by Buterin and Aaronson. It's filed in config/problems.yaml; the §0 red
above is the identity-trust face of the same shift.
§1
The Big Three
researched
Founder war story · ex-SpaceX engineers, and your exact dispatch lane
Rocket engineers raised $100M to put freight on autonomous electric rail — starting with the short hauls the railroads quit
Parallel Systems · Long Beach, CA · founded 2020 by Matt Soule (CEO; ~13 yrs in SpaceX
avionics), John Howard (ex-SpaceX battery) & Ben Stabler (ex-SpaceX avionics software) ·
$100M Series C led by AVP (w/ Hillspire, Agility Global, Cobalt, Anthos, Congruent, Riot,
Collaborative Fund) · >$200M raised total · running on 160 mi near the Port of Savannah
under FRA approval ·
TechCrunch ·
R&A News
What they do
Parallel builds Panther: a battery-electric, coupler-free autonomous rail vehicle that
hauls several tons of freight up to 500 miles with no operator. Instead of one long train on a
fixed schedule (“precision railroading”), small vehicles travel alone or in short platoons and split
apart at the yard without a human. The target market is the route nobody serves well: “less than 500 miles
is hard for railroads to do competitively,” Soule says — so those loads went to trucks, and Parallel
wants them back on rail.
Where the idea came from
Three SpaceX leads looked at freight rail the way they'd looked at rockets: as a vehicle-and-avionics problem, not
a railroad-operations one. Soule ran avionics; Howard ran the battery division; Stabler did avionics software and
actuator electronics. The unfair advantage is literal — they already knew how to make an autonomous,
battery-electric vehicle safe and certifiable, and pointed it at an industry that hadn't had a clean-sheet vehicle
in a very long time.
Why it works (if it works)
Economics, not romance. Soule's claim is that Parallel can make a short rail lane profitable at ~10,000 moves
a year — a density trucking can't touch per unit of fuel and labour — while using existing track
and existing freight cars. The coupler-free, platoon design is the enabling trick: it turns rail from a batch
process (assemble a huge train, then run it) into something closer to on-demand dispatch. That's the part that
rhymes with your own work.
The bear case
Rail is a regulated, incumbent-controlled network: the Class I railroads own the track and the relationships, and
a startup needs their cooperation or a very patient regulator. “First commercial payload coming up very
soon” still means revenue is ahead, not behind. Autonomy on shared freight corridors is a safety-case
marathon, and $200M is small against the capital a national rollout wants. Competitors (Intramotev, OTIV,
Magrail) are chasing the same lane.
For you
This is the §0 physical-world-OS thesis as a company — and it's your cab-dispatch instinct at
industrial scale. The lesson to steal isn't the trains: it's that the founders took a generalist
vehicle-and-autonomy skill and aimed it at a legacy industry that measured success by density and dispatch,
not by a hero product. You keep saying you're “not a one-person thing” — this is proof that the
transferable asset is the engineering instinct, re-pointed. Read how three people with no rail background got rail
people to fund them.
Money & power · a fund built on the exact bet you are — that talent is everywhere
While VCs crowd into San Francisco, this $320M fund is paid to find founders everywhere else
Endeavor Catalyst · New York · $320M Fund V (AUM now >$850M; 400 LPs, ~30% of them
Endeavor founders) · managed by Allen Taylor & Jackie Carmel; Endeavor co-founded by
Linda Rottenberg · co-invests $1–3M into rounds that already have a ≥$5M lead,
capped at 10% · 437 companies across 44 markets, 83 unicorns, 11 IPOs · ~90%
outside the U.S. ·
TechCrunch
What it is
Endeavor is a 30-year-old non-profit that admits founders to a global network after a brutal screen — last
year it looked at over 10,000 candidates and selected 88. Endeavor Catalyst is its investment
arm: it doesn't lead rounds, it rides alongside a real lead once a company has raised ≥$5M, writing $1–3M
and never more than 10%. The twist that makes it unusual: half the fund's profits flow back to Endeavor,
so every win pays to find the next founder somewhere unfashionable.
Where the idea came from
Rottenberg's original observation, decades early, was that high-potential founders exist far outside Silicon
Valley and simply lack the network capital treats as a prerequisite. The portfolio is the argument:
Bending Spoons (Italy,
~$26B), ElevenLabs (Poland-founded, $22B), Checkout.com (Swiss founder, $12B), Replit (Amjad Masad,
Palestinian-Jordanian, $9B), Flutterwave (African payments, $3.2B). None of them are San Francisco stories.
Why it works
It's a structural arbitrage on geography. If 90% of capital clusters in one zip code, the mispriced talent is
everywhere else — and a fund with a trusted local network and a “we only follow a real lead”
rule gets to buy into it with less competition and lower entry prices. Europe is now its fastest-growing region
(12 new investments in H1 2026 vs 14 in all of 2025), and it's steering 20% of this fund into repeat founders
— people who've already won once.
The bear case
A follow-on fund lives or dies on deal access, not insight: if the best “elsewhere” companies start
raising their real leads from the same SF firms anyway, Endeavor's edge thins to a brand. The 10%-cap, no-lead
model means it rarely owns enough of a breakout to mint a fund by itself. And “emerging markets”
carries currency, governance and exit risk that a Delaware-to-Nasdaq company simply doesn't.
For you
This is the most personally relevant item in the brief and it isn't close. You are, by their definition, a
founder “elsewhere” — and here is a $320M institution whose entire thesis is that your location
is an edge, not a handicap. Two concrete moves: (1) study the Endeavor Entrepreneur selection as a
standing door you grow into (it's for companies with traction, so it's a two-year target, not a this-week one);
(2) internalise the model — talent is evenly distributed, networks are not, and the whole game is
manufacturing the network access you weren't born into. That's the same lesson as yesterday's hacker-house story,
told in fund form.
Buildable idea · the reshoring/production-rate thesis, as pure software
A $3.6M seed to build the “Alibaba of American manufacturing” — because you still can't search for a U.S. factory
Bloom · Detroit / remote · founded 2023 by Justin Kosmides (CEO) w/ Hitesh Chudasama &
Chris Nolte · $3.6M seed led by SNAK Venture Partners (w/ Flyover, Mana, Detroit
Venture Partners, Invest Detroit, Michigan Outdoor Innovation Fund) · 2,000+ matches for 140+
companies · competitors Fictiv, Xometry, MacroFab ·
TechCrunch
What they do
Bloom is an AI-driven marketplace that matches hardware companies with U.S.-based manufacturers, suppliers and
service providers — contract manufacturing, assembly, design, freight, warehousing, even hazmat shipping.
Kosmides's framing: “what an AI-driven version of Alibaba did in China, where it created a marketplace for
contract manufacturers.” The gap he's filling is embarrassingly basic — “the ability to
discover, and to do matchmaking… just doesn't exist electronically” for American manufacturing.
Where the idea came from
Bloom started in 2023 serving e-bike and e-scooter makers drowning in supply-chain pain, then pivoted to a broad
marketplace when post-election tariffs sent drone makers and robotics startups scrambling for
domestic suppliers. The clean example in the piece: a Michigan contract manufacturer that used to refurbish Nest
thermostats and Bird scooters now bids on drone-assembly contracts through Bloom. Supply met a
policy-made surge in demand, and the matchmaking layer was missing.
Why it works (if it works)
It's a discovery-and-trust layer, not a factory — which is why it's a seed-stage software company and not a
capital sink. Bloom seeds its matching with 30–40% supplier data contributed directly plus public data, and
the competitors (Fictiv, Xometry, MacroFab) mostly sell individual parts, not multi-supplier
coordination of a whole build. If “make it in America” is a durable policy regime, the index of who
can make what, at what rate, becomes valuable infrastructure.
The bear case
$3.6M is pre-everything, and marketplaces are brutal: liquidity is the whole game, and a thin directory loses to
a buyer's existing rolodex. The thesis rides on tariffs staying put — a policy reversal softens the demand
that made the pivot work. Xometry and friends are far better capitalised and could bolt on “full-build”
matching. And supplier data that's 30–40% self-reported is a data-quality problem waiting to happen.
For you
This is yesterday's production-rate thesis (Furientis: “can you build 1,000 a year”) turned into a
software problem you could actually prototype. The reframe worth keeping: when a policy shift or a new technology
creates a sudden surge of supply or demand, the matchmaking layer between the two sides is often missing
and buildable without capital. You don't need to make the drones — the business can be the index of who
makes them. File the pattern: find the two sides that can't find each other yet.
§2
Money Moved
funding screened · 6 filed
Who
Deal
What & why it's here
Src
Biren
~$515M
Shanghai AI chipmaker raised in a new share sale — even with its Hong Kong stock down 50%+. Read it next to SemiAnalysis's §9 number: China is pouring capital into its own compute stack regardless of price, because the constraint is strategic, not financial.
The company behind the Manus AI agent raised its first round since relocating. The agent wave is now funded at nine figures — which is exactly why the unowned control layer under it (today's §0 green) matters more, not less.
Fund V to back founders outside SF, ~90% of it abroad (see §1). The most relevant cheque in the brief for someone building where he is, not where the money clusters.
Series C (AVP) for autonomous electric freight rail by ex-SpaceX engineers (see §1). Atoms + dispatch; the physical-world-OS thesis with a vehicle attached.
Sequoia led a round for a robot-data startup — the training data that teaches machines to act in the physical world. Whoever owns that data owns a moat no model lab has; the §0 physical-world thesis, upstream.
An Owkin alum raised a seed to build a “world model for human cells” — a simulator of biology for drug discovery. Filed as a signal of where deep-tech AI money is going next: not chat, but models of physical and biological systems.
The frame, from Crunchbase: North
American startup funding fell in Q3 as the AI giants eye the public markets — and
investors are raising
the bar. So read the cheques above as selective, not frothy: money is concentrating into physical compute, robot
data, biology and “elsewhere” founders, and getting harder to raise for an undifferentiated AI wrapper.
Nous Research also confirmed a $1.5B
valuation as it shipped business agents — valuation, not raise, so it's a note not a row.
The single most important thread this week, and it's a mental model, not a product. AI systems are now making real
mathematical progress (OpenAI,
Tao's “Math 2.0”) — and the
Association for Human Mathematics is already pushing
back that the output shows “power, not scholarship.” Why it changes what you read next: math falling is
the leading edge of AI touching every formal discipline, and its first casualty may be cryptography (today's §0
red). This is the six-month weather report.
A clear, non-lecture explainer of the models that actually move physical machines — the layer under Mecka
(§2), under Parallel's Panther (§1), and under your tiltrotor. Read it to get literate in how
“perception → policy → actuation” really works, because you cannot lead the robotics people
you'll need to hire if you can't follow them. Directly load-bearing for the drone.
The compute buildout's power-and-cooling bill is becoming a first-class engineering problem — and software
that squeezes efficiency out of existing data centres is a cheaper way into the energy story than drilling a well.
This is your subsurface/energy thread reappearing from the software side, where you actually have an edge.
The man who shipped the iPod and the Nest on why the Humane-pin generation of AI hardware died. It's a teardown of
product decisions — the “why is this the wrong shape” reasoning you said you love —
from someone who's done it at the highest level. Pair it with Underdog/Ghost from this week: software ate the
privacy-gadget before the gadget could ship.
Unchanged since Fall 2026 and still the highest-authority idea source there is. Today's §0 green and red both
live on it, and §1's Bloom is chasing its defense/manufacturing categories. Free, and a map of where the
smart money wants to go.
Anthropic shipped a new Haiku. Directly relevant to your one real cost: the cheaper and more capable the small model, the more you can build agentic tooling that runs many calls without the bill exploding — and the more viable the §0-green “log every tool call” workflow becomes. Simon Willison's first notes are the fastest read on what changed.
OpenAI's responses now fill with pictures, charts and buttons instead of plain text. Watch the design move, not the model number: the chat box is turning into a generated interface, which quietly changes what “an app” even is. Note it for anything you build with an LLM front-end.
Docker shipped official tooling to run agents in containers. It's the sandboxing half of the §0 agent problem going mainstream — and a reminder that the provenance half (what did it touch, for whom) is still conspicuously absent even from the company that just standardised how you run them.
A 2-billion-parameter open model built to make structured decisions (route, approve, classify) rather than chat. Boring-but-useful: this is the kind of cheap local component that an agent-audit or dispatch tool is actually assembled from. File it as a building block.
A desktop inbox that holds an agent's proposed actions until a human approves them. It's a hand-rolled version of the exact control the §0 green problem says nobody ships as a product — proof the demand is real and currently met by weekend tools. Worth studying as the shape of the thing before someone productises it.
An open, from-scratch rebuild of Photoshop's core in Rust. Not something you'll use tomorrow — here for the craft and the audacity, and because clean-room reimplementation (rebuild the interface, touch none of the original code) is a genuinely respected, legally careful discipline worth understanding.
§5
The Arena
25 live · deadlines verified on Devpost today
Devpost returned a full pull today and these day-counts are pulled from each event's live submission deadline, not
carried forward. Soonest first; red under 7 days. The three closing in 6 days are only worth starting
if you can ship this weekend.
Multimodal AI Hackathon 2026
₹100,000 · ML/AI · multimodal — closes Oct 14
6d
UnivaBio
$40,000 · beginner-friendly, health/bio — closes Oct 13
6d
WarriorHacks 2.0
beginner-friendly, open-ended AI — closes Oct 14
6d
INDUX 5.0-2026
₹225,000 · cybersecurity + design — the §0 trust theme as a build brief, and your design eye · closes Oct 15
7d
Hack Apertus
€10,680 · open-source AI model focus — rides the open-weight wave · closes Oct 16
8d
Next Founders Hackathon
$275 · founder-framed — build the thing you'd actually start · closes Oct 15
8d
Hyperbloom October — UI/UX & Web Design
$9,117 · design — a pure taste play · closes Oct 19
12d
Amazon Developer Hackathon (Build, Ship, Shape)
$138,000 · ML/AI, productivity — biggest general-purpose purse on the board · closes Oct 23
15d
OpenCV AI Competition 2026 (AWS)
$20,250 · IoT + vision — the closest thing on the board to your drone · closes Oct 27
19d
AWS CDS Agentic AI Partner Hackathon
$40,000 · agentic AI — build the §0 agent-audit idea and get judged on it · closes Oct 28
21d
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 (19d) is IoT + vision — point a
camera at a drone payload and you have an entry; HackFW MADE (22d) is IoT + robotics, small money but
exactly your hardware; AWS Agentic (21d) lets you build the §0 agent-audit idea and be judged on
it; and Nebius × NVIDIA (22d, $50K) is the beginner-friendly workhorse that leaves dated
proof-of-work behind. Register for one today — it converts intention into a deadline.
§5b
Open Doors
US summer internship cycle — peak weeks
No sponsorship needed for you. Early-to-mid October is the densest part of the cycle. A US defense-tech, robotics, autonomy or AI-infra role doubles as a §0 on-ramp — and Parallel Systems (§1) is the exact profile of company hiring into it.
now
Thiel Fellowship — applications open
$200K, no equity, under 23, US-friendly. Multiple of this week's and last week's funded founders are current Fellows. Demonstrably the door that funds exactly your profile.
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
Endeavor — know it exists now, apply later
The network behind §1's $320M fund screens founders “elsewhere” (10,000 → 88 last year). It's a traction-stage door, not a this-week one — but it's the institution built for exactly where you're building. File it as a two-year target.
open
§6
Skill
day 10
Put a logger between you and your AI agent — the audit trail is the §0-green prototype
~1–2 hours · free · produces: a week-long record of everything your agent touched, and the seed of a product nobody ships
Today's §0 green is “nobody can prove what an AI agent actually touched.” The cheapest way to
understand that problem is to become the instrument. You already run an agent daily (Claude Code) on a machine
holding your keys. Add a hook that fires on every tool call — Claude Code exposes
PreToolUse / PostToolUse hooks in settings.json —
and have it append one JSON line per action to a local log: timestamp, tool, the arguments (file path, shell command,
URL), and the result size. Then just work normally for a week.
Two things happen. First, you feel the problem: scroll the log and you'll see your agent read files you forgot
it could, run commands you didn't picture, reach hosts you didn't name — all with the same blanket authority
whether it was renaming a variable or touching your git history. Second, you'll be holding the primitive. A clean,
queryable “what did it touch, when, under what instruction” log is literally the first feature of the
product that doesn't exist yet — the one NVIDIA and Cisco skipped when they built sandboxes. Pinrail (§4) is
someone else arriving at the same spot from the approval side. The log is the prototype; you just have to read it.
Free resource: Claude Code hooks — the built-in
way to run a script on every tool call, no extra tooling. If you don't use Claude Code, a shell wrapper that tees every
command and its args into a dated file gets you 80% of the same lesson for zero setup.
§7
Watch
canon · alternating off yesterday's fresh
Barbarians at the Gate — the best thing ever written about how a deal actually gets done
Book (Burrough & Helyar) or the 1993 HBO film (~1h45) · reads like a thriller, not homework
Yesterday ran fresh (YC's RFS shorts), so today the canon. The pick is set by the week's other theme — the one
the §0 math story crowded out. Dealmaking was everywhere: Paramount
Skydance just closed a $111B merger with Warner Bros. (§9), the Bending
Spoons CEO explained why most M&A fails, and Butterfly Effect raised half a billion dollars (§2).
Barbarians at the Gate is the RJR Nabisco leveraged buyout told move by move — the egos, the leverage,
the rooms getting controlled. If you binged Entourage for Ari working a room, this is the forensic, real-money version
of the same craft: how people with no asset but relationships and nerve move billions. The HBO film is the fast way
in; the book is the one that stays with you.
Why it earns a slot over news: the mechanics of a deal don't expire, and you've told me this is the skill you want to
train. Watch for one thing specifically — how each player manufactures leverage out of information and
timing, because that's the transferable part, whether the number is $25B or your first $25K.
§W
The Wildcard
8 strange things, off the tech wire — full pull today
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.
The world's most-played game escaped the USSR through a tangle of handshake deals, double-crosses and Cold War suspicion. A dealmaking thriller hiding inside a toy — who owned falling blocks was a genuine international mess.
She coined the term “software engineering” when it was a joke, and her team's code caught the error that could have aborted the first Moon landing. The famous photo of her beside a stack of her own printouts taller than she is is still the best image of the craft.
A bracing argument that roughly a million deaths a year from mosquito-borne disease is not fate but a decision — we have the tools to eradicate the species and mostly choose not to. “You can just do things” applied to a whole genus.
The “cloud” is a few fragile cables on a seafloor, and a regional war can physically move where the world's data flows. A reminder that the most abstract infrastructure we have bottoms out in very cuttable glass.
A 17th-century alchemy manual told entirely in pictures — no words at all — so the secret could be passed to initiates and withheld from everyone else. Deliberately obscure information design, three centuries before the term existed.
Jonathan the tortoise is the oldest land animal alive, hatched around 1832, and scientists are finally reading what his cells did differently. Longevity as a slow, stubborn fact rather than a startup pitch.
A map of where the fonts you read every day are actually drawn — heavily the US and a few surprising clusters elsewhere. The geography of an invisible craft, for the part of you that cares why a letter is shaped the way it is.
Around 1200 BC an interconnected world of empires fell apart almost at once, and we still argue about why. A case study in how a tightly-coupled system — trade, tribute, specialised supply chains — can cascade into collapse. Uncomfortably legible today.
§8
Deals
1 free resource worth your time
Audit the best free AI and Python courses for 2026 — and the trick to taking paid ones for nothing
Free · Class Central (back today after timing out yesterday)
No discount-code noise this time — just free education, curated. Class Central refreshed its roundups:
14 best AI courses for 2026 (split by beginner vs
coder) and 10 best Python courses for AI and data.
More useful than either list is the mechanic underneath it:
how to audit Coursera courses for free and
the 8,000+ MIT/Harvard OpenCourseWare courses that are
simply free. For someone whose whole strategy is off-campus, branch-agnostic proof of work, this is the cheapest way to
close a specific skill gap — 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
Paramount Skydance completed its $111B merger with Warner Bros. Discovery. A media behemoth just formed. Filed as the live backdrop to §7's Barbarians at the Gate: the mechanics of how a deal this size gets done — leverage, egos, timing — are the exact thing worth learning, now playing out in your news feed.
The AI boom is making the world's cheapest smartphones disappear. Data centres are buying up memory chips, and the shortage is pricing the cheapest phones out of existence — so the compute buildout has a distributional cost at the very bottom of the market. The AI story has a downstream you won't see from San Francisco but can see from where you are.
Rosalind Franklin didn't miss the helix in her own DNA image. The tidy story that Watson and Crick “scooped” a scientist who didn't understand her own data is wrong, and historians have the notebooks to prove it. A clean lesson in how a convenient narrative can outlive the facts — and why primary sources beat the myth.
Meta and Microsoft are moving to curb employees' use of Claude. Two giants steering staff off a rival's model is a small, telling data point on how the model wars are being fought inside big companies — not just on benchmarks, but on procurement and policy. The tooling layer you'll build on is also a battleground.