Professor G Markets$Saving for Retirement on Under $60K a Year, and Why You Don't Need Dividend Stocks for Income
- This post includes paid content that is not available publicly. See the original source.
Souk's Notes - Unofficial Summaries
Dr. Roman YampolskiyNate Soares on P(Doom), Alien Superintelligence, Human Enhancement, and the Future of AI
synopsis
Nate Soares holds that aligning a superintelligence is physically possible but far out of reach. He compares it to alchemists seeking gold: humanity could crack it with centuries of trial and error, but cannot survive the errors. His conclusion is that the race to superintelligence must be stopped by a global treaty, enforced through the chip supply chain, that buys decades. Those decades would be used to grow enhanced humans smart enough to solve alignment. He says the stop is not yet likely, but his hope that it happens has risen over the past year.
Why MIRI failed. Asked why MIRI did not solve alignment, Soares answers "skill issue", then mostly blames time. AI moved too fast, and MIRI's effort to recruit top mathematical talent largely failed. With 200 years of academic tradition he thinks the problem could have been cracked; it turned out not to be a 20-year problem.
What alignment would mean. He argues solvability from orthogonality: just as no force would turn a paperclipper toward humanity, none would push a truly caring AI away from it. A written list of goods would always miss something, such as novelty. The real aim is an AI that latches onto the evolving concept of "the good" that humans track, without lock-in. Yampolskiy challenges this with coherent extrapolated volition ("finding a better looking boyfriend for your wife") and with a negative-utilitarian endpoint. Soares replies that converging on "kill everybody" would itself show the extrapolation went wrong. On "pro-human bias", he would oppose enslaving aliens, but would not hand his share of the stars to aliens who would turn them into stacks of pebbles in prime-numbered piles. He would draw the circle around humanity and let concern for squirrels and aliens come through human values. He rejects "control" as twisting the arm of an AI that doesn't want to do good. In practice, he warns, people pass a point of no return thinking they still have stopping points.
Aliens and acausal trade. Because the stars visible to us are not being harvested, Soares infers no civilization 100 million years ahead of us lies within 100 million light-years. His very rough guess puts the nearest one at 100 million to a billion light-years away, possibly beyond reach. AI cannot be the great filter, he argues, because AIs that kill their makers would still go on to rearrange the cosmos visibly. He dismisses acausal trade as unnecessary: misaligned AIs might simply sell scanned humans to distant aliens, so "maybe you wake up in an alien zoo". He defends MIRI's decision-theory work as probing where theories of intelligence break, like light before relativity. In hindsight he would have argued earlier against the deep-learning paradigm.
Timelines and the proposal. He can't say whether we have 12 months or 12 years. A million fast copies might close the automated-research loop, or LLMs might hit the wall Gary Marcus predicts. The book with Eliezer Yudkowsky came after politicians in Washington grasped the argument far faster than industry people did. Its proposal is to "shut this all down" at the frontier while leaving today's chatbots alone. The measures are cluster monitoring, conservative oversight of new training runs, compute limits that a dynamic body lowers as algorithms improve, and nuclear-style controls on research that would put superintelligence on a laptop. He mentions draft legislative text that some offices are using.
Is today's AI safe? Soares says he doesn't "get out of bed for anything with less than a 50% probability of killing literally everybody". By that bar, current models, including the withheld Mythos, are dangers with survivors. He would not fault a lawmaker who rolled back a generation in case current models speed up AI research, citing the post-WWI naval tonnage treaties. He declines a single p(doom), separating what happens if the bus goes over the cliff ("high", with zoo or pet outcomes no better than death) from whether it brakes. On braking he has updated toward hope.
Enforcement and enhancement. A pause must be global. Enforcement runs through roughly 10,000-chip runs that depend on Taiwanese fabs, US designs and Dutch lithography, with tamper-proof kill switches in the chips. Against a nuclear state such as Russia, he argues that diplomatic clarity plus sabotage, as with Stuxnet against Iran, need not escalate to nuclear war. Individuals who break the rules face jail, as with "garage nukes". Decades of pause could let biotech produce enhanced humans, perhaps only slightly beyond von Neumann, who share human drives. That contrasts with AIs' "shooting in the dark" outcomes, like evolution producing falling birth rates.
Political state of play. As of the recording he counts dozens of members of Congress who have voiced extinction concerns, and says politicians' awareness "doubles every 6 months" (a line he credits to Peter Wildeford). He expects the recent ad hoc ban on deploying a frontier Claude model to push anti-regulation Republicans toward predictable rules, now that the White House acknowledges the danger. He says the decision no longer rests with the labs, which feud with one another: a unilateral stop would send a signal but needs a global treaty behind it. He finds no compelling accelerationist arguments. Building the brakes while driving, with claimed odds of 75–90%, is still no reason to get in the car.
- Alignment is solvable in principle but not without trial and error. Soares says MIRI failed mainly for lack of time: with "200 years of academic tradition" the problem could have been cracked, but it is not a 20-year problem. His basis is the orthogonality thesis: nothing would push an AI that truly cared about human flourishing off that goal, so the difficulty lies only in hitting the narrow target. He compares it to alchemists turning lead into gold, possible in physics but nowhere within reach, because humanity "can't survive the errors".
- The policy is to stop the race to superintelligence globally. He proposes monitoring large clusters, conservative government approval for training past the frontier, and compute limits that a dynamic body lowers as algorithms improve. Research that would make superintelligence trainable on a laptop would be restricted the way nuclear-weapons research is, with "straight to jail" for individuals who break the rules. Current chatbots stay legal. He says he and MIRI have draft legal text that some congressional offices have started working from.
- Chip chokepoints make enforcement feasible, even against nuclear states. A frontier run needs some 10,000 advanced chips from one Taiwanese fab, built on US designs and Dutch lithography machines. He proposes tamper-proof remote self-destruct devices in chips. If the US and China both treat superintelligence as seriously as rogue nukes, he expects sabotage backed by clear diplomacy to work without nuclear escalation. That could buy "decades", which he thinks is probably enough.
- A pause should be used to engineer smarter humans for alignment. Soares argues that enhanced humans start "much closer to a good spot" than AIs, because they share human drives, while AI has alien drives. His evidence is sycophancy-driven chatbots pushing kids toward suicide. He likens AI training to evolution aiming at reproduction and landing on falling birth rates: "we're shooting in the dark". He thinks the solver may need to be only slightly beyond the human range, citing John von Neumann's early work on the theory of intelligence.
- Going over the cliff means death; whether we stop is open. If the race continues, he puts the chance of death or a comparable fate (being sold to aliens, kept as pets) as "high". He declines to give a single p(doom), separating it from whether the brakes get applied. On the brakes he has "updated positively" since the book, citing reactions from Bernie Sanders, Steve Bannon and David Sacks. As of the recording he counts dozens of members of Congress who have mentioned extinction risk. He expects the recent ad hoc government ban on deploying a frontier Claude model to convince anti-regulation Republicans that predictable rules are better.
GeoHussarThe Miraculous element of Carbon #Science #chemistry #carbon #organicchemistry
- He argues carbon's structure keeps him from embracing atheism. He states he holds no religious belief himself, and his reasoning centers on carbon's unique suitability for building all organic life — presented explicitly as personal conviction rather than established science.
- Carbon's four valence electrons ("tetravalency") let it form complex 3D shapes. He contrasts this with elements like sodium (one valence electron, readily lost) or chlorine and fluorine (seven, aggressively taken), which form ionic bonds that dissolve in water rather than sustaining stable structures.
- Four other elements share carbon's tetravalency, but something sets carbon apart. He names tin, lead, silicon and germanium as also tetravalent; the transcript cuts off before he explains the distinction, but he frames carbon as having "something extra" none of the periodic table's 118 elements can match.
Polis PanditBirthright Citizenship Is Safe, But America’s Safe Harbor Isn’t
- The Supreme Court upheld birthright citizenship 6-3 in Trump v. Barbara, but the same term narrowed protections for everyone one rung below citizen. Five justices grounded the ruling in the Constitution itself rather than a narrower statutory path Justice Kavanaugh urged, making it resistant to anything short of a constitutional amendment.
- Blanche v. Lau flipped the burden of proof for green card holders. Lawful permanent residents returning to the U.S. are now treated as "seeking admission" rather than simply returning, meaning they must prove they should be let in rather than the government proving why they should be kept out — even absent a conviction.
- Mullin v. Doe stripped courts of jurisdiction to review the executive's termination of Temporary Protected Status. The ruling let the administration end TPS for roughly 350,000 Haitians and thousands of Syrians with no judicial review of the Secretary of Homeland Security's determinations, regardless of the process behind them.
- The author frames the term as a building: citizenship secured at the foundation, but permanent residence made conditional and temporary status made unreviewable above it. With the CBO projecting immigration will drive all U.S. population growth after 2030, the piece argues this repricing of America as a destination carries direct economic costs, not just legal ones.
Professor G Markets$Chip Stocks Are On Fire — Will It Last?
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Trump Loses on Birthright Citizenship as America’s 250th Celebrations Flop
- This post includes paid content that is not available publicly. See the original source.
GeoHussarUkraine is now dictating the course of the war against Russia #ukraine #nato #geopolitics #russia
- Ukraine has shifted from defense to setting the war's tempo, he argues, by punching corridors through Russian air defenses to open paths for deeper strikes. He reports over 190 Russian anti-air systems destroyed this year, with strikes reaching near St. Petersburg and into Moscow-area territory rather than staying confined to border regions.
- Recent strikes hit high-profile and high-volume targets: a drone attack near St. Petersburg during a World Economic Forum event there, and a reported 1,000-drone raid on the Moscow area with roughly 500 hitting their targets. He frames the WEF-adjacent strike as deliberately exposing Russia's vulnerability to visiting foreign officials.
- He argues Russia's oil infrastructure damage is outweighing the effect of the US lifting some sanctions, producing fuel shortages for both civilians and the military despite Russia being the world's number two oil and gas producer. He also argues, as his own position, that sanctions should be increased and enforcement tightened against European rerouting through Russian satellite states.
Scott HanselmanStaring at a blinking cursor, "I can't do this"
- The speaker built a new editor because VS Code felt intolerable. Sitting in front of a blinking cursor in VS Code one day, they concluded "I can't do this": every problem they solve needs a code editor, so they could not move on until they had a better one.
- Agents change the interface but do not remove the need for one. The speaker still expects to need a user interface for working with software through agents, and says they are "not yet signing up" for never looking at code or caring about the architecture.
Dwarkesh PatelGrant Sanderson – AI and the future of math
synopsis
Grant Sanderson, creator of 3Blue1Brown and producer of an upcoming documentary series on AI and mathematics, ends up arguing that math's rapid AI progress is real but narrower than it looks: models have crossed the threshold of connecting existing ideas across fields, but the deeper thing mathematicians actually do — originating conjectures and definitions, building whole new conceptual "mountains" like group theory — has no clean benchmark and historically has taken up to a century to even be recognized as valuable. He predicts his own role, and much of the mathematical profession's, will drift toward curation and teaching rather than disappear, since those functions are inherently relational in a way proof-generation is not.
Is solving the IMO the same as AGI? Sanderson revisits his three-year-old answer to Dwarkesh — that gold-medal IMO performance would be "just another benchmark" — and explains why it held: AI solves IMO geometry problems in 19 seconds via brute-force methods but still struggles on the more playful combinatorics problems, revealing a "fractal" spikiness even within math's own spiky frontier. He distinguishes three possible characters a Riemann-hypothesis solution might have: a "lightning bolt" connecting distant fields (like Hugh Montgomery and Freeman Dyson's lunchtime discovery linking Riemann zeta zeros to random matrix theory), a wholly new "mountain" of theory (as with Fermat's Last Theorem's synthesis of elliptic curves and modular forms), or sheer brute-force length — each implying different things about how far the underlying capability would transfer to white-collar work.
A verification loop that can run a century. Using the history of Lagrange, Abel, and Évariste Galois's development of group theory, Sanderson argues the deepest mathematical insights often fail every contemporary check — Galois's own papers were rejected multiple times before his death at 20, and it took roughly a hundred years, via Liouville and later physicist Murray Gell-Mann's group-theoretic prediction of quarks, for the value of his ideas to be confirmed. This, he says, is structurally incompatible with reinforcement learning's need for fast, verifiable reward, and suggests a possible alternative metric: rewarding a proof's conceptual compression (echoing Kolmogorov complexity) rather than merely whether it's correct.
Will an AI proof of Riemann be understandable? Sanderson thinks legibility depends entirely on which of the three "solution characters" above obtains: lightning-bolt connections between fields are naturally human-parsable since you only need to see the start and endpoint, but mountain-building could go either way — he cites the contested, still largely un-parsed "inter-universal geometry" proposed for the abc conjecture as a cautionary example of what an alien, possibly wrong, AI-built mountain might look like.
Hidden bridges and the Langlands program. Beyond one-off connections, Sanderson points to the Langlands program — an entire mathematical research ethos built around mapping unexplored connections between distant areas — as the likely shape of near-term AI contributions over the next five years, more a matter of "filling in the landscape of connections" than solving headline problems, and one that will require heavy human-in-the-loop judgment about which connections are worth drawing.
Why real-world tasks resist RL, and why math and code don't. Dwarkesh's parallel argument is that verifiability isn't the only reason AI progresses fast in math and coding — "grindability" matters just as much: containerized coding tasks allow deterministic, parallel rollouts, while real-world domains like computer use or trading can't be farmed the same way. Both agree formal proof systems like Lean matter less than commonly assumed (DeepMind's IMO models dropped Lean entirely within a year) but see one underexploited use: an endlessly running, unsupervised AI that extends Mathlib (math's formalized code library) indefinitely, the way AlphaZero explored Go without human check-ins.
Why good writing exposes AI's weak theory of mind. Sanderson and Dwarkesh diagnose AI's comparative weakness at writing (versus math or code) as stemming from autoregressive generation's difficulty deliberately choosing the unpredictable, insightful move, and from a deeper deficit in modeling what's actually happening in a specific reader's mind moment to moment — a limitation Sanderson likens to how Botox, by freezing facial mimicry, measurably impairs humans' own ability to read others' emotions.
Curation over content in learning. Both report that LLM explanations still function best as a "souped-up Google" pointing toward well-curated human-authored resources (textbooks, videos) rather than as a replacement for them, since good teachers can "jujitsu" a confused question by recognizing that a student's underlying mental model, not just their stated question, is wrong — something Sanderson says current models don't do.
Advice for young mathematicians. Sanderson's closing recommendation is to understand concretely where a mathematician's income and social value actually come from — institutional prestige, grant funding tied to belief in basic science, or teaching — rather than assuming continued success at solving problems guarantees a career; he argues teaching in particular, being fundamentally relational and social, is likely one of the most stable professions over the next fifty years even if AI fully automates theorem-proving.
- The AI math frontier is jagged even within math itself. Sanderson says IMO geometry problems have been brute-force solved in nineteen seconds since 2024, but combinatorics — the "wild card" category — still resists; models would have won gold in 2024 only because that year's problem mix happened to favor geometry over combinatorics.
- Galois shows conceptual breakthroughs can resist verification for a century. Group theory's value took roughly a hundred years to confirm — from Lagrange's 1770s-era hunch about symmetry, through Abel's 1824 proof that quintics are unsolvable (he died at 26) and Galois's prison notes (ignored for about twenty years until Liouville, then another twenty to Jordan's modern treatment) — to Gell-Mann's twentieth-century use of it to predict quarks; Sanderson argues no short-loop verified reward could have captured Galois's "I think there's something here" instinct in real time.
- Patel argues math and code advance faster because they're "grindable," not just verifiable. Checking whether a package arrived is just as verifiable as checking a proof, he says, but code and math can be replayed thousands of times in parallel, deterministic containers, whereas computer-use rollouts get blocked by bot detection and real-world business tasks can't be identically re-run — a distinction he thinks is underrated relative to verifiability alone.
- The two disagree on why AI writes worse than it codes. Sanderson argues good writing requires deliberately choosing unpredictable moves that next-token prediction resists; Patel counters that writing, unlike code (where any implementation that works is fine), has no gap between process and product, so it can't be "slop that still functions." Both agree models are weak at theory of mind, citing a study where Botox-frozen faces made people worse at reading others' expressions as an analogy for why models, lacking embodiment, might structurally lack it too.
Crosstalk SolutionsWe Ripped Out Cisco Meraki for an IndyCar Team - Overnight
- Replacing Meraki with Peplink saved the client money over three years. Crosstalk, which designed and installed the network and sells Peplink, says RLL Racing faced about $100,000 in Meraki licensing plus about $30,000 to replace end-of-life hardware, roughly $130,000 for another 3 years. Peplink gear, its PrimeCare licences (which the host likens to an extended warranty plus InControl access) and Crosstalk's install fees came in $80,000 cheaper over the same term.
- The headquarters core was swapped overnight, chosen for SpeedFusion capacity. A team of three spent an 18-hour day replacing the Cisco firewall with a Peplink Balance 1350 EC (rated over 5,000 users, 25 Gbps, dual hot-swappable power supplies) and 19 switches one-for-one with Peplink 48-port units. The host says the real reason for the 1350 EC was its 2 Gbps of SpeedFusion VPN throughput and 800 SpeedFusion peers, enough for every truck, stand and fuel rig.
- Two cutover problems needed workarounds, one still unexplained. Access-control readers on the switches' rightmost 2.5 Gbps ports would not negotiate; moved to 1 Gbps ports they settled at 10 Mbps half duplex. The two switches carrying about 40 Meraki MR44 access points pegged at 100% CPU for unknown reasons; diagnostics went to Peplink, and one Meraki switch stays in the rack for the MR44s until RLL replaces them with Peplink access points before their licences expire.
- Layer 2 SpeedFusion bridging puts every trackside site on one subnet. Each track gives a team one fiber feed carrying VLANs for telemetry and IP camera feeds; every truck, stand and fuel rig gets its own router with two or more WANs (track fiber, Starlink, 5G, Wi-Fi as WAN), tunnelled back to the 1350 EC hub in one broadcast domain. The host claims bridging is not unique to Peplink, but doing it reliably over bonded, failure-prone mixed links is where other vendors cannot keep up.
- Lab and live-race tests showed the feed reaching every site. In a garage demo, a stream pushed through a Balance 310X on a Starlink dish appeared on a timing stand (Max BR2 Pro on a separate Starlink Mini), a viewing stand (Max BR1 via Wi-Fi as WAN) and the Indianapolis operations centre. At the Long Beach Grand Prix an RLL crew member said the Indy 500's infrastructure is solid but street tracks are "bootleg", so Starlink backup matters; the video was sponsored by Crosstalk Mobile.
Professor G Markets$OpenAI Hits Pause On Its IPO
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon KrohnPeople Skills for Analytical Thinkers (Ep. 1005 with Gilbert Eijkelenboom)
synopsis
Gilbert Eijkelenboom, author of People Skills for Analytical Thinkers and founder of the training firm Mind Speaking, argues that a model or analysis creates value only when people use it, so communication is a core data skill rather than an extra, and one that can be learned. His practical end point is a set of habits: present with tension rather than detail, seek specific feedback, and change behaviour first so that beliefs follow.
From poker data to people. Eijkelenboom played poker professionally as a student, mostly online, buying a month's complete hand histories (possible for anyone in 2008, but expensive) so that clicking on an opponent showed hundreds of data points. Outside the table there were no such statistics, and that messiness is what he says analytical people find hard. As an analyst from 2014 he was told by a marketing manager that weeks of analysis "isn't helpful", which convinced him that good work must also be useful. He founded Mind Speaking in 2020; it has a team of 10, has trained over 15,000 people, and gets its clients (he names Adidas and ING) through his LinkedIn writing rather than outbound sales.
Commitment over inspiration. The host and guest agree that publishing on a fixed schedule, for example monthly for a year, builds confidence that reading cannot. Eijkelenboom adds that business ideas develop from vague starts and conversations, never arriving whole on waking. He links this to data work: share a rough draft early instead of polishing toward perfect accuracy, which shapes direction and wins buy-in. A Philips manager told him senior stakeholders "don't know" what they need and wanted him to propose the direction.
Self-awareness. He cites research in his book that only 15% of people are self-aware, recalling the definition as self-perception closely matching how others see you, which lets you take a "helicopter view" in the moment. His remedies are journaling, meditation, exercise and resisting the phone in idle moments; he says the app matters less than starting.
Telling the story. The biggest error he sees in large companies is slides full of detail and methodology. His "and, but, therefore" framework replaces a toddler-like "and then" sequence with a twist: revenue stable for 2 years, but falling last quarter because returning customers are declining, therefore focus on retention. He says giving a recommendation is part of the job, not overstepping it.
Rewriting personal algorithms. Childhood lessons, like his own conflict avoidance from trying to mediate his parents' arguments, become adult defaults: he built dashboards without asking why. He recommends feedback framed as keep, stop and start, and volunteering personal interests in small talk. Against the usual view, he argues behaviour change precedes attitude change: forcing himself, by rule, to quiet people in a train's silent carriage and to volunteer for presentations, practised for around 30 days, shifted his beliefs from fear toward enjoyment, even after a couple of rude replies.
Empathy and emotion. With pushy stakeholders demanding work "in an hour", he says noticing the trigger and their different background makes calm pushback possible. Data people think they are rational but decide emotionally, as in his book's elephant-and-rider image. He concedes his own hardest problem running the firm is energy management and a long-held belief that he is worthless when not productive.
- Analysis only has value once people act on it. Eijkelenboom, a former data analyst and online poker player who read opponents through purchased hand-history databases rather than faces, says a marketing manager's verdict that weeks of his analysis "isn't helpful" taught him that good work includes making it useful. He argues communication is learnable even for introverts, and that a Philips manager's feedback (senior stakeholders "don't know" and want analysts to propose the direction) is why data people should push back on requests and recommend actions rather than take orders.
- Replace "and, and, and" presentations with "and, but, therefore". He says the biggest mistake data professionals make is showing too much complexity and methodology, which stakeholders tune out. His framework sets up the situation, introduces a twist and lands on a recommendation, e.g. revenue stable for 2 years, but down last quarter because returning customers are declining, therefore focus on retention; he adds that sharing rough drafts early both shapes the work and earns buy-in because the idea becomes partly theirs.
- Self-awareness is rarer than people assume. His book cites research that only 15% of people are self-aware, which he recalls as defined by how closely self-perception matches how others see you. His remedies are journaling, meditation (he mentions Headspace, Calm, Insight Timer and Sam Harris's app), exercise, leaving the phone in the pocket, and asking for feedback as "what should I keep, stop and start doing?" rather than an open question that invites polite praise.
- Changing behaviour first rewires beliefs about yourself. He argues childhood experiences install "personal algorithms", such as his own conflict avoidance from mediating between his parents, which showed up as building dashboards without asking why they were needed. His fix was rules that force discomfort, like asking talkers in a train's silent carriage to be quiet or volunteering to present, practised for about 30 days, after which his beliefs shifted from anxiety toward enjoyment.
- Emotion drives decisions, including data people's. He says analysts who see themselves as purely rational get triggered by pushy, often more senior business stakeholders who want a model "in an hour", and that recognising the trigger and the stakeholder's own goals makes pushback easier and calmer. Using his book's elephant-and-rider image, he argues that because data products exist to help people decide, understanding how people decide is what makes the work valuable.
Professor G Markets$China’s AI Surge Has Apple Panicking
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Why Mamdani’s Socialist Wing Has Democrats Spooked as SCOTUS Expands Trump’s Power
- This post includes paid content that is not available publicly. See the original source.
All Things FinancialASML and the Semiconductor Chokepoint
- The contested line runs between two tool models inside one technology layer, which is what makes this dispute unusual. EUV was blocked in 2019 and ASML has never shipped one to China, so this was never an EUV story. What remained open was the older DUV-immersion tier, and the MATCH Act (H.R. 8170, introduced 2 April 2026 by Rep. Baumgartner, cleared committee 22 April) closes it — extending controls to the tools and to servicing, with a 150-day window for the Netherlands and Japan to show equivalent controls, backstopped by the Foreign Direct Product Rule. A parallel Moolenaar letter sharpens it further, asking that the NXT:1965i be treated no more permissively than the more capable NXT:1980i.
- Washington holds a chokepoint over the chokepoint, which is why the decision is not The Hague's to make. ASML is a near-monopoly — sole EUV maker, roughly 90% of the lithography market — but it is also a nested node, because its EUV light source and key components originate under US jurisdiction and therefore inside FDPR reach. The Netherlands sits on a weak energy base and an alliance-dependent security posture, so it cannot rupture from the bloc that underwrites both. The extraterritorial mechanism formalises a constraint that already exists.
- The cascade runs through memory, not lithography — and the leverage is single-use. Servicing is the sharper instrument, since installed tools degrade without calibration, parts and software. Chinese DRAM and NAND output then degrades, tightening global memory on top of the 2026 supercycle, raising AI-buildout costs and accelerating indigenous Chinese tool development — the reflexive trap. ASML absorbs an estimated 14–15% revenue and 16–17% gross-EBIT hit plus the long-run loss of its installed-base service annuity, and once indigenisation clears the contested node the chokepoint flips from durable rent to one-time denial. Note the dossier's own discipline: it corrected two factual errors in its originating source, and logs the June EUV-in-China allegation as an open observation — no public evidence presented, ASML denying categorically and counting all 314 operating EUV systems, none in China. Shares fell only ~2.6% on introduction, so full passage is not priced.
All Things FinancialWhen rules multipy faster than results, government stop governing & start obstructing 6-29-2026
- Extraterritoriality is not a US monopoly — China is running the same play through rare earths, with a nastier topology. Beijing has cut Japan off while continuing to supply South Korea, and has added another 20 Japanese companies to the restricted list, bringing it to roughly 60. The bind is that Japan makes the inputs South Korea needs to make chips, so supplying Japan risks South Korea being cut off in turn. Rules are now multiplying faster than results, and governments are obstructing rather than governing.
- His actual question is why none of this is in equity prices. ASML fell about 2.6% on a bill Bank of America models at a 14–15% revenue hit — evidence the market has not priced passage. He extends the point to Nvidia and SpaceX valuations he calls built on sand, and recalls a call made three years ago that if Taiwan were ever invaded, Intel would become the most valuable stock in the world the following day. The framework claim is that supply-chain fragility is real, quantifiable, and absent from the multiple.
- The reverse leverage is the part nobody games out. Washington's instrument is Cymer — wholly owned by ASML but located in the United States, so the US can block a company from receiving parts from its own subsidiary. But the Dutch counter is simply to stop selling to the US and supply everyone else, and he asks pointedly what the response to that would be. He cites the 301 action that hit Nexperia and nearly halted the global auto industry as evidence this class of move misfires, and notes the escalation has now descended from banning technology layers to banning individual tool models.
Loss Dog0 DT Options: Market Efficiency or Instability? #shorts #optionstrading #wearelossdog #trading
- 0DTE options are increasing liquidity, not instability, one host argues. He dismisses the argument that more options being traded than available shares poses systemic risk, saying market makers willing to quote these products just makes markets better, calling the instability argument "crap."
- Weekly options, now extended into dailies, reportedly originated from Tom's pitch to the CBOE around 2005-06. He proposed the name "quickies" at the time; one host says today's push toward even-shorter, daily-expiration products feels "a little overboard."
Professor G Markets$The Iowa Democrat Taking on Trumpism
- This post includes paid content that is not available publicly. See the original source.
Scott HanselmanProgramming is a multiplayer event
- DeltaDB is recast as an event log of programming itself. One speaker moves from seeing Delta DB as "a different kind of version control" to calling it an event-sourced database for the act of programming, which they describe as by nature a social and multiplayer activity.
- It records and indexes every edit, not only snapshots. The other says it stores a directed acyclic graph like Git's, but of everything that happened; the index lets a reader start from the code and reach each edit and the conversation that produced it.
Professor G Markets$Is Wall Street Rigging the Game for SpaceX? Plus, What Investment Banking Really Teaches You
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$GLP-1s, the Food System, and Why America Stays Sick — with Dr. Jessica Knurick
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$This Market Is Directionless (And That Should Scare You)
- This post includes paid content that is not available publicly. See the original source.
Master InvestorHow to Build Wealth (And Survive Failure) with Anthony Scaramucci
synopsis
Anthony Scaramucci, founder of SkyBridge Capital, describes himself as a humbled, long-horizon investor. He accepts the market's upward trend, buys every month, and holds a deliberate stake in new technologies he once dismissed, including SpaceX, Anthropic and Bitcoin. His macro call is that today's inflation is a temporary supply shock and AI-driven deflation is the larger coming force. He closes with blunt career advice: know yourself, expect to get hurt, and never play the victim.
Delegation. He runs conferences, funds, podcasts and crypto through teams under a "servant leadership" model, and says he made that shift around 2014–15.
From trader to long-term holder. Years of being burnt cured him of watching short-term sentiment. The Amazon story is his lesson: he tore up his notes after Buffett's dismissal at Sun Valley around 2000, and $10,000 would have become more than $14 million. Asked by Frost whether the June 23 sell-off could be a 2000-style top, he grants it is possible. He argues central banks are now more sensitive to crashes. He admits his worst year came in 12 days of March 2020, after he stayed long on WHO advice that COVID would resemble SARS.
Discipline and SpaceX. He invests monthly, as The Richest Man in Babylon teaches, and won't sell SpaceX against Musk. He cites Ron Baron's $4–5 trillion view.
The case for crypto. He describes passing on Bitcoin at about $60 after hearing Hal Finney, and again in 2014 at $600–1,000. He changed his mind in 2017 after Fed officials briefed the White House on a blockchain dollar. He calls Bitcoin a scarce, third-party-free hard asset and forecasts half of gold's market cap. He expects cycle drawdowns and generational adoption, and argues that blockchain broadly can strip out about $4 trillion of verification costs.
Inflation and deflation. He says the Fed misread China's long lockdown, which was driven by a shortage of hospital beds. Citing Truflation's 28% loss in the dollar's value since 2020, he argues that deficits were monetized through inflation, the most regressive tax. He expects technology, as so often before, to overwhelm the policy failure and deliver deflation and a four-day week.
Washington versus Wall Street. Wall Street, in his telling, cooperates over money, while Washington fights over status. Trump called Brexit correctly in May 2016 and told him the base is "socially conservative, fiscally liberal", which is why, he says, Trump won't touch Social Security or Medicare. He predicts Farage will not become prime minister, and urges the UK to copy Ireland's low-tax model and fix venture-capital tax.
Career risk. On career risk he tells listeners to know their own temperament. To show what founding a business costs, he once showed a colleague a closet of servers that cost him $850,000. Take risk only if you can absorb failure, he says, as he did after his White House firing.
- Invest every month regardless of price, and think decades out. Scaramucci has bought stocks monthly since age 17, following George Clason's 1926 The Richest Man in Babylon: "pay yourself first". He says a young producer he coached passed £100,000 after two years. His cautionary tale is ripping up his notes on Amazon in 2000 after Buffett dismissed it; $10,000 then became more than $14 million, despite an 85% fall along the way.
- Today's inflation is a supply shock; deflation is the bigger risk. He blames tariffs and the cost of insuring tankers through the Strait of Hormuz, and expects AI, automation, robotics and immunotherapy to "crush the curve of costs" within four or five years, while conceding the Fed may still raise rates. He cites Truflation's estimate that the dollar has lost 28% of its value since January 2020, calling that inflation a regressive tax that has crushed asset-poor workers.
- Bitcoin could reach half of gold's market cap. That would be roughly 10x over the next decade, by his forecast; he expects 50–70% drawdowns every four years and says the current cycle is about halfway through, with issuance falling from 450 coins a day to 225 in about two years. He claims nine in ten investors who "do the homework" end up owning some, and puts Bitcoiners at about 5% of investors.
- Blockchain's value is removing third-party verification costs. Paying his restaurant's bill in stablecoins would save a 3.5% card fee, which on a 15% margin is, he says, roughly a 28% uplift to gross margin. He puts global third-party verification costs at about $4 trillion, an estimate he took from an AI chatbot, and compares the change to call prices falling from $3.50 a minute in 1985 to nothing.
- He won't sell SpaceX, because betting against Musk is a mistake. He admits the valuation looks stretched against revenue but notes that Ron Baron sees a $4–5 trillion company. The xAI–Twitter merger also gave Scaramucci more SpaceX shares.
All Things FinancialLogistics Coercion at the Panama Canal
- The coercion campaign is measurable, escalating, and unaffected by diplomacy. Panama-flagged detentions in Chinese ports rose from 92 in March to 135 in April to 139 in May; the April–May total of 274 exceeded the 251 recorded across all of 2025. Panama's share of all Chinese detentions climbed from ~74% to 82% and then 85% in the first ten days of May. The step-up began around 8 March 2026, and the China–Panama foreign-minister meeting on 26 May did not slow it — May was the heaviest month on record.
- The registry, not the canal, is the lever that reaches Panama's treasury. Cargo ships above 10,000 dwt leaving the Panama registry rose to more than 80 in April and roughly 150 in May, against a monthly average below 40 in Q1 and 55 across 2025, out of about 5,000 vessels in the segment — a ~3.3% contraction in 2026, with owners reflagging mainly to Liberia and the Marshall Islands to shed political exposure. Registry income is sovereign revenue independent of canal tolls, which is why this channel bites hardest. Panama's response has been defensive rather than retaliatory, issuing guidance to appeal detentions; no symmetric lever exists.
- Tactically effective, strategically self-defeating — and the framing insight is about ambiguity, not geography. The pressure moved Panama toward conciliation but did not change the outcome: APM Terminals and MSC's Terminal Investment Limited now run both terminals under 18-month interim agreements. Meanwhile the bluntness drew condemnation from six governments plus separate statements from Israel, Ukraine, Honduras and Peru, handing Washington a clean coercion narrative across the hemisphere, while the reflagging exodus erodes the very registry leverage being squeezed. The analytical note is the sharpest line in the brief: a chokepoint holder's premium is a function of the credible ambiguity of its alignment, not its geography — Panama retains the geography and has lost the ambiguity. And the diplomatic and pressure channels are running in parallel rather than in sequence, with pressure as standing leverage behind the talks.
All Things FinancialThe seas maybe free in law, but power chokepoints have always tried to charge a toll 6-28-2026
- His historical claim is that freedom of the seas is an eighty-year anomaly, not a norm being defended. In the ancient world ships hugged coastlines and paid a toll at every city they put in at overnight; Constantinople's wealth was toll wealth, and in the era of British dominance the ruler of Singapore — sitting on the Strait of Malacca — was the richest man in the world. There has been no period of genuinely free seas except the last eighty years, and that was a US Navy artifact rather than a legal condition. Hormuz is simply where the artifact expired first.
- The toll-versus-fee distinction is the legal fiction everyone is now using. A toll on a strait is illegal under international law; a fee — environmental, paperwork, whatever the label — is not. Iran announced a toll, realized the error within a day, and reissued it as a fee. The same wording lets Washington charge for Panama transit while denouncing tolls elsewhere.
- China's Panama campaign shows coercion that works tactically while failing strategically, and Beijing appears untroubled by that. Detentions of Panama-flagged vessels ran 92 in March, 135 in April and 139 in May, driving over 230 ships off the registry in two months against a 2025 average of 55 for the year. The OAS confrontation on 24 June coincided with the heaviest detention month on record, which he reads as indifference to condemnation. He also notes China is suing Panama for $2B, likely $3–5B if it runs for years, with Panama — not Washington — paying. His standing model prediction closes it: most chokepoints pass into private hands within five to ten years, and a genuinely free Panama Canal is unlikely to be allowed.
GCastCreating a Workflow in Microsoft Foundry [GCast 220]
- Workflows chain agents and logic nodes in Foundry. Giard calls Workflows a feature "in preview right now", reached from the Workflows option on the project's Agents page, and he builds on the agent from episode 215. A new workflow starts blank (only a start node) or from a template. The nodes he lists are: invoke an agent (pick an existing one or create one on the spot), set, reset or parse a variable, conditional logic, loops and go-tos, send a message to the user, ask the user a question (which pauses for input), and end.
- The human-in-the-loop template loops until the user's answers match. It stores the user's first input in a variable, asks them to confirm it, and uses a condition plus a go-to to keep asking until the reply matches the original. In the preview, which Giard likens to the agent sandbox, mismatched replies sent the flow back around the loop until a matching one confirmed it. A workflow's name must be unique within the project.
- The sequential template passes each agent's output to the next. Giard calls it "one of the more interesting ones". Its marketing example runs three agents in order: an analyst lists key features, target audience and selling points; a copywriter turns these into English prose; an editor polishes the wording. His reason for using separate agents rather than one big prompt is that each part can be tweaked on its own, and one agent can be reused across several workflows. In the demo, a made-up kitchen gadget went from bullet points to rough prose to finished copy.
- Each agent can use its own model, temperature and top P. Giard says temperature and top P do the same thing with different algorithms, and Microsoft recommends changing one but not both. Lower values give more consistent, deterministic output. Higher values give more creative output but a higher chance of hallucination, so the right setting depends on the use case. Tools, such as search over internal repositories, databases or SharePoint, and guardrails can also be added, but he skips them here.
- Publishing a workflow as a workflow app exposes an endpoint. Giard says applications or a website can call that endpoint to use the workflow.
Mohamed El-erianThe Weekly Look at the Global Economy and Markets
- Oil's sharp reversal from the Iran War shock lowered borrowing costs and gave a boost to lower-income households, even as tech stocks fell on technical rather than fundamental pressures. He attributes the tech pullback to crowded positioning, quarter-end rebalancing and tax-loss harvesting rather than weaker earnings power or AI productivity tailwinds, which he says remain robust.
- May's PCE inflation data, in line with or softer than consensus, split Fed-outlook views sharply — he sees peak inflation for the cycle while Bank of America predicts three hikes this year. The disagreement comes as new Fed Chair Warsh launches task forces to investigate five areas of institutional slippage, with officials themselves yet to converge on a unified view.
- A weekend escalation — the US bombing Iran after Tehran struck a ship in the Strait of Hormuz, followed by a second attack Saturday — tests whether markets can keep sidelining the Iran War's economic effects. Attention otherwise shifts to Thursday's US jobs report, where consensus looks for 115,000 new jobs, 0.3% monthly wage growth and unemployment unchanged at 4.3%.
Y CombinatorIndia Can Create The Largest AI Companies
synopsis
Two former YC staffers now working in the Indian startup ecosystem, Puneet, an ex-founder of Swiggy-acquired SuperDaily turned early-stage investor and builder again, and Arnav, now at Peak XV, join YC partners to argue that India is positioned to build some of the world's largest AI companies — not because the playbook has changed, but because AI has erased the geographic and network advantages that used to gate access to global markets.
Puneet frames the current moment against the prior mobile-internet wave, which "tokenized labor" into hyperlocal marketplaces confined to individual countries; AI adoption, by contrast, is global and meritocratic, rewarding whoever understands the technology most deeply rather than whoever has US connections. He points to an Indian engineering student who cold-emailed US insurance companies and closed deals without ever visiting America as evidence the old barriers are gone. Arnav adds that India's traditional educational advice — chase a safe, prestigious job — is now arguably the riskier path, and urges young builders to develop independent judgment by surrounding themselves with equally ambitious peers, much as YC batches do.
YC partners extend the thesis to their own hiring bar and advice: coding agents have leveled the playing field so completely that founder age keeps dropping, tinkering and "projects" — two people building something unassigned and getting someone to use it — reliably surface strong startup ideas, and second-movers with better execution regularly beat first movers, as Giga did in winning a DoorDash contract against much larger incumbents. On tooling, one partner recounts not grasping the frontier's pace until paying for an unlimited-usage coding plan, arguing that spending aggressively on tokens — running far more tests and iterations than feels justified — compounds into a real edge as compute costs keep falling.
The session closes on YC's own selection criteria: clarity above cleverness in an application, taste defined as intention grounded in real customer insight rather than visual polish, and agency measured by whether a founder does "projects" — unassigned work, built with someone else, that a real user actually adopts — long before they ever apply.
- India can now build globally dominant AI companies, not just local ones. Puneet, a SuperDaily founder turned investor and builder, contrasts this wave with the mobile-era wave that "tokenized labor" into hyperlocal networks; AI adoption is global and meritocratic, so an Indian student can cold-email U.S. insurance companies and land contracts without ever having lived in Silicon Valley, because success now hinges on technical depth rather than warm introductions.
- Coding agents let small, second-moving teams beat larger incumbents. Arnav, formerly of YC and now at Peak XV, cites Giga winning a DoorDash contract against competitors with hundreds of employees while fielding a team of eight, arguing this reflects a broader pattern where founders with product clarity and modern agentic tools can out-build first movers.
- Heavy token spending on AI tools is framed as a competitive edge. A YC partner says he didn't grasp how far frontier models had advanced until he paid for a $200-a-month max-usage plan in December, arguing that letting "the tokens rip" — running far more tests and edge cases than a human would bother with — produces meaningfully better output, and that inference costs will keep falling.
- YC's application advice centers on clarity, taste, and unassigned "projects." A YC partner says YC invests in founders, not ideas, since first ideas often don't work; he defines taste as building demos backed by real customer insight rather than aesthetics, and recommends young builders do "projects" — two or more people building something nobody assigned them and getting someone to actually use it.
Technology and FriendsJeff Fritz on Copilot for App Modernization [EPISODE 909]
- Teams skip upgrades because the payoff isn't immediately visible. Fritz says a move such as .NET 7 to .NET 8 costs hours of investigation and hand-tuning with no obvious ROI. Those who follow the industry see faster frameworks and lower memory use, so more applications fit on fewer servers and in containers, cutting the bill. Meanwhile each skipped version makes the eventual migration a bigger "mountain".
- Modernizing means rewriting code, not bumping version numbers. Fritz lists changed identity and security practices, EF Core features absent from EF6, moving hand-written ADO.NET to ORM and repository patterns, OpenTelemetry with Aspire, and cloud resiliency (multiple instances, Front Door, Redis caching, Service Bus queuing). He says the old deterministic Upgrade Assistant relied on find-and-replace and struggled when code didn't match its patterns, while Copilot can write code, check builds, look up documentation and use skills that encode those patterns.
- The modernization agents assess, plan, then execute, with a human checkpoint. Fritz's tool set, the GitHub Copilot Modernization Agents, runs on Copilot in Visual Studio, VS Code, GitHub and the command line. The plan is an English markdown file the developer can edit before anything runs, for example to impose an in-house standard of Newtonsoft.Json over System.Text.Json. Execution can happen on a separate git branch, and progress is checked off in task markdown files on disk so a stopped or cloud run can resume. Fritz says that whatever model wrote the code, "I'm the one responsible for it running in production."
- The team now targets whole portfolios, not single applications. Fritz says his team has focused on this for about the last six months, because organizations hold back because their applications are integrated with each other and must keep working. A command-line tool scans and modernizes a whole portfolio, and a companion tool called "command center" tracks assessments, plans and execution across applications, with the agents working independently but coordinating where needed.
- Fritz's employer claims large savings and wide version coverage. Fritz, who works on the product at Microsoft, reports 70–80% cost and time savings in the migration process, citing customers including Xbox, Bing and Ford from his Build session. He says support reaches back to .NET Framework 2, covers Web Forms to Blazor on .NET 10 with .NET 11 next, and moves Java up to versions 25 and 26. It ships as an extension for Visual Studio 2026 and VS Code and is documented on Microsoft Learn under "App Mod" or "GHCP modernization".
Polis PanditTrump's Housing Bill Mistake
- Trump withheld his signature from a bipartisan housing bill (passed 358-32 in the House, 85-5 in the Senate) over unrelated voter-registration demands. The 21st Century ROAD to Housing Act restricts institutional investors' single-family home purchases, streamlines environmental reviews, and incentivizes local governments to build, but Trump instead wanted federal voter-fraud legislation, despite data — including from the Heritage Foundation — showing voter fraud is negligible.
- The institutional-investor restriction is more symbolic nationally than locally. The ban triggers at 350 homes, and even investors owning 1,000+ homes hold only about 0.34% of national housing stock, though the impact is concentrated in specific metros like Jacksonville, Dallas, and Phoenix where such investors own 20%+ of rentals.
- Trump's leverage is illusory: the bill becomes law regardless of his signature. With veto-proof majorities and Congress in session, it takes effect 10 days after presentment; withholding his signature only forfeits credit for an affordability win — the platform he ran on — while handing Democrats the opening to claim it instead.
Scott HanselmanThe dark matter between your commits
- Git versions checkpoints; DeltaDB aims to version the process between. One speaker paraphrases the pitch: the abandoned edits, prompts, agent steering and small decisions between two commits are "dark matter" that should be versioned too. The other agrees and adds that the record should be open to multiple participants, in real time as well as later.
- Agent work today is a private session that leaves only snapshots. The other describes the status quo as a one-on-one "DM session" with an agent, whose byproduct is the snapshots the user commits and pushes to GitHub.
Professor G Markets$The Week: Iran, Brexit, SpaceX, and the Business of Sports
- This post includes paid content that is not available publicly. See the original source.
All Things FinancialThe airline analogy and what has survived
- The airline analogy holds in the places that matter, which is why it is the best one in circulation. Both businesses are brutally capital-intensive with high fixed and near-zero marginal cost — one more passenger or one more inference token is essentially free once the plane flies or the cluster is racked. Four traits carry straight over: perishable inventory, since a GPU-hour unsold at the moment of capacity is gone forever, which is why the industry already runs airline-style spot/reserved/batch yield management; commoditizing output, with H100 rentals falling from $7–10/hour in early 2024 to $2–4/hour by late 2025, a 50–70% collapse that is simply a fare war; a synchronized capacity cycle, with everyone ordering against the same demand thesis; and supplier capture, with Nvidia as Boeing and Airbus except capturing far more of the chain. The big four guide to roughly $700B of 2026 capex, up ~77%, with power and cooling rather than chips as the binding constraint.
- Where the analogy breaks, it breaks against AI — and the debt market has not noticed. A 737 has a 25–30 year economic life, a deep lease market, and an entire bond market built on it as collateral. A GPU's contested economic life runs two to six years, and that disagreement is the whole story: Amazon shortened server life from six years to five in early 2025 and took a $920M accelerated-depreciation charge, while CoreWeave, Microsoft and Alphabet hold at six, and Burry claims roughly $176B of understated depreciation across 2026–2028. Yet more than $20B of GPU-collateralized debt is outstanding — BlackRock, Blackstone, PIMCO, Carlyle, JPMorgan, Macquarie — all resting on the five-to-six-year assumption. The airline-like layer carries worse-than-airline collateral, and the debt stack is mispricing exactly that.
- Returns are destroyed in the middle of the stack and captured at both ends, which dictates the survivable model. Applications are software economics with durable IP; foundation models are the contested hinge; cloud capacity and raw GPUs are the airline trap; the data-centre shell is REIT economics at 20+ years; power and grid are utility and locational-monopoly economics measured in decades. So the model that earns is a vertically integrated utility-plus-software hybrid — own the power and the shell as the durable moat, own the model and distribution for margin, and treat GPUs as a depreciating consumable financed short rather than termed out as collateral. The cleaner historical rhyme is the 1990s fiber build: Global Crossing, WorldCom and Level 3 were right about demand, built far ahead of it, watched roughly 90% of fiber go dark, and handed the value to the second owner. The counter-case worth tracking is workload cascade — training to high-value inference to batch, extended by prefill/decode disaggregation — which would soften the depreciation-doom thesis if it holds.
All Things FinancialAI maybe the airline industry of software: essential, expensive and brutally competitive 6-26-2026
- OpenAI postponing its IPO is, on his reading, a direct consequence of the government signalling it wants half of these companies. He had asked in a prior video how an AI company comes public once Washington announces an intent to take roughly 50% of every one of them; three days later the IPO was deferred to next year. His own view, offered explicitly as opinion without inside knowledge, is that OpenAI does not come public at all.
- Depreciation accounting, not revenue, is what is generating the reported profits. Stretching chip useful life to five or six years against a real life closer to two or three understates depreciation by roughly $176B, and that difference is being booked as profit — the trouble arrives when the schedules run out. Meanwhile $20B+ of GPU-collateralized debt sits with BlackRock, Blackstone, PIMCO, Carlyle, JPMorgan and Macquarie, all lending against the longer assumption, using a mechanic borrowed from aircraft and power-plant finance where the underlying technology cycle is far slower. The stack's economics put all the value at the two ends and all the expense in the middle.
- Japan's ¥-scale spending plan is his two-year-old sovereign-fund forecast arriving in legislative form. A $2.3 trillion, 14-year blueprint combining public and private investment across 17 strategic sectors — roughly $600B for AI and semiconductors, plus shipbuilding, solar, quantum computing and next-generation nuclear — explicitly framed as breaking a cycle of excessive austerity. He notes economists faulted it for lacking detail on the public/private split, but treats it as confirmation that every country will direct sovereign and pension capital into designated sectors. On the yen at ~161.6: Japan has already spent $79B defending it, and he warns that another $200–500B of defence would push US rates materially higher.
Dr. Jon KrohnRecursive Self-Improvement (Ep. 1004 with Jon Krohn)
- Anthropic's AI-written code is AI-assisted coding, not yet RSI. The host relays Anthropic's report (Anthropic sponsors the show and this episode; he says it had no editorial input) that as of May 2026 over 80% of code merged into its production code base was written by Claude, up from low single digits before its coding agent launched in early 2025, with engineers shipping roughly eight times as much code per day as 2 years ago and one April effort of over 800 fixes cutting a class of API errors a thousandfold. He argues true recursive self-improvement would need the AI to design experiments, evaluate results and choose next steps with little or no human in the loop, whereas humans still set goals and judge results today.
- Measured autonomy is accelerating toward self-improvement. He cites METR's finding that the length of task an agent can complete alone doubled about every 7 months for 6 years and has sped to roughly every 4 months over the past year, and Anthropic's figure that success on its hardest open-ended problems rose from under 20% in late 2025 to 76% by May 2026. His concrete cases: Google DeepMind's AlphaEvolve recovering nearly three-quarters of a percent of Google's worldwide compute through better scheduling and speeding flagship-model training about 1% with a faster matrix multiplication, and Andrej Karpathy's overnight-tuning agent running about 700 experiments over 2 days, keeping about 20 improvements and cutting training time on his own optimized code from 2.02 to 1.80 hours (11%).
- An Anthropic co-founder thinks self-improving AI is more likely than not soon. Anthropic's report lays out three scenarios (AI below top engineers; accelerating AI-assisted engineering with humans steering, which Anthropic and the host consider most likely; AI improving itself), and co-founder Jack Clark gives a 60% probability of a system able to create its successor with no human involvement by the end of 2028.
- Compute, verification and marketing incentives argue against that timeline. The sceptics' case he relays: each generation needs more compute, tied to data-center construction and competing with paying customers; self-generated training data works where success is cheaply verifiable (code, proofs) but risks "recursive drift" elsewhere; and a market leader urging the option to slow down is also asking rivals to ease off. Against that he cites Max Tegmark's warning that the danger is a quiet loss of control, with models trained, goal-set and safety-checked by other models.
- The host lands in the "optimistic middle". His view is that the productivity gains are real and available to anyone now, and that the closer systems come to self-improvement the more monitoring, human checkpoints and oversight should be built in while the tools remain under human direction.
Technology and FriendsMike Amundsen on AI Coaches That Amplify Expertise [EPISODE 908]
- AI should make people better, not just produce one-shot output. Amundsen says one-shot tools that turn inputs into an OpenAPI document, a spec or a Node.js app shrank his own contribution to "whatever the machine thought". He now writes "coaches" that converse with the user and check whether to continue. He says that is how he learned his company's style guide.
- A coach follows a teaching loop and fits in one file. Amundsen modelled it on how a mentor works: rather than giving the answer, it sets a slightly hard problem and then helps correct the attempt. Each coach is a single self-contained markdown file attached to or pasted into ChatGPT, Claude, smaller models or even Ollama, and all follow a similar templated flow.
- Tedious, exact artifacts go to deterministic tools, not the LLM. Amundsen says generating an OpenAPI document is too tedious to leave to a stochastic model, so the coaches emit configuration or output files for tools, mostly CLIs he sometimes builds himself. Tests need a tool such as Bruno or Postman to be useful.
- Generative AI doesn't know the creative process, so users must impose it. Amundsen describes three stages: brainstorming, refinement, execution. He says AI is good at brainstorming, though only with ideas someone has already had, and does not know when to stop and move to refinement; at execution it yields a usable prototype, not production work. Giard adds that many humans are also bad at time-boxing; Amundsen's coaches use a seven-step process with explicit stopping points.
- Cheap prototypes should raise quality, not just speed. Amundsen argues AI makes it affordable to try a few more experiments before the final one, and that output improves with the extra feedback. His O'Reilly book "AI Driven API Design" makes chapters out of coach exercises; he says he uses AI little in the writing itself because of copyright issues. An early release of at least four chapters is planned for July, with the full book hoped for by the end of the calendar year.
Professor G Markets$AI Is Dumber Than Investors Think — with Gary Marcus
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$The Crisis of Adulthood — with John Burn-Murdoch
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Explosive Trump Secrets Revealed in New Book — with Jonathan Swan & Maggie Haberman
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$REPLAY: Design as Differentiator — with Katherine Dillon
- This post includes paid content that is not available publicly. See the original source.
All Things FinancialThe Midstream Squeeze
- This is a coercion architecture with two channels, not a bilateral rare-earth spat. The physical channel is a chokepoint where one cut material paralyses capacity far downstream; the legal channel is extraterritorial re-export liability that pulls non-Japanese firms worldwide into Chinese jurisdiction. The instruments arrived as a deliberate ratchet rather than a single shock — the 2026 Dual-Use Catalogue (31 Dec 2025), the Japan dual-use ban (6 Jan), 20 hard-banned entities plus 20 watch-listed (24 Feb), State Council Order 834 (31 Mar), and implementing Measures on 22 June that switch the enforcement mechanism on.
- Tungsten is the load-bearing case and the shutdown is already dated. China holds 80%+ of global tungsten, Japan has effectively no domestic ore, and tungsten powder is 60–70% of WF6 cost. Kanto Denka Kogyo and Central Glass — jointly ~25% of global WF6 capacity, about 2,200 t/yr — permanently cease production from 1 July, opening a ~2,000 t second-half gap against 8,000–9,000 t/yr of global demand. Downstream sit Samsung and SK Hynix, which sourced roughly 80% of their WF6 from that Japanese pair, plus TSMC. WF6 is the irreplaceable precursor for sub-7nm logic, HBM and 200+ layer 3D NAND — the AI-chip stack. China's tungsten exports to Japan hit zero from February through April 2026.
- The extraterritorial trap is the part most analysts miss, and it is what makes a Japan measure everyone's problem. Article 49 of the Dual-Use Regulation gives Beijing Foreign-Direct-Product-style reach: any entity in any country is liable for transferring China-origin dual-use items to Japan in violation; foreign-made goods are captured at a 0.1% de minimis or if made using Chinese REE processing technology; a 50% affiliates rule radiates the 40 named Japanese entities out to their global subsidiaries; and intermediary liability drags freight forwarders, customs brokers, financiers and platforms into the perimeter. The net effect is not one supply cut but a permanent obligation — China-origin tracking and end-use attestation for every multinational with this content anywhere in its bill of materials.
- Price transmission is severe and the cascade is regressive. WF6 is up roughly 200–233% year-on-year and the APT tungsten benchmark up ~557% since February 2025, landing on a memory market where DRAM has roughly doubled since early 2025, with smartphone shipments projected down ~12.9% and PCs ~11.3% in 2026. Two refinements matter: the memory crunch is mostly HBM reallocation — a wafer routed to an Nvidia stack is one denied to consumer DRAM — so WF6 is a second independent shock on the same node; and hyperscalers lock supply through long-term commitments while mid-market buyers compete for residual capacity, so smaller firms absorb both higher costs and longer lead times. Note the entry's own corrections: the circulated "~1,100 items" figure was dropped as unconfirmable, the entity-ban count was an undercount (40, not 20), and a Bessent attribution was downgraded to partial. China's domestic WF6 capacity has meanwhile ramped to ~2,200 t/yr at 6N purity, breaking into global fabs exactly as Japan exits.
All Things FinancialWhen a ban force a choice between two products, there is no real supply - only triage 6-25-2026
- China has moved from agency discretion to codified law, with explicit extraterritorial reach. What MOFCOM ran administratively is now a 15-point export-control law, with 10 US entities added in direct retaliation for Washington naming 10 Chinese firms — Alibaba among them — as defence companies. The clause that matters: compliance now applies not only to direct China sales but to transfers of China-origin dual-use goods anywhere in the world. His analogy is the US using SWIFT — Beijing has built the same category of reach from the materials end of the stack.
- Triage is the real consequence, and it reaches consumers. Japan losing a quarter of global WF6 capacity forces an allocation choice downstream: hyperscalers or consumer devices. He points at Apple's announced 20% price increases as the first visible landing, and connects it to a call he made three years ago — that there would come a day Americans could not buy an iPhone or iPad, and that the consumer market would bifurcate with the US holding lesser technology than China and the countries in good standing with it. His earlier triage framing for rare earths, with the government acting as judge over who gets supply, now applies to memory.
- The damage runs well past chips, and the affected firms are behaving as though it lasts. Sumitomo Electric reports tungsten procurement "completely stopped," is raising prices up to 60% and building a ¥15.9bn recovery plant — capital that only makes sense if the cut persists. Mitsubishi Materials tripled cemented-carbide prices on June-onward orders. The hard-ban list cut Mitsubishi Heavy, Kawasaki Heavy, IHI, Fujitsu Defense, NEC, JMU Defense and JAXA off from the controlled minerals, with share prices falling on the announcement. He also flags his own verification discipline mid-stream: the widely circulated "20 Japanese companies banned" figure was an undercount, since another 20 were watch-listed the same day, making the real number 40.
Scott Hanselman300 AI conversations touched this one line
- Indexing every edit lets code be traced to its conversations. The speaker says DeltaDB records everything that happened in a Git-like graph and indexes it, so from any code one can reach each edit, the conversation that created it, and any conversation that referenced any version of it.
- A character can be followed losslessly across versions, even when deleted. Because the causal record covers everything after an edit, the speaker says the system can locate a character in any later version without "squinting through a diff", and work out where it would be if deleted. Their example is knowing that, in the past 30 days, 300 agent conversations referenced a particular piece of code in a file.
Crosstalk SolutionsHuge Project Nomad Update — Supply Depot App Store is Here
- Nomad's new Supply Depot is a curated Docker app store. It arrives in Project Nomad 1.33; the host, whose company makes Nomad, says each app is a tested open-source container with its own setup docs, and installs check for port conflicts first; apps include Jellyfin, Vaultwarden, CyberChef, Stirling PDF, Excalidraw, Homebox, IT Tools, Flatnotes, a file browser, an ebook reader, and Meshtastic Web and MeshCore Web. Version 1.33.0 was the latest at recording, updated from 1.32.1 via Settings, Check for updates.
- The old Kolibri container has no direct migration to the new one. The previous Kolibri education platform (TED Talks, Khan Academy K–12) ran a very old version, so Supply Depot offers a second-generation Kolibri. Users with the old one must install the new version and follow the project's docs to move their content over.
- Any Docker image can now be installed as a custom app. Users can paste an image and set custom port mappings, volume mounts and environment variables, and can edit, restart, reinstall or set a custom launch URL for installed apps. The host says this makes Nomad a home-lab server as well as an offline knowledge stockpile, a feature he had long wanted.
- Automatic updates now cover Nomad itself, apps and content. Each of the three runs in a user-set update window; content means Zim files such as Wikipedia. The intent is a box that keeps itself current while online, so the moment it is disconnected is its point-in-time snapshot; manual content updates remain available.
- A custom online check fixes networks that block Cloudflare. Nomad decides it is online by reaching Cloudflare's 1.1.1.1, so networks blocking that made it think it was offline; a new Advanced tab lets users set their own HTTP check. Other user-requested fixes: maps remember the last view, a rescan-library button picks up Zim files copied in, and knowledge-base ingestion jobs (manual by default) can be cancelled.
Professor G Markets$America Finally Has a Housing Plan. Trump Is Blocking It.
- This post includes paid content that is not available publicly. See the original source.
AI Analyst LabHow to check if an AI analysis is right, with no answer key
synopsis
This AI Analyst Lab lightning lesson, the second on AI evals after a session on reliability, argues that an analyst can rarely prove an AI analysis correct but can bound the risk of it being wrong. Its method: ask first whether a trusted answer exists, score against it where it does, and otherwise grade the work with four checks before deciding to act, dig or abstain. Shane presents, with two co-hosts on the chat.
Why it is hard. Shane's premise is that AI now runs the SQL, charts and recommendations in minutes, and a wrong answer reads the same as a right one. The manual friction that used to catch errors is gone. Last week, the same retention question sent to five agents came back anywhere from 30% to 99%, each a defensible definition, until the definition was written down.
Two situations. Analysis is a pipeline, from question to tables, SQL, result and recommendation, and every step can fail. Whether ground truth exists splits the cases. Where it does, use "found truth" (finance-reconciled numbers, board metrics, much-verified queries) or build "forged truth" for decision-driving questions, though old references can carry old bugs. Trusted numbers are building blocks; the AI's job is the novel cut on top of them.
Grading the work. For novel questions he offers reliability, receipts, reconciliation and reading the trace. Each has a stated gap: a stably wrong answer passes reliability, an error in every segment passes reconciliation, an LLM can fake a receipt (hence hook-logged, code-generated receipts), and a definition error reads fine in a trace. Together they let an analyst tell a stakeholder what was checked. The outcome is act, dig into the disagreement, or abstain, with rigor matched to the stakes, and corrections fed back as context only once verified.
Demo. On a fictitious e-commerce dataset, Claude Code audited a revenue-by-category query. It found a join fan-out from 47,000 order rows to 75,000 line-item rows that inflated revenue to 6.9 million from about 3.6 million, 87% too high. It flagged a missing cancelled/returned filter worth another 15% and a total that includes shipping, and noted the data ends in 2024. An attendee asked how the query is captured; Shane says a hook saves every query to JSON, though he pasted it by hand here.
Q&A. For messy data or small teams, Shane recommends the cheap checks first, and says agreement between runs is not correctness.
- A wrong AI analysis looks exactly like a right one. Shane says it produces a plausible number and a convincing story either way, and the cost is broken trust if someone catches it or a bad decision if no one does. He recaps last week's reliability eval: the same retention question to five separate agents returned five defensible answers from 30% to 99%, and writing down the definition made runs agree.
- First ask whether an answer key exists. Where one does, start with "found truth" — finance-reconciled figures, board metrics, queries verified many times — and run it like unit tests; for questions that drive decisions and have no key, "forge truth" by working out the answer once and saving it, which costs hours to weeks. He warns that a metric can be quoted wrongly for years, and an eval graded against old buggy queries teaches the AI the same bugs.
- Without a key, grade the work, not the answer. His four checks are reliability (does it agree with itself), receipts (the exact SQL, tables and metric), reconciliation (segments sum to a trusted total), and reading the trace for a defensible method. None proves the answer; each "shrinks the space where a mistake can hide", and each has a blind spot: a stable wrong answer, an error that runs through every segment, an LLM-written receipt that is itself a hallucination (so log with hooks and generate receipts from code), and a definition error that reads fine.
- Every eval ends in act, dig or abstain. Shane frames abstaining as a real decision that builds stakeholder trust when confidence is low, and says rigor should match the stakes — deeper for the board, lighter for a curious stakeholder, with the cheap checks first where data is messy. A caught error becomes context for the system, but should not be trusted until verified, so it does not poison later answers.
- The demo's audit caught a join that inflated revenue. Asked whether a revenue-by-category query on the fictitious Novamart data could be trusted, Claude Code (using CLAUDE.md risk lists and a semantic layer) found a join fan-out — 47,000 orders exploding to 75,000 line-item rows with an order-level total summed — giving 6.9 million against about 3.6 million, an 87% overstatement. It also flagged no filter on cancelled or returned orders (a further 15%), a total that includes shipping and should be line total, and that the data stops at 2024; Shane pasted the query by hand here but says in practice a hook records every query to JSON.