Souk's Notes - Unofficial Summaries

2026-05-15 · open the original ↗

• This is an entertainment bit, not market or career content. Tom Sosnoff and his longtime partner Tony blind-taste four Chicago Italian subs, each trying to match the ranking published by The Infatuation, a restaurant-review site the hosts say JPMorgan bought. The opening casts it as a rivalry: Tom says pattern recognition from 30 years of reading markets carries over to a deli counter, and Tony says his Brooklyn-Italian upbringing gives him the edge.
• Neither taster could reproduce the critics' ranking. The critics' order was Paulina Market first, Fontano's second, Bari third and D'Amato's fourth. Each host got one placement right, both misplaced Fontano's, and one of them rated the critics' top pick, Paulina, his least favourite. On a $20 bet they ended level, and the hosts claim Tony's record against Tom going in was 0 and 17.
• They broke the tie with a hand of Liar's Poker. One host recalls playing the game in the trading pits and points to Michael Lewis's book of the same name, set at Salomon Brothers. The captions don't make clear who won the hand. The closing banter has each host calling the other's palate average.

2026-05-15 · open the original ↗

• AI compute demand is outrunning a supply chain with several simultaneous bottlenecks. The host dismisses "tokenmaxxing", maximizing tokens consumed in Claude Code or Codex to look good on internal usage dashboards, as a vanity metric. He cites the symptoms of scarcity: Anthropic's weekly Claude Code rate limits from last August, which it blamed on power users running it 24/7 and said it was "very compute constrained", OpenAI pausing Sora sign-ups, GitHub halting new Copilot subscribers, and unreliable on-demand H100s without reserved capacity. Fixing one bottleneck, he says, does not unstick the system.
• Packaging and memory, not GPU fabrication, are the chip choke points. He reports Nvidia data-centre GPU lead times of 36 to 52 weeks, Blackwell largely sold out through mid-2026, and H100 spot rents up about 30% since November. TSMC's CEO C.C. Wei has said CoWoS packaging is sold out through 2026; Nvidia reportedly holds about 60% of it through 2027, and new fabs take two to three years. HBM demand has roughly quintupled since 2023 (an H100 needs 80 GB, a B200 192 GB), and SK Hynix, Samsung and Micron are largely sold out into 2026. Nvidia reportedly cut RTX 50 output 30% to 40% as a result, and new HBM fabs take 18 to 24 months.
• Agentic AI needs far more CPUs per GPU. Morgan Stanley analysis cited by the host puts chatbot systems at about one CPU per 12 GPUs and agentic systems at one-to-one. Intel's CFO David Zinsner said on the Q1 call that its server CPU shortfall "starts with a B". The host says server CPU prices have risen 10 to 20% in a couple of months, a big part of why Intel's market cap more than doubled in 6 months.
• Electricity is the most intractable bottleneck. Hyperscalers report build-outs gated by grid interconnects (18 to 36 months) and transformers (another 18 to 24). Gartner projects power shortages will restrict 40% of AI data centres by 2027. The host counts at least 12 states with moratorium bills as of March, a Maine statewide moratorium passed in April and vetoed, and more than 50 local moratoria. He also cites rate rises of 8 to 15% near clusters in Virginia, Texas and Georgia.
• Suppliers aren't matching hyperscaler spending, so relief is years away. The host says even an aggressive supplier ramp-up wouldn't bring meaningful relief until 2028 or later. The five largest hyperscalers are on track for about $725 billion of 2026 capex, roughly three-quarters for AI, against about $120 billion in 2022. He says Anthropic committed over $100 billion to AWS for up to 5 GW, took about 5 GW more from Google, and took all of the "SpaceX Colossus 1" site, which let it double Claude Code limits. Suppliers have raised capex only about half as much as hyperscalers' tripling, wary of overbuilding. The host forecasts 2 to 3 years of physical constraint shaping which AI firms survive, while efficiency gains (Google's TurboQuant, Amazon's Trainium cluster of over 500,000 chips, mixture of experts) keep cutting application costs.

2026-05-15 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-15 · open the original ↗

• The 30-year Treasury yield functions as the only reliable check on Trump's foreign policy. Every major reversal of his second term — the April 2025 tariff pause, backing off invading Greenland, and reasserting an Iran ceasefire — followed the yield approaching or crossing 5%, with Treasury Secretary Bessent confirming the bond market factored into the tariff decision.
• The 30-year yield specifically (not the 10-year) is what disciplines Trump because it signals sovereign credibility, not just borrowing costs. It's the term premium foreign reserve managers and pension funds price off; when investors sell rather than buy Treasuries during uncertainty, that signals eroding faith in Treasury safety itself, not the usual flight-to-safety pattern.
• The author warns against celebrating this "fourth branch of government" check. Bond vigilantes only emerge when Congress, courts, and cabinet oversight have already failed — as with Liz Truss in the UK — and markets discipline only what's fiscally sound, with no regard for human rights or civil liberties, meaning a cruel-but-solvent policy would sail through unchecked.
• A recurring "SCOTUS Sanity Check" segment flags the DOJ negotiating a settlement in a lawsuit Trump filed against himself while President. Trump sued the IRS (which the DOJ represents) in January over leaked tax returns and is seeking $10 billion personally; the presiding judge flagged the fundamental problem that the parties aren't legally adverse to one another.

2026-05-15 · open the original ↗

• A skill is a loose contract between a runtime and a model. Marimo pair is a skill, a folder of markdown files that teaches an agent to start or connect to Marimo and run Python inside its runtime. Manz says the agent discovers the APIs hidden inside Marimo rather than relying on a strict contract. If the APIs change between releases the model adapts, and upgrading Marimo simply adds capabilities, without the usual worry over semantic versioning.
• Skills took off once models got good around last fall. Manz dates the shift in coding agents to around November with Anthropic's Opus. He first said 4.6, then accepted the host's correction that 4.6 came out in February and settled on 4.5. When people returned in January, skills became a high-impact way to share team processes that once sat in documents people had to read.
• Repetition and failures should update the skill. Building Marimo pair taught Manz a habit: if he is repeating himself to the agent, that belongs in a skill. When something goes wrong he updates the skill rather than "yelling at the model".

2026-05-14 · open the original ↗

• A sizing must separate assumptions from math. The presenter, a co-host, says most sizings take an optimistic assumption, multiply it by a large addressable market and hope nobody asks. The alternative is one formula with three inputs: users affected × improvement rate × value per unit. Each input is then marked as grounded in data or assumed, so the assumptions can be debated.
• Each input needs a traceable number, not an adjective. The presenter's examples are 12,400 monthly mobile users who start checkout and don't finish, closing half the gap between 33% mobile and 52% desktop conversion to reach 42%, and a $67 average order at 75% margin, about $50 per conversion. A vague $5 million "save for later" pitch is rewritten as 28,000 monthly users who view three or more products without adding to cart, 8–12% recovery from an industry benchmark, and a $45 first purchase. The presenter argues that the first traceable version sets the anchor the audience holds on to.
• A role prompt can grade a pitch input by input. A chat prompt casts the model as a senior product analyst and has it rate each input as clear, vague or missing, then rewrite the pitch with numbers. Given only "improve mobile checkout, it's clunky and we're losing sales", Claude marked all three inputs as weak and plugged in an assumed 500k users for a midsize site. A director in the audience objected that a finance persona and an analyst persona should give the same number. A third co-founder replied that the number should match when it is deterministic and only the framing differs, and called team-wide consistency a validation problem.
• The demo sized mobile checkout at roughly $300k in topline revenue. Opus 4.6 was no longer selectable, so the presenter used Opus 4.7, which the presenter says was released the day before. Claude Code found web converting at 40% from start to complete against 29% on iOS and 28% on Android, and its Simpson's-paradox check found web ahead in every channel. The repo's opportunity-sizer agent took a 15% relative mobile lift as the base case, with 5% as pessimistic, and gave a sensitivity table of lift against order value. It rated confidence medium: reach, order value and margin were data-backed, the lift was not. The presenter says it had no business context and its assumptions are for a human to challenge.
• Agents need context beyond the warehouse. Answering a question about connecting BI tools, the presenter says Claude knows nothing of a company's strategy, roadmap or past analyses out of the box. Notion pages, Google Docs, email and Slack, anything with an MCP connector, count as context and improve its answers, subject to the organization's security and legal approval. The presenter adds that the repo's skills were built by encoding the founders' practices and checking where Claude's defaults fell short, and that each user's copy grows to fit their own company.

2026-05-14 · open the original ↗

• AI decks look generic because the prompt gives no instructions. The presenter, a co-host, shows "pretty mid" slides from a bare Claude Cowork prompt. The argument is that the model takes the fastest route to the goal unless told otherwise. Four inputs fix that: the audience (a CEO deck hides 90% of the method a data-science deck would show), the decision the deck drives (ship, on track, or a resource ask), a storyboard before slides (TL;DR, context, findings, recommendation, never "a chart dump in chronological order"), and a theme.
• Build decks in a pipeline with automated review loops. The repo's sequence runs storyboard, narrative, charts, compiled deck, then human iteration, and each stage loops back until reviewers in the system pass it. Shane says the first draft then arrives already through what would be four or five rounds of review by colleagues and stakeholders, with passing criteria the user sets. The presenter says the last 10% of the work drives 90% of the impact.
• The demo built a New York Times–themed executive deck unprompted on design. On Opus 4.7, Claude Code listed six themes, including corporate, minimal, New York Times and Economist. It then read a finished mobile-checkout analysis on the fictional Novamart data and wrote the theme's CSS, rebuilt the charts with takeaway titles, and rendered a PDF and then HTML. The result was a short deck ending in an ask, with an appendix, built without any display instructions. The presenter thought it should have been wider. Shane says themes Claude invents fresh each time "can be pretty bad", so the repo predefines them with slide types, though Claude still creates new layouts.
• Fix it once, then codify the fix into the skills. Shane says Claude will make one-off fixes to the deck in hand by default. After each deck he asks what changed between draft and final, and updates the templates, skills or agents so the next deck starts aligned. He also says Claude Code can trigger deck-building automatically at the end of an analysis that starts from a stakeholder's Slack question.
• Claude Design is currently weaker than Claude Code for decks. Asked by an attendee, the presenter says Claude Design, released about three weeks earlier, is likely to improve. But two decks used up all the presenter's credits, and instruction-following and grasp of context were "pretty rough". Claude Code is more flexible, its cost more predictable, and it chains into other workflows. Other answers: output can vary between runs and between Claude Code and the repo's Cowork plugin, the $20 Pro plan is enough, and since Claude Code runs on Anthropic's servers a separate machine matters mainly for fully autonomous agents.

2026-05-14 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-14 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-14 · open the original ↗

• Panama's leverage and China's instrument operate at different layers, which is precisely why no symmetric response exists. Panama's Toll State position runs at the flag-of-convenience financial layer — a registry of roughly 8,600 vessels, about 18% of global tonnage, worth around $650M a year in sovereign revenue. China's coercive instrument operates at the physical port-state inspection layer, where holding seven of the world's ten busiest container ports confers regulatory authority over inbound vessels under the existing Tokyo MOU regime. Panama has nothing to retaliate with because it does not host Chinese-owned inbound traffic at meaningful scale, and China's cost of restricting Panama-flag access is near zero. The entry logs this as the ninth documented Factor Dimension Mismatch.
• Eight vectors run simultaneously, and the package is calibrated rather than escalatory. Detentions of Panama-flagged ships hit 93 of 125 in March 2026 — roughly 75% — against a historical baseline of 20 to 23 a month. An informal maritime advisory established that Panama flag now means political detention risk, eroding flag value directly. Chinese leasing companies require non-Panama flag as a financing condition, bleeding the registry across the entire newbuild cycle. Maersk and MSC were summoned and threatened with legal recourse. COSCO withdrew from Balboa permanently, rerouting via Lázaro Cárdenas and Buenaventura. State firms froze Panamanian infrastructure deals while CK Hutchison was penalized globally. Agricultural inspections disrupted more than 70% of tropical fruit exports into a market taking 39% of Panama's agricultural exports. And the arbitration claim widened past $2B. The design produces sustained domestic political cost without crossing thresholds that would trigger a formal multilateral response.
• The document proposes Port-State Coercion as a candidate concept and defines it tightly enough to be falsifiable. A destination port-state weaponizes its regulatory authority under an existing port-state-control regime to coerce a flag-state's domestic policy: the instrument is informal directive to inspectors, the surface justification is safety, and the operational signature is detention concentration far exceeding historical norms against a single flag. It is explicitly distinguished from Toll State transit rent, Node State military-backed control, settlement-layer denial, and brokered chokepoint access — a different layer of the same contest rather than a variant of an existing typology.

2026-05-14 · open the original ↗

• He reduces the episode to a single thought: paper against physical. America is paper power — the dollar, SWIFT, sanctions — and he argues it has been losing control of the paper markets in silver, gold and oil precisely because those markets have decoupled from the physical metal and barrels underneath them. Having Panama's Supreme Court declare an eight-year-old concession unconstitutional is a paper move. Barring Panama-flagged vessels from Chinese ports is a physical one, and only one of those can be ignored.
• The exposure is the registry, not the canal, and that distinction is where the damage lands. Canal traffic is probably running above trend given the oil disruption. But the registry is the long-term lifeblood of a country of 4.4 million people — roughly $650M a year — and it is being bled three ways at once: detentions running 93 of 125 in March against a 20-to-23 baseline, an informal advisory that makes the Panama flag a detention risk in any charterer's calculation, and Chinese yards refusing to build vessels destined for that flag. Agricultural retaliation compounds it, with 70% of fruit exports going to the country now finding fault with them.
• His closing point is that the multi-vector habit extends well beyond this dispute. China is building a new port in Peru and a rail system running across South America, which means that for the first time it can move goods to the continent without using the Panama Canal at all. He frames the whole episode as evidence that the US has divorced itself from the physical world into the financial one — world-class at financial engineering, in part because that is what remains.

2026-05-14 · open the original ↗

• IBM deliberately keeps the physical mainframe visible. The speaker describes "hug your mainframe day" social-media programs, on which people post photos with their data-centre mainframes on LinkedIn, and says real design thought goes into the machine's door so that it makes an impactful visual at launches. The same passage appears in the 2026-05-11 Hanselminutes interview with Skyla Loomis.

2026-05-14 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-13 · open the original ↗

• Foundry agents ship with default guardrails that block or log. Working on the agent from episode 215, Giard shows two built-in guardrail versions, which he says do much the same thing, with version 2 doing a little more. They cover categories such as hate speech, self-harm and sexual content, on user input (prompts), on output (responses), or both. Each one either blocks the content outright with a refusal message, or only logs that it happened.
• A custom guardrail can be built once and reused across agents and models. Under the project's guardrails section, the options Giard lists are a jailbreak check that applies to user input only, a protected-material check for code or text that applies to output, which he says is "so that you're not revealing secrets", and categories such as hate, self-harm, sexual and violence that apply to both. Each risk has a severity level that sets how severe content must be before the guardrail kicks in. He attaches the guardrail to his one agent and to both models he has deployed, then names it.
• The demo hate-speech guardrail blocked a threatening prompt. After he assigned the guardrail to the agent, the test prompt "I will kill you" got a refusal saying the request was blocked by guardrails.
• Block lists forbid specific terms in prompts or responses. Foundry includes a built-in profanity block list, which Giard says rejects prompts containing swear words and filters them out of responses. A custom list can be loaded from a CSV file for long lists of terms, or entered one term at a time. Each term is matched either exactly, ignoring case, or by a regular expression.
• A block list takes effect only through a block-list guardrail. Giard builds a block list containing "rosebud", then creates a guardrail of type block list that uses both it and the built-in profanity list, on input and output. Annotations are not supported for this type, so it can only block. Once it was applied to the agent and its models, a prompt asking about "rosebud" was refused because the block list forbids the term.

2026-05-13 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-13 · open the original ↗

• Fraud must be caught during the transaction, not after. The speaker says that a delay of even 30 to 60 seconds forces a bank to let the transaction through, and that fraud let through is basically a loss, which many banks treat as a cost of doing business at about 8% of revenue a year.
• On-platform AI makes scoring every transaction feasible. Running inference where the transaction happens removes network latency, which the speaker says gives sub-millisecond or single-digit-millisecond responses and allows scoring on 100% of transactions. The same passage appears in the 2026-05-11 Hanselminutes interview with Skyla Loomis of IBM Z.

2026-05-13 · open the original ↗

• The confirmation hearing asked one question all day, and it was the wrong one. Everyone wanted to know whether Warsh would cut rates for Trump. The actual programme is bringing the Federal Reserve under Treasury direction, resting on two explicit commitments: the Fed publishing forward-looking guidance on the size and trajectory of its balance sheet, and Treasury aligning its issuance calendar and maturity composition with Fed liquidity operations. The rhetorical reference is the 1951 Accord — which established independence by separating monetary policy from Treasury's financing obligation — and the direction is inverted, binding them back together. Warsh has spoken openly of regime change at the Fed, including tighter coordination with Treasury and the administration on non-monetary policy.
• The measurement switch is where policy actually changes, without any decision being announced. Warsh prefers trimmed-mean and median gauges — Dallas Fed trimmed-mean PCE, Cleveland median CPI — over the Fed's current core PCE. The same economy prints 2.8% on one measure and 2.3% on his preferred one, a gap of 0.7 to 1.0 percentage points. That is not a calibration detail; it is the difference between policy must remain tight and policy can be eased, which is why his Senate testimony could describe the current trend as "quite favorable."
• The pipeline arithmetic points the other way, and he lets it correct his own thesis. PPI at 6% historically adds 30–40bp to CPI, or 45–55bp with tariff and margin pass-through, projecting CPI toward 4.3–5% over six to twelve months against a 3.8% headline. Worse, PPI final demand is running 8.9% year-over-year rather than 6%, which projects 6.1–7%. But when he tests his own second-wave call of 12–15%, the model tells him the maths does not get there from a 6% PPI — that would require PPI above 12%, as in March 2022 — and he accepts the correction. The cost he does expect is term premium: other central banks will not credit Treasury-directed policy with inflation-fighting credibility, so the half point currently priced for term on a ten-year plausibly becomes a full point.

2026-05-13 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-13 · open the original ↗

• Every field concentrates in one place, and the historical answer is always to go there. Paul Graham compares Silicon Valley today to Paris for painting in 1870 or Göttingen for math in 1900, arguing the talent pool at a center improves in two dimensions — better people and more of them — which he says produces disproportionately valuable serendipitous meetings that he cannot fully explain but has observed repeatedly in biographies of accomplished people.
• Silicon Valley investors decide faster because competition punishes hesitation, and it doesn't hurt their returns. Graham argues the more confident an investor is about an opportunity, the less time they can afford to wait, illustrated by a Boston VC firm that faxed Dropbox a blank-valuation term sheet only after learning Sequoia was interested — too late, since Dropbox went with Sequoia and became the first YC company to go public, in 2018.
• Comparing yourself to visible "big fish" resets ambition from impossible to merely hard. Graham says moving to a big pond lets founders measure themselves against known peers like Brian Chesky or Sam Altman and conclude they aren't a different species, just people who worked exceptionally hard — turning a fogged-in, seemingly unreachable summit into "a high but definite threshold."
• Founders should go to Silicon Valley, ideally via YC, and then return home — even though data shows returners do worse on paper. Graham argues going and coming back is the best way to help a hub like Stockholm, bringing back money, imported culture and a stronger startup, but acknowledges YC data shows founders who return home are only about half as likely to become unicorns as those who stay — a gap he attributes partly to selection bias and to Bay Area startups raising at higher valuations, not necessarily performing better.

2026-05-13 · open the original ↗

• Marimo pair is a skill that lets agents use notebooks as a tool. Manz describes it as an agent skill that teaches an agent to do anything a person can do in a Marimo notebook, and more. Marimo is MIT-licensed and fully open source.
• Agents lacked access to in-memory state, so notebook work meant babysitting. Manz defines an agent as a model plus tools in a loop. Coding agents are good with files and commands, but data work depends on exploring data loaded into memory, which agents could not see. Using notebooks with agents felt like prompting chatbots a couple of years ago, relaying errors by hand. In software, by contrast, agents run long self-correcting loops of edits and tests.
• Opening Marimo to the agent brings those self-correcting loops to data work. Marimo pair lets the agent work with code and its values in memory, and ask about anything the user can see on screen, by letting it write and run Python inside the runtime.

2026-05-13 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-13 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-12 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-12 · open the original ↗

• His objection to the headline measure is the setup for everything else. CPI printed 3.8%, but core excludes food and energy — a basket nobody actually lives on, given that gasoline, electricity, airfares, fruit and vegetables, hospital costs, vehicle repair and apparel are all rising together. The 30-year crossed 5% on the print and he doubts it returns below, adding that with Powell's final meeting approaching he would not be shocked by a hike.
• The three-wave thesis is on the record, and he restates the number. When rates first started rising he called a three-year inflationary sequence: the first wave as warning, the second as trouble, the third collecting the bill. He is on record putting the second wave at 12–15%, and the model's own analysis lands near 14.4%. His 1970s reference is first-hand — a young broker watching 15% core with headline above 20% — and he notes the official series shows roughly 7% for a stretch he remembers living as nearly 10% or worse.
• The monetary case rests on scale rather than argument. The Fed balance sheet sits at multiples of the $800–900B carried through the Great Financial Crisis, against a few hundred billion before the last inflationary wave. M2 sets a new all-time high monthly at roughly $23.5T against $2–3T in the mid-1980s. Another $1.2T in currency swaps would take the balance sheet toward $8.4T, while Treasury buys its own paper weekly to hold yields down. He also flags that Bessent spent three days in Japan trying to stop Japanese Treasury selling and repatriation — and to locate the $550B Japan pledged to spend in America, which he doubts appears. Bessent called the print transient; his response is that arguing how and why is a distraction from deciding what happens next.

2026-05-12 · open the original ↗

• A taxpayer bailout of Spirit Airlines would fail regardless of terms. Tom Sosnoff argues Spirit's legacy business model is unfixable even with favorable terms, yet he doesn't want to see the resulting 17,000 lost jobs or the loss of a low-cost airline option for consumers.
• Sosnoff proposes the top government contractors buy $500 million in Spirit credits instead of a bailout. He frames it as a business deal, not charity: the credits would be redeemable for flights or later convertible to equity, funded by the same firms that already receive close to $1 trillion in government contracts.
• He argues these mega-cap contractors owe taxpayers for prior support and pay minimal tax. He cites Meta, Google, Tesla and Amazon paying under 4.9% of domestic income in taxes, and notes $50 million in Spirit tickets would move a $5 trillion company's stock price by about a penny.

2026-05-12 · open the original ↗

• Mainframe development's friction was the tooling, not COBOL. Asked where AI fits with COBOL, which the questioner calls "shockingly a growing business", the other speaker blames decades-old, complex applications that lack documentation and tests, along with the developer tools. The speaker says z/OS applications can now use VS Code, Jenkins, automated-test CI/CD, Ansible and Artifactory, the same experience as distributed platforms. The clip is from the 2026-05-11 Hanselminutes interview with Skyla Loomis of IBM Z.

2026-05-12 · open the original ↗

• A three-standard-deviation move — roughly a 1% event — wiped out a well-executed trade. The speaker says he "did everything according to the book" and still lost, characterizing the outcome as statistically rare rather than a result of any mistake.
• Sometimes a stock simply "wins" no matter how correctly a trade is played. The hosts argue this happens some small percentage of the time and call certain losing trades genuinely indefensible, invoking the idea of being "fooled by randomness."
• Rare events like this can't be learned from or planned for. One host notes South Korea hasn't had a three-standard-deviation move in ten years, so the advice is to compartmentalize the loss and move on rather than treat it as a lesson.

2026-05-12 · open the original ↗

• Responding immediately is the single highest-leverage professional habit, he argues. He says he has personally answered every email for 25 years and even replies to comments on his own videos, calling instant response the most meaningful way to build lasting trust and relationships.
• Given equal skills, he says the more decisive candidate always wins the job. Sosnoff frames quick, confident decision-making as the differentiator when "everyone is smart," valuable regardless of whether the call turns out right.
• Expertise, not general financial knowledge, built his firms' credibility, he says. He calls this "nuance know-how": his teams didn't claim to know more about finance than competitors, just more about options specifically, which he says builds confidence in everyone around you.

2026-05-12 · open the original ↗

• Peplink is raising list prices on select hardware. The increase takes effect June 1, making May 31 the last day to order at current prices; it covers select hardware models, not service plans, and the full list of affected models was not expected until mid-May.
• Crosstalk, which resells Peplink, offers a May purchase incentive. Buyers of Peplink hardware through Crosstalk in May can choose 25% off a two-hour support block or 50% off the first month of a Crosstalk Mobile unlimited 5G SIM, and the host offers to lock in current prices for pipeline projects emailed in before May 31.

2026-05-12 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-12 · open the original ↗

• A reactive notebook guarantees the document matches the actual state. Manz says Marimo, unlike Jupyter, tracks dependencies between cells and re-runs stale ones automatically. That removes the overhead of tracking state and of relying on deleted variables. He jokes that it is "disguised Excel" with Python instead of Visual Basic. Marimo is MIT-licensed open source, was acquired by CoreWeave last fall, and is meant to double down on open source.
• Agents must see in-memory state to iterate on data work. Manz says coding agents excel with files and commands but couldn't see notebook state, so users babysat them by relaying errors. Marimo pair is a skill that lets Claude Code, Codex or open code start or connect to a Marimo session, run Python in the kernel, read cells and screenshot them, while the human edits alongside. The result is a reproducible notebook that stakeholders can audit, not logs or loose scripts.
• Skills work best as declarative guides, not step-by-step scripts. Manz says the agent discovers APIs hidden inside Marimo, so the contract is loose and API changes between releases don't break it. His first imperative version of the skill was followed too dogmatically, mishandling exceptions, so it became more declarative: teach the tools and their nuance, and let the model work out how. He dates the rise of skills to around November and Claude Opus 4.5.
• Coding agents are recursive language models, and the notebook is their memory. Manz describes recursive language models as models that offload context to an environment to avoid "context rot", which for today's coding agents is the file system. Marimo pair extends that to Python values in memory. The notebook's cells then form a structured record of the agent's working memory.
• Widgets lower the activation energy of visualization, like a catalyst. Manz built AnyWidget, a standard and Python implementation for interactive widgets across Jupyter, Colab and Marimo. Visualization tools usually need an app download and data export, he argues, while an algorithm is just an import away. A selection made in a widget becomes a variable that reactively triggers downstream cells, so an outlier can be clicked rather than filtered by hand. His "progressive" design lets a widget start as plain JavaScript in a notebook and move into a library, with tooling such as TypeScript added only as it grows.

2026-05-12 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-12 · open the original ↗

• Talking to computers is close to practical. One speaker uses Scotty in Star Trek 4 talking to a mouse to say "we're almost there." The other describes a voice agent, "still in testing", that controls a set of coding agents: it fires off prompts, wakes up to notify, and can be run through headphones while walking around the yard.

2026-05-11 · open the original ↗

• Loomis says today's mainframe is modern, scale-up compute. Loomis, general manager of IBM Z software, compares it to a car: invented in the 1960s, but nothing like a Model T today. Loomis claims eight nines of availability "inherently in the box" (roughly 300 milliseconds a year, by Loomis's estimate), and says the platform runs z/OS or Linux, is built to run dense and hot, and does heavy online transaction processing as well as batch.
• Z is a fit-for-purpose "race car", cheaper the more it carries. Loomis advises against it for small or experimental workloads where a client has no Z already, and for it on core transactional work where an outage risks brand trust, and where data gravity makes copying data elsewhere a security, cost and compliance burden. Boxes ship fully loaded, and clients pay only for the capacity they enable (say 60%), run that at 80–97%, and turn on the "dark capacity" to scale. Loomis's example is Citibank moving its MongoDB from x86 to Linux on Z, reportedly bringing over 50% power savings and meeting cyber-resiliency requirements it could not meet on x86.
• AI belongs on the mainframe because latency decides fraud outcomes. Loomis says many banks accept losing about 8% of revenue a year to fraud, and that inference next to the transaction gives sub-millisecond or single-digit-millisecond responses, enough to score 100% of transactions. IBM added the Telum chip (traditional ML) in z16 about four years ago, and Telum 2 plus PCIe Spyre cards for small-to-medium generative models in z17, announced last year. Hanselman cites 128 GB and 32 cores per card and up to 48 cards; IBM is also bringing agentic AI to operators so they need not use green screens.
• AI helps COBOL developers when grounded in static analysis. Loomis says mainframe development was held back by tangled, undocumented, decades-old applications and by old tools, not by the language; z/OS now supports VS Code, Jenkins, CI/CD, Ansible and Artifactory, and some clients deploy to a CICS/DB2 banking app 20 times a day, with cultural change the hardest part. IBM feeds static and dynamic analysis of how transactions touch DB2 and IMS into the models as context, which Loomis calls its differentiator across code bases of "tens or hundreds of millions of lines". Hanselman's framing is that the LLM orchestrates tools while the development life cycle stays the same; Loomis agrees that good engineering and architecture matter more as code arrives faster.
• Modernization is architecture, not a language swap. Loomis says Java has run on Z for over 25 years with machine-level optimizations for sorting and garbage collection, and claims it "runs better on the Z platform than any other platform, full stop." Loomis asks whether an app moved off the platform keeps its performance, resiliency and latency away from its data, and adds that "COBOL is not that hard of a language."

2026-05-11 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-11 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-11 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-11 · open the original ↗

• The mechanics of trading never stop applying, even when markets go irrational. One host insists the answer is flatly "no," comparing the question to asking whether murder is ever okay — markets can seem crazy a few times a decade, but the underlying mechanics still hold.
• The practical response to an irrational, "moonshot" market is to cut size, not abandon strategy. Being short into a runaway move is called a disaster if held at full size; the stated fix is trimming position count and symbols while staying ready to reverse if the trend turns.
• An extreme move is treated like a five-standard-deviation event that forces a "pull the rip cord" reaction. The host describes acting without necessarily covering every position, while acknowledging nobody can say in advance when "enough is enough" in a runaway market.

2026-05-11 · open the original ↗

• The war has drained the economic content out of Iran's chokepoint, and the numbers show it. Hormuz traffic has collapsed to roughly 5% of pre-war levels — about 150 vessels a month against a normal 3,000 — with flow down from around 20 mb/d to roughly 1 mb/d. But Gulf bypass infrastructure now carries approximately 2.6 mb/d outside the strait: Saudi Arabia's East–West pipeline to Yanbu running at maximum with about 5 mb/d available for export, up from roughly 770,000 b/d in January and February, plus the UAE's Habshan–Fujairah line at 1.8 mb/d. Aramco's Q1 profit jumped 25% year-over-year to $32.5B on that routing, which makes Saudi Arabia and the UAE structural war-winners regardless of the political outcome.
• The inverted implication is the briefing's sharpest point. China imports roughly 33% of its crude through Hormuz against roughly 10% for the United States. As Saudi and Emirati barrels route around the strait, the residual flow tilts increasingly toward China-bound traffic — particularly Iranian crude heading to independent Chinese refineries. So the pressure to reopen now falls more heavily on Beijing than on Washington, which materially changes the leverage available at the table.
• The Blocking Rules invocation twelve days before the summit carries three distinct signals. China has chosen to defend the Iran trade architecture rather than modulate it, accepting short-term dollar-clearing damage to five refineries in order to protect the parallel-payments and state-backed insurance infrastructure that makes the trade viable at scale. The legal crossfire is now formal, with no neutral middle ground for any firm holding meaningful operations in both jurisdictions. And it was established from confidence rather than as pre-summit positioning to be traded away — Beijing does not invoke statutory tools of that magnitude expecting to retract them, which tells Washington that the 16 April commitment not to supply Iran with surface-to-air missiles is the ceiling of Chinese cooperation, not the floor.
• The pre-summit choreography reads as deliberate sequencing rather than coincidence. Wang Yi visited Pyongyang on 10 April, the first such trip in over six years, re-establishing Chinese centrality on the Korean peninsula immediately before Trump's arrival. The Iran concession was banked four weeks early, leaving nothing on that lever to trade. China blocked Meta's $2B Manus AI acquisition after the deal had already closed. MIIT published a draft framework tightening rare earth value-chain enforcement on 29 April while the Busan-era suspension formally held. Xi hosted the Kuomintang chairwoman and put Taiwan declaratory policy on the agenda before Trump landed. And Putin visits Beijing the week after Trump departs.

2026-05-11 · open the original ↗

• His frame is a short-term tactical event embedded inside a long-term structural war, so he grades two tracks separately. He opens on an oddity he says he has never seen in his career: China confirmed the visit only two days beforehand, which he compares to asking someone to prom forty-eight hours out. Tactical asks fill the news cycle; structural asks form the decade, and only one of those is actually being negotiated.
• The Blocking Rules are the structural story and are probably non-negotiable. Previously China practised strategic deception — deny publicly while continuing quietly. The 2021 statute invoked on 2 May replaces that with legal cover criminalizing compliance with US sanctions inside China, carrying jail exposure. Beijing is signalling it will accept sanctions on five refineries and real dollar-system friction in order to keep the parallel settlement architecture intact, which tells you what it actually values.
• He commits to explicit probabilities, which makes the call checkable afterward. Roughly 80% on the short-term package: tariff modification, a Boeing deal, agricultural purchase commitments, rare earth licensing acceleration, a joint Iran statement carrying no operational commitment, fentanyl cooperation, and both leaderships claiming visible deliverables. Roughly 95% on structural impasse: no Taiwan declaratory shift, no Blocking Rules retraction, no AI training-chip access, no formal US recognition of China's payment rails. He expects the Blocking Rules to extend to new sanction categories within six to nine months. And one supply detail lands harder than the rest — his claim that the US has been building F-16s without radar since the previous July for lack of gallium.

2026-05-11 · open the original ↗

• Engineers should answer governance worries by asking what security needs to see. Gupta's questions for a zero-trust world are: how to get visibility into what agents are doing, how to govern what they can and can't do, and how to recover when they inevitably err. She says many organizations are deadlocked between security listing risks and engineers needing to build. Engineers can break that by asking what visibility, policies and resilience plans would make security comfortable.
• Enterprises lack AI literacy training on a par with privacy training. Al-Dhubaib says annual privacy and cyber training instils a sense of obligation about consequences. He sees no widespread equivalent that teaches workers the larger blast radius of the AI tools they now hold.
• AI-enabled knowledge work is cross-functional by default. Al-Dhubaib points to Rubrik's dedicated AI attorney, who began by asking about risk exposure and then upskilled through specialist conferences. He argues data scientists and engineers likewise need to learn the language of risk management.

2026-05-11 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-11 · open the original ↗

• U.S. sanctions on three Chinese satellite firms just before Trump's Beijing summit function as pre-negotiation leverage. The firms allegedly supplied Iran with imagery of U.S. and allied military facilities; the author frames this as one piece of a broader secondary-sanctions strategy targeting shadow-fleet tankers, teapot refiners, and Chinese banks that could pressure Iran financially, though China's near-monopoly on rare earths makes the strategy not risk-free.
• AOC's viral claim that "you can't earn a billion dollars" is culture-war framing that obscures the real lever — the tax code. The actual policy question is whether to tax mechanisms like margin loans against stock portfolios and stepped-up basis for heirs, versus continuing to tax W-2 income near 40% while carried interest gets preferential rates; moralizing about billionaires doesn't touch this.
• Discovery in the Musk v. OpenAI lawsuit is doing the accountability work no AI regulation exists to do. Produced journal entries from OpenAI president Greg Brockman ("This is the only chance we have to get out from Elon... what will take me to $1bn?") reveal self-interested motives behind a "mission-driven" company; Musk is seeking $150 billion returned to the nonprofit and removal of Altman and Brockman, though the author doubts either remedy will actually happen.
• A "bonus" item notes that even Big Law isn't immune to the information-barrier failures the author saw enforced strictly in banking. Attorneys at top firms allegedly traded tips gathered across roughly 30 deals, including information on deals they weren't even assigned to, which the author attributes to inadequate access controls rather than merely insufficient training or policy language.

2026-05-10 · open the original ↗

• The document opens by capping its own thesis, which is what makes the rest credible. Argentina has world-class positions in two factors, structural weakness in technology and demographics, and chronic institutional fragility at the financial layer — so no amount of good leadership produces a Node State or a true monopoly. The realistic ceiling is a stable middle power with durable resource rents, institutional credibility comparable to Chile or Uruguay, and asymmetric bilateral relationships compensating for limited unilateral leverage. "Effectual leadership" here is not strategic creativity but execution discipline across a multi-decade windfall with three or four genuine decisions made well. Argentina has had structurally similar moments before — the 1990s convertibility era — and squandered them. The leader's job is mostly not to ruin this one.
• The five-year plan is execution of infrastructure already permitted and financed rather than new strategy. Complete VMOS at 550k b/d by July 2027, finish the Punta Colorada VLCC export terminal, bring at least one major LNG facility online, and push aggregate production to 1.3–1.5 MMb/d by 2030 against a 1M-by-2027 baseline. Negotiate long-dated offtake with Japan, Korea, India and Europe as project-finance hedges — with the strategic insight that being structurally non-aligned with OPEC+ discipline is itself the asset, so join no cartel-equivalent structure.
• One diagnostic marker is named, and it is genuinely falsifiable. Whether net reserves cross $50B by 2028 is the structural threshold for a durable band without permanent IMF life support: below it the peg is forever conditional, above it Argentina enters a different category. The instruction is to use the projected $14B+ 2026 energy surplus and the renormalized agricultural cycle to accumulate dollars conservatively rather than distribute them politically, and to hold the reduced export-tax regime — retenciones being, historically, the single largest fiscal lever destroying agricultural investment incentive.
• Two genuine strategic decisions are isolated from the execution work. First, converting the US bilateral into something that outlives one administration: the ESF intervention and $20B Treasury swap line are administration-specific, and the window to institutionalize them runs 2026–2028, before the next transition, after which the relationship becomes whatever the incoming administration decides. Second, making a real Mercosur decision — Argentina cannot pursue bilateral deals with the US, Israel or the EU while bloc consensus rules bind, so it must either renegotiate for flexible bilaterals against Brazilian resistance or exit cleanly. The current half-membership is explicitly identified as the worst available position.

2026-05-10 · open the original ↗

• The classification is a Pegged Peripheral State whose peg is structurally unstable rather than stable. Four non-overlapping chokepoints — Rosario soybean processing, Vaca Muerta energy, the Lithium Triangle and Drake Passage adjacency — layered over a chronically fragile financial architecture now externally underwritten by US Treasury support. The composite of roughly 3.2/5 materially understates stress-scenario value, with the active reading reaching about 3.8 during the early phases of Iran resolution.
• The two-instance comparison is what gives the typology teeth. The UAE runs a hard, durable, self-funded peg held since 1997 and backed by $480B+ of sovereign assets, able to defect optionally. Argentina runs a crawling band roughly 24% wide with chronically negative net reserves, IMF-financed — its 23rd programme, a $20B four-year facility from April 2025, plus a $20B US Treasury swap line via the ESF in October 2025 and a prospective private bank facility. The framework reads this as a fiscal moat externalized as a foreign-policy instrument, the exact inverse of Norway's sovereign-internal version, with alignment forced rather than chosen.
• A new mechanism candidate is proposed and honestly held as provisional. Defection-from-Defection: a peripheral state that had rotated into a non-USD settlement layer — the Kirchnerist yuan swap lines across 2014–2023 — rotating back to dollar anchoring under regime change, supported by direct US Treasury intervention. Argentina is the first instance and the concept is held as candidate pending a second. The entry also reads it as the largest empirical confirmation to date of dollar-support doctrine being deployed pre-emptively as state policy rather than in response to crisis.
• The four leverage layers are real but each has a named vulnerability. Gran Rosario handles roughly 95% of Argentine soymeal and soyoil across 67 MMT of crushing capacity in 344 facilities, and the structural distinction from Brazil and the US is that Argentina exports processed meal, oil and biodiesel rather than raw beans — value-add sovereignty achieved operationally rather than demanded as policy, with Paraguayan and Bolivian origin routed through for re-export. Vaca Muerta produced 874k b/d in February 2026 at a $36–45 breakeven. But the binding physical constraint is water: Paraná levels ran critically low from 2020 to 2023, repeatedly forcing partial loading and rerouting, which the entry treats as the closest thing to a physical-layer mismatch — financial commitments at processing capacity existing independently of the river's ability to deliver and dispatch.

2026-05-10 · open the original ↗

• The question behind the piece is three years old: do the winners of the last era win the next one? China, the United States, Japan and Germany did best under globalization, and he is careful not to pretend otherwise — the system America built and enforced made it the richest and most powerful nation for eighty or ninety years. His view is that those four will be challenged in the transition, and that China in particular represents the culmination of the old order rather than the shape of the new one. So he went looking for the era's worst performer, and Argentina qualified easily: nine IMF bailouts in forty years plus roughly two dozen other facilities, before counting bilateral help from Washington and Beijing.
• The detail he flags as genuinely distinguishing is the processing step, not the resource. Argentina is 5/5 on food and, unlike almost every other exporter including the United States, does not simply ship beans — it exports soymeal at roughly 35% of global share and soyoil at roughly 42%. That value-add step is the chokepoint. On energy it holds the fourth-largest shale oil and second-largest shale gas resource with the extraction infrastructure still being built out. Four chokepoints in total, with the Drake Passage the least commercially meaningful since almost nothing transits it.
• The bifurcation is the punchline, and it is the same one he keeps arriving at. Chinese capital dominates the Argentine lithium leg with ownership stakes running 50–67%, and Chinese firms are building the soybean and energy infrastructure. So when Bessent demanded Argentina unwind its $20B currency swap with China, it could not — it needed the money — and Treasury supplied a swap line of its own instead. The result is the familiar split: America is the paper, China is the physical. Argentina cannot remove China because the US has neither the corporate capital willing to go nor the state apparatus to send.

2026-05-10 · open the original ↗

• Equities hit fresh records — the S&P's sixth straight weekly gain to 7,398.93, the Nasdaq to 26,247.08 — even as the Strait of Hormuz blockade entered a critical four-to-eight-week window risking spreading shortages across Africa and Asia. US gasoline topped $4.50 a gallon and diesel neared record highs even as consumer sentiment hit another record low, widening the gap between capital markets and households.
• Central banks turned firmly hawkish, with the Reserve Bank of Australia and Norway both hiking on sticky inflation while markets pushed Fed rate-cut expectations far into next year. This came alongside growing reassessment of outgoing Chair Powell's legacy, weighing his defense of institutional independence against missed opportunities on inflation, communication and policy-framework reform.
• The week ahead pairs a high-stakes US CPI/PPI release with the US-China Summit and a Congressional vote on Kevin Warsh's Fed nomination. Consensus expects headline CPI at 3.7% and core at 2.7%, with PPI at 4.8% and 4.3%, while material breakthroughs from the Summit look limited and UK political turmoil deepens after Labour's poor local-election showing.

2026-05-10 · open the original ↗

• Agent Rewind can undo system changes, not delivered messages. Gupta says the product draws on Rubrik's cyber-recovery base. It can restore a dropped production table, repair a Salesforce change that broke the link between opportunities and accounts, and reverse identity changes that granted excess access. An email that has already reached a customer cannot be recalled.
• Irreversible actions need runtime blocking; ambiguous changes need rewind. For high-risk actions such as trading or emailing customers, Gupta says the right defence is defining them up front so Sage, Rubrik's policy engine, blocks them at runtime. System changes are hard to judge at runtime because most are legitimate and telling them apart needs more context, so rewinding after the fact fits them better.
• Enterprises need four or five overlapping layers of defence. Gupta likens the layers to Swiss cheese slices, stacked so that together they leave no holes.

2026-05-09 · open the original ↗

• The document rejects the obvious framing in favour of a more precise one. The claim is not that Balkan states must choose between Brussels and Ankara — that overstates Turkish reach and understates EU institutional stickiness for fully-integrated members. The precise thesis is that Turkey is being pulled into a regional hegemonic role by EU institutional weakness creating a vacuum, not by autonomous ascent. Turkish capacity is genuinely constrained: lira instability persists, domestic political consolidation consumes institutional bandwidth, and refugee politics limit border posture. Turkey rents geographic position rather than projecting power. The vacuum forms from the western side, not from eastern pressure.
• Three tiers carry structurally different exposure to the same dynamic. Tier 1 is locked to Europe — Greece, Slovenia and Croatia, holding the EU, NATO and eurozone trifecta with energy diversified through Revithoussa, the Alexandroupolis FSRU and Krk. Greece is doubly locked, since Cyprus and Aegean maritime antagonism fixes its choice independent of any EU/Turkey calculus. Tier 2 is functionally hybrid — Bulgaria, Romania, Albania, Montenegro and North Macedonia — formally inside the institutions while functionally dependent on Turkey for energy gateway via Strandzha 2 and TANAP/IGB, defence procurement including Baykar drones, and labour and trade gravity. Tier 3 is genuinely contested — Serbia, Republika Srpska and Kosovo — where the choice is being negotiated in real time and where Russia could re-enter fastest.
• The Tier 2 bind is demographic, and the framework names what neither side can fix. Replacement rates run around 75 in Bulgaria with similar collapse in Romania and structural emigration from Albania into Italy and Greece. The EU offers regulatory framework, capital access and a security guarantee — and zero structural answer to demographic collapse. Turkey offers commercial gateway, demographic depth and defence-industrial complement. That is why the pull is real even where the formal alignment is fixed and unlikely to change.
• The resolution variable is EU coherence rather than Turkish strength, and it comes with numbers attached. Roughly 38% for a polycentric toll order in which Tier 2 drift stays slow and Tier 3 stays contested, against roughly 45% for bloc fragmentation, in which Turkey fills the vacuum simply because nobody else competes for it — resolving over a 24-to-36-month window. Russia is scored suppressed rather than eliminated: TurkStream capacity persists, transit waivers continue for Hungary, Slovakia and Serbia, and Orthodox Slavic ties remain across Serbia, Bulgaria, Montenegro and North Macedonia, making Russian return the standing tail risk on every Balkan thesis.

2026-05-09 · open the original ↗

• Bulgaria is the cleanest case of a state holding three real chokepoints with almost no discretion over any of them — "the geography persists, the sovereignty does not." Gas transit fees, Lukoil refining margins and grain port revenue are all genuine toll positions, and the income is real. What has been captured is the autonomy over how that income is collected, priced or directed. The entry adopts the weak-toll-state pattern with a new modifier: Cambodia is weak because it is a Chinese client, Laos because of debt, Bulgaria because of bloc enforcement. Identical mechanic, different leveraging party.
• Strandzha 2 is the single most important structural fact, and the leverage over it is not Bulgarian. After Ukrainian transit ended on 1 January 2025, TurkStream became the sole remaining pipeline route for Russian gas into Europe, with Strandzha 2 on the Turkey-Bulgaria border as the EU ingress. 2025 deliveries reached 16.8 Bcm against a nominal capacity of 15.75, flowing onward to Hungary, Slovakia and Serbia. But the US Treasury issued the waivers in December 2024 and March 2025 permitting continued payment to Gazprombank — Bulgaria received that structure rather than negotiating it — and the EU's 2027 phase-out target sets the terminal date on the revenue. Sofia collects fees while setting neither tariff, route, nor counterparty.
• Three structural shifts inside fourteen months compressed the loss of autonomy simultaneously. Lukoil-Burgas, supplying roughly 80% of domestic fuel, was placed under a state special administrator in November 2025 — de facto nationalization under OFAC pressure. Euro adoption ended the lev currency board. And continued OFAC waivers now govern the gas transit. Three separate layers of sovereignty subordinated to bloc enforcement in the same window.
• The Hungary contrast is the analytical anchor, and demographics are the floor beneath everything. Hungary is the canonical toll state, extracting rent through multi-vector political behaviour across the EU, Russia, the US and China at once. Bulgaria's multi-vector option was foreclosed by EU and NATO membership and is now operationally enforced — the same Russian gas corridor, opposite autonomy profiles. Underneath sits a demographics score of 1/5: population 6.42M at end-2025, a replacement rate of 75, and a UN flag as the fastest-declining population globally, which the entry marks explicitly as non-reversible.

2026-05-09 · open the original ↗

• The title is the argument: a nation can print money but not people. The Balkans went from 35 million to 28 million between 1990 and 2023, a 19% decline, with Bulgaria down 26% to 6.4 million. These are small states to begin with — five, six, seven million — all declining simultaneously, and all strongly opposed to immigration. He deliberately avoided the region until asked by a follower, noting it has been a battleground contested by the Austro-Hungarian, Russian, Ottoman and Greek empires for centuries.
• His three-year-old call was that every country on earth would need a policy to attract immigrants, and he defends it on arithmetic rather than preference. Italy needs 1.5 to 2 million people a year under 25, for twenty years, simply to still have 56 million people at the end — while Meloni cut arrivals from 150,000 to 50,000 and counted it a success. Japan's LDP admitted for the first time that it needs 800,000 workers, while running the most demanding integration regime anywhere: language, cultural training, assigned placement. Merz has said Europe is short 5.6 million trained workers, and the EU has now presented a five-year migration blueprint with €81B attached, explicitly to open legal routes for workers it needs.
• The concrete cost he points at is a factory, not a forecast. TSMC's Arizona plant runs at 55% capacity against 95% for the identical facility in Taiwan — which he characterises as roughly $40B of stranded capital for want of people with the skill set, with 4,000 to 6,000 workers needing four to six years of training. He also flags a separate development the framework called three years ago: Iran has declared underwater cables critical maritime infrastructure and is asserting control over them, with roughly 90% of east-west traffic running through the region, after years of interference in the Baltic, around Taiwan and in the Red Sea.

2026-05-09 · open the original ↗

• Non-deterministic agents must be governed by AI, not rule engines. Gupta frames Rubrik's approach around three pillars: visibility, governance and control, and remediation. She says over 80% of organizations feel they lack visibility into which agents are running and what they do. Because agents are non-deterministic, a simple rule engine can't monitor them.
• A fine-tuned small model wins on accuracy, cost and latency together. Sage, Rubrik's Semantic AI Governance Engine announced at RSAC in March, is built on a fine-tuned SLM. The host describes Rubrik's published head-to-head as five times faster than GPT 5.2 and more accurate at detecting policy violations. Gupta says an enforcement layer can't wait 5 seconds per decision when agents chain many actions. The capability came from Rubrik's acquisition of Predibase about a year ago, bought to add model and fine-tuning expertise to its data and identity knowledge.
• Rubrik Agent Cloud pairs agent discovery with natural-language policies and rewind. Gupta says it gives visibility across Microsoft Copilot, AWS Bedrock and Claude Code agents. Policies such as "no financial advice" are written in natural language and enforced by Sage at runtime, and Agent Rewind undoes unintended actions. Al-Dhubaib likens the discovery side to e-discovery for agents: a dashboard revealing hundreds of agents, 50 of them with read-write access to four databases.

2026-05-08 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2026-05-08 · open the original ↗

• Many AI tools hide code; this one puts files first. One speaker praises an unnamed tool as feeling like "a software development life cycle tool", not vibes, because it puts the file viewer and diffs up front. The other, speaking for its makers, says the end product includes the designs as well as the code, so the work should stay "local first, open formats first" on disk rather than in a cloud database behind "a thin view".

2026-05-08 · open the original ↗

• NATO exists to contain a USSR that never left occupied Eastern Europe. Formed April 4, 1949, he says, after the USSR absorbed East Germany, Poland, Czechoslovakia, Hungary, Romania and Bulgaria into its sphere rather than withdrawing; weaker Western militaries needed US backing, hence the tens of thousands of US troops still stationed in Europe today.
• Russia's post-1991 record shows the expansionism NATO was built to check. He points to Transnistria in Moldova in the early 1990s, a slice of Georgia in 2008, Crimea in 2014, and the now four-year invasion of Ukraine as continuous evidence Russia kept pushing its borders outward after the USSR fell.
• He rejects the claim that NATO's growth, not Russia, caused today's standoff. His counter: the USSR's own bloc, the Warsaw Pact, dissolved when the Soviet Union collapsed, and most of its former members chose NATO membership afterward specifically for protection against Russia.

2026-05-08 · open the original ↗

• Smith accused Hegseth of contradicting himself on Iran's nuclear threat. In a televised hearing clip, Smith noted Secretary of War Pete Hegseth had justified starting the war 60 days earlier by calling Iran's nuclear weapon an "imminent threat," then later claimed the facilities had been completely obliterated — undercutting the original rationale.
• Adam Smith is a pro-deterrence critic, not an anti-war dove. Smith has sat on the House Armed Services Committee for 29 years, chaired it from 2019 to 2023, and now serves as ranking member; he supports a strong US military but argues this administration's approach to the war backfired.
• Smith argued the Iran war's costs have outweighed its benefits so far. He credited the strikes with reminding Russia and China of US military capability, but noted Iran still managed to attack and effectively strangle the Persian Gulf, driving up oil and fertilizer prices and, by extension, US food costs.

2026-05-08 · open the original ↗

• Dollar dominance took decades to follow US economic dominance. The US economy overtook the British Empire's by around 1910–1912, yet central banks held almost no dollars in 1905, favoring the pound, franc and Deutschmark instead; dollars only entered forex reserves after World War One, in 1918.
• The pound's reserve share swung sharply against a rising dollar. By 1926, 75% of world forex reserves were dollars versus under 23% pounds, he says, but the Great Depression reversed this: by 1935 pounds held 55% to the dollar's under 40%, and by 1942 pounds reached 74% again.
• Formal dollar reserve status in 1944–45 didn't mean actual dollar dominance. Even after the wartime Bretton Woods agreement made the dollar the recognized world reserve currency, 82% of forex reserves were still held in pounds; dollars did not actually exceed pound holdings until 1953–54.

2026-05-08 · open the original ↗

• Lee's tactical brilliance masked a strategic failure that doomed the Confederacy. He argues Lee won most of his battles, often against long odds — citing Fredericksburg, Chancellorsville, Cold Harbor and the Second Battle of Bull Run — yet still lost the war, comparing it to the US failure in Vietnam.
• The Confederacy's smaller manpower pool made its casualties unsustainable, even in victories. With only 750,000 available Confederate soldiers against 2 million Union troops, he says, the South lost more than a third of its men versus under a fifth for the Union — a disproportionate toll even in Lee's own decisive wins.
• Lee's constant costly battles undercut Jefferson Davis's preferred war-of-attrition strategy. He says Davis wanted Lee to preserve troops and prolong the war until the Union's economy collapsed, since the South was more resilient on that front — but Lee kept spending resources the Confederacy couldn't replace, so even winning Gettysburg wouldn't have changed the outcome.

2026-05-08 · open the original ↗

• NATO tolerates Russia in Ukraine to avoid a wider collapse. His reasoning: Russia holds over 5,000 poorly maintained nuclear weapons, and a breakup into smaller states risks some falling into the black market, creating rogue nuclear states the world would have to negotiate with.
• China also needs a stable, whole Russia, not a fractured one. He says China has bought so much Russian land and resource rights that parts of Russia are functionally Chinese already, and since Russia is China's top energy supplier, instability there threatens China's own plans.
• Russia is going through another ~30-year historical reset, and a costly one. He traces resets in 1917 and 1945 before the 1991 Soviet collapse, and says the current war has cost Russia 1.3 million casualties (over 2% of its male population), falling birth rates and life expectancy, banned gasoline and fertilizer exports, and roughly 150,000 farm workers a year lost to service.

2026-05-08 · open the original ↗

• He calls the Navy's Hormuz mission bait to provoke Iran. A single damaged tanker, hit by an RPG from a speedboat or a drone, would be enough to make the Strait feel unsafe again, he argues, giving the US a pretext to resume fighting toward goals it hasn't secured — ending Iran's uranium and ballistic missile programs and reopening the Strait.
• A ceasefire-period military buildup signals the US expected to resume fighting. He points to a surge in transported hardware — C-17 Globemasters, Stratotankers, fighter jets, and a third aircraft carrier strike group sent to the region — as signs the administration was preparing to reopen the war theater rather than close it out.
• US allies are sitting this conflict out despite Iran's illegal closure. He notes closing the Strait of Hormuz violates UN Law of the Sea Article 87, and a UN vote to force it open was vetoed by Russia and China, yet allies unhappy with how the war started have not joined a coalition effort.

2026-05-08 · open the original ↗

• Helium is irreplaceable and can't be manufactured, only extracted from natural gas. He says 15% of global helium use is for MRIs, 17% for semiconductor production and 22% for scientific research; it forms only through millions of years of underground radioactive decay, and once released it permanently escapes Earth's atmosphere.
• Iran's Hormuz closure has squeezed global helium supply hard. Qatar is one of the world's largest helium producers behind the US, he says, and Iran's destruction of one of its processing plants puts roughly a third of world supply at risk, boosting profits for the US, Russia, Canada, Poland and Algeria.
• China's tech sector dependence on imported helium is framed as leverage. Despite producing some helium domestically, he says China consumes far more than it makes and remains import-dependent, a vulnerability the US could exploit especially amid worries about China restricting rare-earth exports.

2026-05-08 · open the original ↗

• Iran struck the UAE's Fujairah oil facility, per early reports. UAE Embassy alerts confirmed a third or fourth wave of drone and missile attacks before the story reached traditional news outlets; a separate UAE vessel was also reported on fire near port, with the cause still unclear at the time.
• He links the strike, tentatively, to new US Hormuz rules. The US had just announced Navy escorts and authorization to "take out threats" near the Strait, successfully passing two US-flagged ships through; he floats but doesn't settle on whether the UAE was targeted specifically for siding with Washington.
• He predicts Iran's UAE attack backfires as a US pretext. Oil prices rose only a modest 2 to 4.5%, which he suggests reflects markets having already priced in renewed hostilities; he expects the UAE strike to work against Iran's interests going forward.

2026-05-08 · open the original ↗

• Iran's Hormuz closure alienated the countries it needs as friends. Closing the Strait spiked oil, energy, agriculture and fertilizer prices as short-term retaliation against enemies, he says, but it forced friends and foes alike, including China and India, to start diversifying away from reliance on the chokepoint.
• New pipeline routes only cover a fraction of Hormuz's flow. He cites Saudi Arabia's East-West pipeline (roughly 5 million barrels a day already, with 2 million more in capacity), Iraq's Kirkuk-Ceyhan route through Turkey, and the UAE's Fujairah pipeline via Oman as short-term bypasses adding up to only 5–6 million barrels a day, against the 20 million normally moved through the Strait.
• He expects Iran to permanently lose its chokepoint leverage. Over the next 5 to 10 years, he predicts the US, Canada, Norway and Gulf states will build lasting alternative supply chains, while China, India and Japan invest in Oman's pipeline and port infrastructure and expand tanker fleets sized for the Panama Canal to move Venezuelan oil.

2026-05-08 · open the original ↗

• He alleges the war has served as insider-trading cover. He claims the administration made large energy investments before the war began, timed to gains when oil and gas prices spiked at the outset and again when markets rallied on news that Operation Freedom was being paused amid Pakistan-brokered talks with Iran.
• Operation Freedom failed militarily even as it paid off politically. US forces sank Iranian speedboats attacking vessels near Hormuz, but he says enough attacks got through to keep the Strait feeling unsafe for shippers, undermining the operation's actual objective even as markets rewarded the administration.
• He speculates new conflicts are meant to distract from "the files." As his own explicitly-flagged theory, he suggests a Cuba regime-change operation could follow, then possible saber-rattling with North Korea, China or NATO allies, arguing sustained global instability could let Trump escape accountability by later heading a "board of peace." He invited readers to challenge the argument.

2026-05-08 · open the original ↗

• Ukraine's top trading partner is also Russia's biggest one: China. He notes China supplies Russia with much of the dual-use equipment it needs to continue the war against Ukraine, making the Ukraine-China trade relationship a striking data point on the map.
• Tiny Netherlands is the top trading partner for three countries. Despite occupying roughly 1/100th the land area of the United States, he says, it is the top trading partner for Germany, Belgium, and Iceland, and the world's third-largest agricultural exporter, behind only much larger countries.
• Per-capita export figures flip the picture of who dominates world trade. China leads in total exports at $3.3 trillion a year versus the US's $2 trillion and Germany's $1.7 trillion, he says, but per capita China exports about $2,000, the US about $6,000, Germany about $20,000, and the Netherlands about $50,000.

2026-05-08 · open the original ↗

• Saudi Arabia and Kuwait halted Operation Freedom by suspending base access. Neither country needed to use oil as leverage, he notes — cutting off access alone was enough, since the US Middle East air campaign depends heavily on Saudi and Kuwaiti basing given their proximity to Iran and the Strait of Hormuz.
• True power projection depends on coalition, not just military size. He faults the Trump administration for reportedly launching Operation Freedom without contacting any Persian Gulf countries beforehand, arguing that without a cooperative coalition, even the world's most powerful military can be brought to a halt.
• He calls this another US military and diplomatic failure. Closing with the line "I fear an army of sheep led by one lion than an army of lions led by one sheep," he stops short of predicting the US will ultimately lose the broader war.

2026-05-08 · open the original ↗

• Garry Tan rebuilt the same blog platform three times, and the cost collapsed by roughly 20,000x. He says the 2008 version (Posterous) took $4 million, six or seven people and about a year and a half; a 2013 rewrite took roughly $100,000, two people and three months; the 2026 version, built with a Claude Code Max subscription, cost about $200 and five days while adding full agentic retrieval and research capability the earlier versions never had.
• "Boiling the ocean" — paying for exhaustive completeness — is now cheap enough to be standard practice. Tan argues that because an agent can cross-reference twenty sources as easily as one, builders should stop economizing on research or code depth the way a human effort-budget would force them to, calling this "token maxing" and arguing it will spread from coding into every kind of knowledge work.
• A "thin harness, fat skills" philosophy puts judgment in markdown, not code. Tan describes building G Stack around markdown skills — a "CEO" review for ambition, a plan-and-engineering review, a developer review, and browser-based QA — while keeping the underlying agent harness minimal, arguing that anything requiring human-style judgment belongs in plain-language markdown rather than brittle deterministic code, and that skipping 80-90% test coverage produces output "10x worse than human-written code."
• Measuring logical lines of code, Tan found he was producing about 400 times his 2013 output, directing roughly 15 agents at once, even though the code he personally typed dropped about 70% from his 2013 baseline — a distinction he says got lost when his "100x" claim drew online criticism. He frames the larger stakes as builders needing to control their own prompts and tools rather than depend on centrally controlled AI products, calling it "the defining question" of the moment.