In partnership with
In Today’s Issue:
🇰🇷 Seoul makes AI free for every citizen
🎓 MIT says AI can pass most of college
🦞 OpenClaw 2.0 shipped itself, almost by accident
🍎 The Mac mini is suddenly AI's favorite computer
📊 Nvidia's $96 billion quarter, explained
✨ And more AI goodness…
⚡ The Signal
The AI economy has split into two models: in America intelligence is a subscription, in Korea it just became a public utility.
On Friday, Seoul picked SK Telecom, KT and Kakao to hand every citizen free, unlimited AI agents, backed by state GPUs and a mandate that half of the workload runs on homegrown models. The same week, Nvidia reported a $96.2 billion quarter and MIT concluded that AI can credibly complete most undergraduate assignments. The pattern is hard to miss: AI is turning into infrastructure, someone has to pay for it, and governments have started answering the question of who. Watch whether "free and unlimited" survives contact with real demand; today's dustup over Claude's usage caps shows how expensive those two words are.
All the best,
Kim Isenberg
(Forbes)
🎓 MIT Says AI Can Now Pass Most of College
An MIT committee has concluded that generative AI can now credibly complete most undergraduate assignments, and is calling for profound changes to how the university teaches. The report urges instructors to move toward oral exams, portfolios and in-person assessment, and asks MIT to make course redesign faster while rebuilding the community life AI is eroding, from study groups to office hours. President Sally Kornbluth called the moment a watershed for higher education: "it's imperative that we get this right."
👉 tl;dr: If AI can do the homework, the homework has to change. MIT just said so, officially.
(OpenClaw)
🦞 OpenClaw 2.0 Shipped Itself, Almost by Accident
OpenClaw, the open-source platform for building personal AI assistants, released version 2.0 on Sunday: 933 contributors, 569 of them first-timers, and more than 16,000 pull requests in seven weeks. What began as a cleanup of installation and the browser experience cascaded through the whole codebase into a full rewrite, one the team says it built largely with OpenClaw itself. The release plugs into your existing ChatGPT or Claude subscription, makes the browser the main home for your "Claw," and adds shared cloud sessions so teams can run agents together.
👉 tl;dr: Personal AI agents are leaving the hobbyist phase: one workflow first, then your digital life, now your whole team.
(Apple)
🍎 The Mac Mini Is Suddenly AI's Favorite Computer
Apple has stumbled into an AI hardware boom, The Information reports: OpenAI has bought tens of thousands of Mac minis and Mac Studios to train its computer-use agents, and Anthropic rents Macs through Amazon's cloud. The boxy desktops win on shared memory and cooling, and demand ran so hot that Apple pushed out new M6 Mac mini and M5 Ultra Mac Studio models last week, well ahead of its usual autumn refresh, pitching clustered Mac Studios for running giant frontier models. The catch: a historic memory-chip shortage has kept the top configurations out of stock for months, while Nvidia courts the same customers with its Mac-mini-lookalike DGX Spark.
👉 tl;dr: Nobody at Apple planned this. An accidental enterprise product is now one of the hottest boxes in AI.
🎬 Watch This
Long before ChatGPT, Zoubin Ghahramani spent three decades building machines that know what they don't know. Now a Cambridge professor and VP of research at Google DeepMind, he sits down on DeepMind's own podcast, "The mathematics of AI uncertainty," to argue that calibrated doubt, his life's work, may be exactly what today's overconfident models are missing. The stretch on correctness versus confidence is the clearest explanation you will hear of why chatbots state wrong answers so fluently, and his take on whether Bayesian thinking is the missing piece for ever-improving AI is worth your time
The "20x" on Anthropic's $200 Max plan is now a federal case: a proposed class action in California alleges the $200 tier delivers only about 1.1 to 1.7 times the weekly usage of the $100 plan, because the advertised multipliers apply per five-hour session, not to the undisclosed weekly caps. OpenAI's Codex side pounced, noting its own 20x applies to the weekly limits themselves.
(U.S. District Court filing via @kimmonismus on X)
Free AI for Every Korean: Seoul's "AI for All" Gamble
The Takeaway
👉 Korea's science ministry picked SK Telecom, KT and Kakao to build free, unlimited AI services for every citizen and resident; beta tests start in September, nationwide rollout is planned by December.
👉 The state supplies 512 Nvidia B200 GPUs this year and subsidizes running costs from 2027; at least half of every service must run on homegrown Korean models.
👉 The services are built as task-completing agents: booking hospital appointments, navigating public services, taxes and housing, inside KakaoTalk and even by phone call.
👉 The stated goal is an "AI-basic society"; the unstated one is independence from American and Chinese AI.
In South Korea, access to AI is about to come with residency rather than a credit card. The science ministry on Friday selected three consortia, led by SK Telecom, KT and Kakao, to build AI for All: free, unlimited AI services for every citizen and resident, with beta tests starting in September and a nationwide rollout planned by December.
(Korea Herald)
Each winner brings its own distribution. Kakao wires agents into KakaoTalk, the country's dominant messenger, and plans a marketplace of specialized agents; KT plugs into portals and shopping and housing platforms for end-to-end tasks; SK Telecom adds an assistant that will even place phone calls to offices with no digital interface. The state's side of the bargain: 512 Nvidia B200 GPUs this year, operating subsidies from 2027, and a rule that at least half of every service runs on domestic models such as LG's K-EXAONE and SKT's A.X, all under a 2026 AI budget of roughly 10 trillion won ($7.2 billion), triple last year's.
(Ministry of Science and ICT via Trending Topics)
South Koreans were already unusual customers: they spend roughly 1 trillion won ($6.8 billion) a year on generative-AI subscriptions, more than the country spends on Netflix. The catch is scale. 512 GPUs split three ways is a rounding error next to frontier clusters, so "unlimited" will lean heavily on the smaller domestic models, and routing health, tax and housing chores through state-subsidized agents hands the government a sensitive new role. The ministry says essential functions stay free even when demand outruns state-funded capacity; December will test that promise.
Why it matters: Korea's service costs its citizens nothing; it is the $20-a-month subscription the rest of the world pays for ChatGPT and Claude that suddenly looks like a choice rather than the natural business model of AI. Governments can make intelligence a utility, and decide whose models power it.
Sources:
🔗 https://www.techspot.com/news/113664-south-korea-giving-entire-population-free-access-ai.html
🔗 https://www.koreaherald.com/article/10855962
🔗 https://www.trendingtopics.eu/south-korea-free-ai/
Winning Startups Aren't Bigger. They're Leaner.
Founders deploying AI agents across sales, marketing, and customer service are closing more with smaller teams. 65% more sales leads. Less headcount.
Download the free Practical Guide to Agentic GTM for Startups and start building the stack your competitors dream of.
(Epoch AI, CC-BY)
The chart: Epoch AI tracks computing infrastructure as a share of US GDP, quarterly since 2015. The teal base is ordinary compute spending; the colored wedge on top is the AI boom: 0.56% of GDP in above-trend compute hardware, 0.15% in data center construction and 0.12% in networking gear. Together the sector now runs at about 1.5% of GDP, double its 2015 to 2022 average of roughly 0.7%.
The lesson: Nvidia's $96 billion quarter and Seoul's state-funded GPUs are the same phenomenon seen from two ends: AI spending has outgrown tech budgets and now moves national accounts. Nearly a full percentage point of the world's largest economy has been redirected in about three years.
The caveat: The chart measures spending, not returns. The dashed line shows where the old trend would have landed; everything above it is a bet that has yet to pay off. If AI revenue disappoints, that 0.8-point wedge is exactly what a reversal would have to give back.
💰 One Company's Quarter Is Now the AI Economy's Heartbeat
⚡ Bottom line: Nvidia reported $96.2 billion in quarterly revenue on August 26, up 106% in a year, and guided next quarter to about $108 billion.
💡 Why it matters: Nine of every ten of those dollars come from AI data centers, so this one income statement is the cleanest live reading of the whole AI buildout.
🔎 What it means: As long as the number keeps climbing, the boom is funded; the first quarter it stalls, everyone finds out together.
The simplest way to read the AI economy is one company's sales line. For the three months to July 26, Nvidia took in $96.2 billion, more than double the same quarter a year earlier, and told investors the current quarter will land around $108 billion. Jensen Huang's framing was blunt: "AI has reached its inflection point. It's doing useful work. Its tokens are productive and profitable. Now, compute is revenue."
(Getty via TechCrunch)
Where the money comes from is the striking part. $89 billion, about 92% of the total, came from the data center business: the chips that cloud giants such as Amazon, Meta and Google buy by the rack to train and run models. The new Vera Rubin platform is ramping into full production across those clouds, and Nvidia still handed $26 billion back to shareholders in the same three months.
(Chart: Superintelligence; data: NVIDIA)
The guidance is the tell. That $108 billion assumes zero data-center sales to China; Nvidia is forecasting acceleration while writing off an entire superpower's market.
The uncomfortable part is concentration. One firm, selling mostly to a handful of cloud buyers, now sets the tempo for the whole AI economy, from Seoul's subsidized B200s in today's Featured Story to every startup's compute bill. Its earnings day has quietly become what the Fed meeting is for bond traders: the moment everyone recalibrates at once. The number to watch next: whether that $108 billion actually lands.
10x the context. Half the time.
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.
