Souk's Notes - Unofficial Summaries

2026-06-17 · open the original ↗

• Commitment insurance is pitched as good for clouds and customers alike. One speaker says customers get "an out" on unused commitments and can reinvest the money rather than getting mad at Amazon or Azure, likening it to Ford wanting Geico to exist so people will buy the F-150.
• Abuse is held off by underwriting against moral hazard. Asked how they stop a failing startup buying cover just before it collapses, the speaker says the answer is underwriting, which exists partly to avoid moral hazard: buying insurance and then taking more risk because of it. The clip ends before any specific safeguard is described.

2026-06-17 · open the original ↗

• Mark holds that AI will not eliminate software engineers. He relays asking Ion Stoica of Berkeley at a Build session, who answered no at once, for the reason Mark gives himself: nothing can be fully specified, and where the spec has gaps AI makes things up that may violate implicit assumptions. Mark calls engineering sculpting, needing feedback from use, and concedes it is opinion he cannot win from first principles; Hanselman adds that spec taken far enough is just programming ("Word documents don't break the build").
• System complexity decides where AI can stand in for engineers. Mark says simple, well-understood systems can be built from a few sentences but medium complexity "falls apart". Hanselman's analogy, from a meeting with Japanese CTOs, is IKEA against Japanese carpentry; Mark maps IKEA to low-complexity web-plus-database apps and bespoke state retirement systems, and Hanselman offers his podcast admin site: "classic late 2025 software" with a god object and a thousand tests, fine for an audience of one.
• Evaluation is the bottleneck that will not go to zero. Mark says evaluation means the tests plus human oversight of the output, and that a moderately complex Azure service has tests taking hours, so a commit that kicks off 5 hours of tests forces more evaluation up front, costing time and money. "Cost of software is going to zero" comes, he says, from people seeing only the simple end; both hosts say they could be wrong.
• Hanselman rejects a skills file as a shortcut to seniority. Mark puts the "controversial take", heard from smart people, that a skills file turns a junior into a senior. Hanselman is irritated by "you're an expert" prompting ("I am more than the sum of my context") and unsure whether to be offended by a request to turn his 20-plus years of blog posts into a skill.
• The bar for shipping depends on who ships it. Hanselman calls his OpenClaw Windows companion app alpha despite weeks of trouble-free use, since odd builds or non-English systems will break it; Mark says a solo open-source project can have rough edges but software from Microsoft to many people cannot. Mark relays Simon Willison trusting only software used by many people for a while, and Hanselman says to treat LLM code like a stranger's pull request.

2026-06-17 · open the original ↗

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

2026-06-17 · open the original ↗

• Founders should stop hunting for the "perfect" idea and commit to one. John, a YC partner, says overthinking startup ideas — either waiting for the ideal concept or doubting your own founder-market fit — produces paralysis; citing Boom Supersonic's Blake Scholl, who came from ad tech with no aerospace background, he argues deep customer knowledge can be built quickly if you commit and go deep.
• "Burning the other boats" means fully abandoning alternatives, not just prioritizing. He points to GovDash, which pivoted at least five times, changing its name and even email addresses each time, before its fifth idea in government-contract software worked so well the company raised a Series B — evidence, he argues, that half-committing to multiple ideas at once produces unreliable signal.
• The bar for "going deep": could you actually run your customer's business? John's test is whether you know their daily crises, whether a specific problem is a top-five pain point, and what they'd pay to solve it — a level of knowledge he says should let you "teach a class" on the problem, built through tight loops of customer conversations and shipped product together.
• Good AI-era startups should verticalize into owning outcomes, not just software. He argues software costs are heading to zero, so durable value shifts to trust, licenses and outcome ownership — citing Corgi Insurance, which acquired an insurance carrier during its YC batch rather than staying a tech-enabled broker, as a model for "being the insurer" instead of selling tools to insurers.

2026-06-17 · open the original ↗

• When the maritime commons stops being a guaranteed public good, geography returns as the binding constraint on power. Political strategy can open doors and accumulate goodwill, but it cannot manufacture reach or compel a guarantee another nation has chosen to stop providing. The Iran conflict is the demonstration: it presented as political — regime, enrichment, civilisational stakes — and was decided as geographic, ending at a twenty-one-mile aperture where the only question was who could keep crude moving and who could survive it stopping. The sanctions architecture had already reached the end of its travel; it could still inflict pain but no longer dictate terms. Ideology was the costume; the factors were the actual terms.
• The eighty-year era was rent, and most tenants mistook the subsidy for a law of nature. One navy underwriting every sea lane flattened the map, so distance stopped billing anyone and alliances could be drawn anywhere regardless of terrain. Land empires never received that subsidy and never forgot the older arithmetic — a continental power holds what it can physically reach. But the paper is careful not to overclaim: contiguity is a multiplier, not a winning side, since fourteen borders is also fourteen potential fronts, and a hollow core hemorrhages through its frontiers rather than projecting through them.
• The retrenchment is real, staged deliberately, and Canada is the cleanest refutation of political strategy. June 2026 cuts to European commitments reduced committed fighters by roughly a third and thinned maritime reconnaissance — but the disclosure was withdrawing every aerial-refuelling tanker, because refuelling is reach, the connective tissue letting power travel past the horizon and stay there. Canada has a genuinely strong factor base yet touches exactly one neighbour, earns roughly seven of every ten export dollars selling south, and would have to double combined sales to China, Germany, France, Mexico, Italy and India at once merely to move a tenth of its exports — every one of them across water no longer guaranteed. So it is not diversifying; it is swapping a continental dependency it can drive a truck across for a maritime one requiring the exact commons being withdrawn. The paper coins a typology for it — the Vault Node, high on four factors and hollow on Security — and stages the reversion in three regimes: repricing by ~2031, regionalisation with in-region providers by ~2036, and a multipolar continental order supplanting the treaty map by ~2051.

2026-06-17 · open the original ↗

• His organising claim is a dating: before 2019 decisions were political, after it they are economic. He recorded a video the day before the Iran war began saying it was the wrong war on Five Factor grounds. His account of how it resolved is the argument in miniature — it began political, turned economic the moment Hormuz closed, and then the only live question was whether an undefined political win exceeded a very defined economic loss. It did not, and that is what forced the end.
• The European drawdown numbers are specific, and he singles out the least-covered one as the real signal. Fighters cut from 150 to 100, P-8A maritime reconnaissance from 26 to 15 — a 42% reduction — all eight KC-46 aerial refuelling tankers withdrawn entirely alongside carrier and warship reassignment, plus 5,000 troops out of Poland. Everyone covered the troops; he argues the tankers matter more because they are attached to a task force that is leaving. His blunt summary of the alliance position: NATO exists in name only.
• He translates the framework into an instruction for reading decisions. Over five years to 2031 the guarantee frays but persists, so geography becomes a repriced risk factor — every commitment now carries the question of whether the guarantor can actually arrive. By 2036 sea-lane security is partly privatised with in-region providers emerging by necessity, Turkey for the Sunni near abroad being his example. By around 2051 a multipolar continental order organised around geographic clusters supplants the treaty map. The practical takeaway he leaves: any country or company still making decisions on political rather than geographic and economic grounds is making a mistake.

2026-06-17 · open the original ↗

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

2026-06-17 · open the original ↗

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

2026-06-16 · open the original ↗

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

2026-06-16 · open the original ↗

• Most teams measure activity instead of customer value. Shane says logins, sign-ups and revenue go up and to the right without showing users got value, which shows up as a vanity metric or a dashboard of 30–50 numbers no one rallies around. Because the North Star sits above every decision, a wrong one passes its mistake to everything beneath it; he cites a former employer whose weekly active users counted anyone who opened an email, which iOS auto-opens inflated.
• Four of seven North Star tests are deal breakers. The seven are customer value, fit with strategy, leading rather than lagging, movable by the team, understandable, measurable, and free of vanity; failing customer value, leading, movable or not-vanity disqualifies a metric regardless of the rest. Airbnb's nights booked is his passing example; the frameworks are drawn largely from Amplitude's North Star playbook, and the lab says it is partnering with Amplitude on this session.
• Input metrics must decompose the North Star, not rename it. The levers are breadth, depth, efficiency and frequency (how many customers, how much each does, how reliably it goes through, how often they return). Shane warns that "active buyers" under a "weekly active buyers" North Star is the same number in a different hat, so pick a countable event such as orders or nights booked that splits into separate levers.
• The demo audit refused DAU and graded completed orders weak. A North Star skill in the lab's course-only "AI analyst plus" repo, not yet pushed to main, grounds each verdict in a cited knowledge base via a "librarian" agent. It refused daily active users as a canonical anti-pattern (logins, not value received) and suggested a value-delivering action instead; weekly completed orders passed five of seven checks with no fatal failure but was flagged "issue-blind" to returns, defects and late delivery, with the fix "weekly completed orders with no issues", plus a generic strategy fit that more business context would sharpen.
• A strong headline number hid growth built on one lever. The drivers mode on the synthetic Novamart data for 2024 (served from an earlier cached run) found orders grew six times, with about 99% from breadth, flat frequency, average order value down about 9% and the membership program at 2–3%. Shane reads this as "renting new customers" and says the AI can grade and decompose a metric but the choice of what to measure and which lever to prioritize stays a human decision.

2026-06-16 · open the original ↗

• Packaging costs more than analysis because its steps run in sequence. The presenter says formatting a finished analysis into a document — charts, titles, placement, coherence between sections — often takes longer than the 30–60 minutes of thinking, and every analyst on a team pays that tax each time. The proposed fix is to analyze once and repackage per audience, with formats and templates decided in advance so Claude Code can publish in minutes; the presenter concedes that scaffolding takes time to build.
• Every publish prompt should name four things. They are the audience, the format, the length and depth, and what to include. The demo's map gives each format a different job: a Google Doc for a VP as a decision record (executive summary, findings, recommendation, the decision needed), a Notion page as a detailed team wiki carrying all charts plus SQL and methodology for review, and slides for a 5–10 minute slot with one message per slide, a headline that states the point and a chart as the proof.
• The hard part is choosing what to cut. The presenter says junior analysts put everything in, leaving 50–60-slide decks to cut to 15; the host adds a new graduate whose 200-chart deck came down to seven. The generated documents titled every chart and section with its insight rather than a description, which the presenter singles out as the point.
• The demo published one finished analysis three ways. Working from the synthetic Novamart dataset (13 tables, 6 million records, a local DuckDB), Claude Code summarized 2024 as revenue up 5.4x to 3.15 million with growth "leaky and getting more expensive", then wrote the VP Google Doc through the Google CLI, a multi-tab Notion page through an authenticated MCP (shown from a run made before the session), and a first-draft slide deck. The presenter says the analysis was already done, which is why it was fast, would still iterate on the Notion layout and on the slides' thin text, and skipped Slack and email.
• Audience feedback should flow back into the next version. Answering an attendee, the presenter describes feeding an Otter transcript via MCP to Claude Code after each meeting to plan a V2, and having it read and answer leadership comments in Notion or Coda. The host adds that he analyzes Claude Code workshop chats and transcripts the same way and built profiles of his main stakeholders from their documents and Slack messages to tailor what he presents.

2026-06-16 · open the original ↗

• A/B tests go wrong because the reader wants a result. Shane says bad tests come less from weak tooling or rigor than from the person reading the result being the one who hoped for it. He names four habits: moving the threshold to where the result landed (a 2% minimum detectable effect quietly becoming 1.5% after a 1.6% lift), peeking and stopping while winning, explaining away broken guardrails, and promoting a secondary metric when the primary is flat. He calls this how brains work, not dishonesty.
• Let code decide, LLMs translate, and humans judge. Experimentation platforms compute the statistics correctly but, in Shane's account, act as calculators that leave the call to people at the moment their judgment is most compromised, and LLMs pattern-match toward hoped-for results the same way. His system targets design and analysis and deliberately leaves instrumentation alone as too high-risk. Deterministic Python issues the verdict, LLMs draft and explain, and judging agents get less context on purpose — "blind judges" — because the failure being defended against is motivated reasoning, not ignorance.
• Metric definitions live in files the agents cannot invent around. The open-source Agent XP repo has five agents (understander, designer, critic, SQL, narrator), design and analyze skills, and a whitelist of stats functions the critic checks were used. Project-level YAML files define semantic models and the demo's 14 metrics, each carrying its type and defaults; revenue per user, for example, brings a −1% guardrail to every experiment that uses it. Shane says deciding those definitions remains a human conversation with the team.
• Analysis runs as ordered gates before any lift is shown. The steps check that the split matches the brief (a 57/43 split on a planned 50/50 halts before lift is computed), that no pre-set guardrail broke, and that the sample is large enough. Only then does it read the primary lift's direction and magnitude, then whether the effect held over the window (for example a 25–30% drop in the last third flags novelty), before a ship, don't-ship or learn verdict.
• The live demo produced a sealed design but never reached analysis. The run took 13 minutes. For "move the buy now button above the fold", the designer picked conversion rate as primary from the catalog because it matched the word used, proposed buy-now click rate as secondary, and ran power analysis on the primary and on the guardrails, finding a 5% relative MDE feasible. A critic agent then reviewed it and the brief was sealed. The analyze stage was not run, and Shane says the tool does not do randomization or assignment and is early, actively changing work.

2026-06-16 · open the original ↗

• Pushing back on the question is the highest-value analyst skill. The presenter says students in the lab's first five-week cohort named reframing the question as the single most valuable thing they learned, ahead of the repo of 60 skills and agents. A vague ask like "analyze user engagement" yields generic observations no one can act on; "which segment drove the engagement drop last month so we can decide where to focus" names a segment, a quantified driver and the decision.
• Every question must pass three checks before analysis starts. The presenter's 30-second gut check asks whether the question is tied to a decision (if none can be named, stop), grounded in data that actually exists, and specific enough to know when you are done. The worked ladder runs from "why are signups down?" to "did the pricing page redesign cause the 12% signup drop in the 2 weeks post launch?", with the decision being whether to roll it back.
• Interview prompts are vague on purpose and reward clarifying first. Asked how to measure a new search feature, candidates who list DAU, retention and engagement fall into "the trap", the presenter says, while those who get offers ask success for whom (power users or everyone) and against what baseline. Only then do they set a primary metric, a guardrail and a 30-day bar for continued investment, confirming the decision with the interviewer as they go.
• State in advance what result would prove you wrong. For a hypothesis that mid-market churn comes from failed onboarding rather than pricing, the presenter lists confirming evidence (churn clustered in accounts that never finished setup) and rejecting evidence (churn spread evenly, pricing dominating exit interviews). Requests are then scored on impact and feasibility: do high-impact, data-ready ones first, plan ones that lack instrumentation, use quick wins as buffer, and decline the low-low "graveyard of side quests" with written reasons.
• The demos turned a churn Slack into a bounded question. A "question quality coach" prompt in the Claude web app (work, interview and rank modes) failed "can you look into our churn numbers?" on two of three checks and coached it toward a segment-ranking decision. A question-framing gate in the lab's Claude Code repo scored it 0/3 and, filling gaps with assumptions for the demo, produced: has monthly voluntary logo churn risen above a 3% target over two quarters, in which plan-tier and tenure segment, so leadership can decide on funding a retention intervention — with "below 3% and no single segment" as the stop condition. The host adds that vagueness often hides in windows, filters and thresholds.

2026-06-16 · open the original ↗

• Shrinking attention spans are pushing retail toward moonshot bets. The question posed is whether hypergrowth "moonshot" investing, aimed at quick exponential wealth, has replaced Warren Buffett's moat style of capital preservation and steady growth. One host blames social media for cutting attention from 5–15 minutes to 20–30 seconds, and calls Gen Z more of a gambling, quick-fix crowd. That host credits the GameStop, Koss and AMC meme-stock episode with showing that retail can move markets.
• SpaceX is trading on price, not value, and will break. The same host says retail piles into a stock or two each day, and SpaceX buyers keep calling it cheap at 135, 150, 200 and 220 regardless of company quality. The forecast: moonshot investing is the norm "for the foreseeable future", until the market breaks and something like moat-style investing returns.
• The hosts split on whether moonshot investing is permanent. A second host calls the question silly, says moonshot investing "will always be the norm", and reminds the first of past advice to young traders: risk all of a $5,000 or $10,000 account instead of setting money aside in bonds. The clip ends with someone insisting moat style "is not going away". The captions do not show who said it.

2026-06-16 · open the original ↗

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

2026-06-16 · open the original ↗

• Crosstalk Mobile runs on T-Mobile as an authorized MVNO. The company, launched about two months before the video, has a direct agreement with T-Mobile, its only carrier for now, with multi-carrier SIMs planned "in the near future". The host contrasts this with gray-market resellers buying wholesale SIMs without a carrier agreement, which he says get shut down when caught.
• The unlimited SIMs really have no caps or overages. Plans start at $119 a month, and the host says no amount of data changes that. Cheaper limited tiers are planned as a priority, for backup links, IoT and RedCap devices such as Netgear's $99 5G backup device and the GL.iNet Comet 5G, with no date given.
• The SIMs suit 5G routers, not phones or tablets. The host says they work well in Netgear 5G and Peplink gear and he has not yet found a router they fail in, but phones and tablets are hit and miss and discouraged. He cites a reseller who got a Peplink router chosen, preconfigured on InControl 2 and drop-shipped with the SIM installed, all by text in a few hours.
• The SIMs get real public IPv4, not CGNAT. The host argues CGNAT makes 5G failover incomplete, since inbound services such as a VPN server for remote staff stop working when the office fails over. Crosstalk SIMs get real IPv4 addresses, and a static IP is $5 a month extra; he personally tested port forwarding through the UniFi 5G Max and the Dream Router 5G Max.
• There is no contract lock-in. Plans are month-to-month, or year-to-year if bought annually, and can be cancelled at any time. The host presents engineers answering customer calls and texts directly as the service's most underrated feature, contrasting it with Verizon Business or AT&T support.

2026-06-16 · open the original ↗

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

2026-06-16 · open the original ↗

• AI erased his technical moat but raised engineers' value. In this role-reversal episode, hosted by founder Kirill Eremenko, Krohn says the machine-learning skill set he built from his Oxford neuroscience PhD onward stopped being a moat a few months before recording, when Claude started doing that work better than him. His argument is that engineers who stop typing code by hand and use tools like Claude Code can deliver far more, so hiring them is worth more, not less.
• AGI needs breadth into the physical world, not consciousness. Krohn points to Google DeepMind's five-levels paper (his episode 748), which grades AGI by the share of humans a system outperforms, and argues that AI now matches or beats many people on computer-based work like coding and writing but is years from acting in the real world, hence world-model startups' huge seed rounds. He sees no reason consciousness is required, and says biological brains are useful inspiration but silicon has different limits and advantages of scale and speed.
• An AI bubble bursting would still leave useful infrastructure. Citing his episode 974, he argues that, as with 19th-century railways and dot-com-era fibre, investors might lose their shirts but data centers, model advances and talent would benefit everyone; he personally holds diversified investments rather than individual AI stocks and would bet this is probably not a bubble. He invokes Jevons paradox to argue more efficient architectures will raise total compute demand, and says the hardware won't become obsolete because any successor to the Transformer will still need parallel matrix multiplication.
• Most AI projects fail because the enterprise data isn't there. Krohn says the number-one failure he sees is a CEO proving a capability in ChatGPT or Copilot when the company lacks the underlying data at scale (Eremenko adds: or the pipelines to use it). He prioritises projects with RICE scoring (reach, impact and confidence over effort), estimated for production rather than prototype, and argues the people doing the workflow plus the technical team score it better than executives; he cites MIT NANDA's 95% failure figure while noting the exact number has been debunked.
• The money is in bridging AI to real workflows, not pure SaaS. He says his consulting firm Y Carrot, about eight people and founded a little over a year earlier, typically shows a prototype in week two and has cut multi-week processes to minutes without job losses so far, because freed time goes to higher-value work. He argues the 80%-margin pure-SaaS model breaks down when Claude Code can replicate it, leaving durable opportunity in forward-deployed or consultative models like Palantir's, now used by Anthropic and OpenAI.

2026-06-16 · open the original ↗

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

2026-06-15 · open the original ↗

• Gold and silver rose on the peace announcement instead of falling. That's the opposite of the usual safe-haven reaction to easing tensions, which the video treats as a reason to question whether the deal is substantive or timed for effect.
• The timing lines up with Kevin Warsh's first day as Fed chair. Caught between pressure to cut rates to refinance $39 trillion in debt and pressure to hold given 4.2% inflation, the peace news dropped oil more than $4 a barrel, easing inflation optics right before his first meeting.
• The deal's substance is contested by a key party. Israel says it was excluded from the negotiations and won't withdraw from Lebanon, a stated condition of the deal, and the Strait of Hormuz reportedly remains blocked pending mine-sweeping.

2026-06-15 · open the original ↗

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

2026-06-15 · open the original ↗

• One host calls a short-gold leg the trade's real "widow maker." He pairs it with a long silver position (two gold micro futures short to one silver micro long); the combined trade had moved $4,000 since Friday, with the long side sitting near the top of a three-to-four-month range around $65.
• A liquid, hyped IPO like SpaceX will price with wildly inconsistent volatility. One host expects different market makers to quote implied volatility anywhere from 40 to 90 on day one, and predicts most of the street will wait for Citadel to post the first markets before committing.
• New IPOs should not be traded until options exist and settle. The advice is to wait a few days for options (hedging with short stock if needed), then give volatility time to find its level — the speaker recalls once profiting and once losing badly by pricing volatility too early.

2026-06-15 · open the original ↗

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

2026-06-15 · open the original ↗

• The author ties the rushed Iran peace framework directly to Fed Chair Kevin Warsh's first FOMC meeting. Rising energy prices from the Iran war have pressured inflation and borrowing costs ahead of the midterms, and the piece argues Trump is willing to make concessions in Iran specifically to help Warsh start with lower rates.
• The deal favors Iran, whose regime remains intact and emboldened. The Strait of Hormuz operated without incident before the war began, so ending the war doesn't recover lost ground, it just removes a self-created risk — while Israel, the author argues, now has less leverage than before fighting started, given the economic toll on the West.
• The author predicts Iran will stall negotiations to maximize leverage into the midterms, forcing Warsh into a defensive, quiet posture. Disrupting Hormuz costs Iran far less than it costs the U.S. and Israel to defend against, and continued disruption risk keeps Treasury yields elevated — pressure the author expects Warsh to manage by minimizing public communication rather than through policy moves.

2026-06-15 · open the original ↗

• He opens on his own distribution risk, which is the title's point: suppression now works by disappearance, not deletion. Views have fallen from 6,000–8,000 per video six months ago to 300–500, with follower counts frozen near 17,000 on TikTok and 11,200 on X for two years. Nothing is removed; it simply stops being shown. His response is to push the audience toward Substack for a semblance of control he does not have on a platform someone else owns.
• The addressable-market arithmetic is his actual investment point about US AI. The administration froze two Anthropic models to foreign nationals — 90 minutes to comply, no detailed rationale given — which means labs have been told they may serve 330 million people out of 8 billion. Set against DeepSeek at roughly 19.66 cents per million tokens versus Anthropic at $15, and OpenAI cutting prices 40–50% while unable to service existing debt, he concludes there is no revenue model that makes sense. He also relays the widely-agreed technical caveat: jailbreaks cannot be fully fixed in these models, and other US labs are presumably equally exposed.
• His post-war macro read hinges on one pair, with an explicit tail risk attached. Warsh's arrival reopens two parked Fed cuts; Europe benefits most from energy normalising; the BOJ is at 1% and expected near 2% by end-2027, narrowing the differential the carry trade depends on. He gives USDCNY at 6.8088 as the level to watch, with above 7 signalling trouble. His highest-conviction expression is dollar-yen, because both legs push the same way — but he names the paradox: a gentle yen appreciation delivers the clean yen-up, dollar-down, risk-off picture, while a violent disorderly unwind lets the dollar catch a safe-haven bid against everything except the yen, producing both rising together and breaking the dollar-down thesis temporarily. Gold he flags as the outlier, supported by lower real rates even in a risk-on tape.

2026-06-15 · open the original ↗

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

2026-06-15 · open the original ↗

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

2026-06-15 · open the original ↗

• Groww's first idea, a robo-advisor, failed before pivoting to full transparency. Lalit Keshre, Groww's co-founder and CEO, says the 2016 robo-advisor model stalled despite India's low investor penetration; customers kept asking "why this product, why not that," which led the team to open the full catalog with transparent pricing instead, Flipkart-style.
• He chose to operate only in regulated territory as a risk strategy. Rather than exploit unregulated or gray-zone areas — which he estimates cover roughly 40% of the space — Groww obtained licenses upfront, which he says removed decision variables and made execution simpler even though it was slower.
• Groww ran four years at zero revenue before customer demand unlocked monetization. Early commission-based mutual-fund revenue was abandoned after vocal power users demanded zero-commission "direct" funds; the company bet that near-zero acquisition cost, high retention and heavy money flow would eventually produce a business model, and stock trading later became the monetization lever.
• Four co-founders stayed aligned for a decade via one values document. Keshre says strategy can change "like pencil," but core values like "customer first," written down at founding, should not; he also stresses genuinely enjoying each other's company as a practical requirement, not a nicety.

2026-06-15 · open the original ↗

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

2026-06-15 · open the original ↗

• One prompt built and deployed a conference sessions site. Padua, who has been at Microsoft for four years, gave the new GitHub Copilot app a JSON file of Build session data, and it created a repo, built the app and set up deployment to Azure. He puts the time at about five minutes at one point and "10 minutes tops" at another. The app ran a team of about 15 agents under a "director". His only manual step was copying the Azure static web app token into the repo's GitHub Actions secrets.
• An internal agent tool matched conference sessions to a partner. Padua says Scout, an internal Microsoft tool built on the open-source OpenClaw, ran parallel agents that read his site and researched the partner Solo Network. When Playwright failed, they fell back to simulated mouse and keyboard input. They picked 16 best-fit sessions, built a table with links and levels, and drafted an email offering a 30-minute call. He has done this for three partners, in Portuguese where he remembered to ask.
• Agents act as you, so limit what they may do. Padua advises running OpenClaw in a virtual machine you can destroy and letting it only draft, never send or auto-reply to email, citing stories of agents deleting emails or a whole operating system. He had Copilot write a PowerShell deploy script for him to review rather than letting it deploy through the Azure MCP server tied to his credentials, to avoid unwanted resources and costs.
• Copilot digs answers out of dense internal documentation. Padua uses Microsoft 365 Copilot to summarize meetings he missed and to draft LinkedIn posts. Searching the documentation this way, he found why a partner lacked the Anthropic model quota it expected in Foundry: he says those models work only under an EA contract, and the customer was on CSP. He also tries competitors' tools, such as OpenAI's DALL-E, Anthropic and Perplexity, because customers ask his opinion of them.
• AI amplifies technical knowledge rather than replacing it. Padua argues that results depend on the user knowing what they want to do, and that AI then delivers "the best of the people can do with a prompt". Giard adds that the agents narrate their failures and change approach, as when Playwright failed, which he says is what a person would do, only faster.

2026-06-13 · open the original ↗

• No spoken content. The 50-second short has no captions, and the whisper transcript of its audio is empty; its subject is in its title alone: "Silver's Weekly Snapshot for June 13, 2026".

2026-06-13 · open the original ↗

• The Overton window frames how radical ideas become mainstream policy. He credits the concept to Joseph Overton, a VP at a Michigan free-market think tank whose ideas were only published after his death by a colleague; it describes the sliding range of what's politically acceptable, from "unthinkable" to enacted law.
• He cites abolition and public education as past shifts, and today's immigration and identity debates as ongoing ones. Slavery's abolition and universal public schooling were once branded radical before becoming policy; he points to democratic socialists winning office and the push for mass deportations and ending birthright citizenship as evidence the window keeps moving on both the left and right.
• He argues most Overton shifts are reversible unless locked into the Constitution. He notes women's suffrage and the freeing of slaves stuck because they were constitutionally codified, while other shifts, including growing distrust of the FBI, CIA and NSA, could still swing back toward the fringe.

2026-06-13 · open the original ↗

• SpaceX's record $75 billion IPO at a $1.8 trillion valuation overshadowed the week's macro news, instantly making it the seventh most valuable US company. The offering, nearly three times the previous largest IPO, made Elon Musk a trillionaire and minted roughly 4,400 millionaires among current and former employees, despite the company generating no profits and carrying significant key-person risk.
• Hot headline inflation prints contrasted with cooler core readings, yet the ECB hiked rates anyway — its first increase since 2023. US headline PPI came in hot at 6.5% for May while core cooled to 4.9%, and CPI posted its first 4% handle (4.2%) in three years even as core eased to 2.8%; he notes the pass-through to core has consistently undershot forecasts.
• The week ahead brings a wave of central bank decisions, led by an unusual Fed meeting where new Chair Warsh runs his first session with predecessor Powell still on the committee. Markets have abandoned hopes for 2026 rate cuts and are now pricing a possible hike by year-end, with the Bank of England, Swiss National Bank, Bank of Japan and Reserve Bank of Australia also deciding this week.

2026-06-12 · open the original ↗

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

2026-06-12 · open the original ↗

• WF6 is a tiny line item that gates output worth orders of magnitude more, and it has no drop-in alternative. It is the gas that deposits metallic tungsten to form a chip's contacts, vias, and the word lines stacked through 3D NAND — nicknamed "chip blood." It requires 5N–6N purity, qualifying a new source takes 12–18 months, and there is no substitute chemistry at scale. It is consumed on essentially every advanced wafer and is heaviest in 3D NAND and HBM, with each added memory layer meaning more tungsten per wafer — precisely as roughly 70% of 2026 memory output routes to data centres.
• The chokepoint sits upstream of the gas makers, so switching supplier does not escape it. China holds about 80% of global tungsten and supplied close to 100% of Japan's powder, and powder is 60–70% of WF6's production cost. China added tungsten to export controls in February 2025 and extended them to high-purity powder under the January 2026 dual-use catalogue; shipments to Japan fell to roughly zero from February 2026, with tungsten powder up ~325% year-on-year and the APT benchmark up 557% since February 2025. Every listed producer — Kanto Denka, Central Glass, Korea's SK Specialty and Foosung, even China's own PERIC — depends on the same upstream powder.
• What makes this different from an ordinary shortage is that it destroys the midstream rather than rationing it. Because most of WF6's cost is the now-unobtainable powder, the Japanese gas makers exit rather than ration — converting a shock into a permanent supply-base migration. Kanto Denka and Central Glass burned roughly five months of inventory and halt production around 1 July. Exposure downstream is uneven: Samsung is most exposed on heaviest reliance, SK Hynix is cushioned by sister company SK Specialty, DB HiTek and Magnachip have been notified of 70–90% price hikes, and TSMC is racing the 12–18 month qualification of non-Japanese alternatives.

2026-06-12 · open the original ↗

• The familiar AI supply chain describes only the compute, and the two missing pieces are both Chinese-gated. ASML to TSMC to Nvidia to the data centre delivers silicon, but an AI supercomputer is thousands of GPUs acting as one — which requires an interconnect to wire them together and memory to feed them. Neither comes from that chain. Indium phosphide supplies the interconnect and tungsten hexafluoride the memory, and both trace back to materials China controls.
• Neither can be substituted for reasons of physics rather than commerce. Silicon is an indirect-bandgap semiconductor and physically cannot emit laser light efficiently; InP is direct-bandgap and is the only material making high-speed optical lasers at scale — even silicon photonics still needs an InP laser bonded onto it. WF6 requires 5N–6N purity on a 12–18 month requalification cycle. And demand per GPU is rising structurally: as bandwidth climbs through NVLink generations from 800G to 1.6T, copper fails at ever shorter distances, pushing optics closer to the chip, while cluster topologies add optical links faster than they add GPUs.
• The decisive finding is convergence — two unrelated chokepoints terminating at the same endpoint. Different minerals, different midstream firms, different licensing gates, but China can squeeze the buildout from two independent directions at once. The severity grid scores indium/InP at ~70% dependency, +250% price shock and a substitution lag running to ~2029, against tungsten/WF6 at ~100% dependency, +325–557% and 12–18 months — with no soft cell anywhere in the matrix and tungsten critical on both dependency and price. On concentration, AXT (US-listed but with fabs in China) and Sumitomo hold roughly 80% of InP substrate with JX Advanced Metals near 10%, but AXT is gated at China's border and Sumitomo consumes much of its output internally, so the merchant market is tighter than the shares imply — Lumentum is already sold out through 2028 with wafers up 250%.

2026-06-12 · open the original ↗

• The constraint binds at the interconnect layer, not the compute layer — which is where the famous chip chain stops describing the problem. ASML prints, TSMC fabricates, Nvidia designs, data centres cluster. But a large training run is thousands to tens of thousands of GPUs acting as one machine, and you can have every GPU the silicon chain can produce: if you cannot optically wire them together, you cannot form the machine. The physics is unavoidable — silicon is indirect-bandgap so light emission is wildly inefficient, InP is direct-bandgap, and the one partial alternative (GaAs VCSELs over multimode fibre) covers only short-reach links. Even silicon photonics needs an InP laser bonded onto it.
• Concentration stacks three deep, and the export permit is what converts latent leverage into disruption. Refining is ~70% China — latent leverage. Substrate is the physical bottleneck, with roughly three firms holding ~90%: AXT is US-listed but fabs inside China and gated at its border, Sumitomo holds ~30% but consumes much of it internally, JX Advanced Metals ~10% — so the merchant market is far tighter than headline shares suggest. The active instrument is China's per-order export-permit regime, added 4 February 2025, requiring a government permit for every overseas shipment. Permits have run 60+ business days and AXT only began receiving them again in March 2026. It is metering rather than embargo — but metering is enough.
• The cascade runs in order, and the gap is structural for the rest of the decade. Permits slow, AXT shipments are gated and backlog builds past $100M in Q1 2026 — so revenue recognition becomes a function of Beijing's regulatory clock rather than customer demand. Transceiver and laser makers then hit substrate scarcity, with Lumentum sold out through 2028 and average six-inch InP wafers up ~250% to about $5,000 since February 2025. The interconnect layer caps GPU clustering, throttling the AI-specific part of the buildout with a non-compute input. Then forward-buying and stockpiling ratchet prices that persist even if permit volume later loosens — and the first large Western refined-indium addition, Korea Zinc's Tennessee complex, is not expected before ~2029. The asymmetry is the point: a single-digit-kiloton market sits beneath a multi-hundred-billion-dollar buildout.

2026-06-12 · open the original ↗

• He opens with an unusually direct admission about the state of American supply-chain knowledge. Having imposed tariffs on the premise that the US held all the cards and the world would rearrange around domestic manufacturing, his verdict is that "we were basically deaf, dumb and blind men when it came to the supply chain." The evidence he points to is that US officials have visited China repeatedly to try to resolve the InP export controls — and that Japan asked Washington to intercede on its behalf, which he reads as a statement about who actually holds leverage.
• He explains the physics plainly and lands on the same conclusion the paper reaches. Direct versus indirect bandgap: in a direct-bandgap material an excited electron drops across the gap and releases the energy cleanly as light, while in silicon that same drop needs a lattice vibration to conserve momentum, making emission wildly inefficient. As bandwidth per GPU climbs from 800G toward 1.6T, the distance at which copper works keeps shrinking, pushing optics from the building edge into the rack and onto the package itself. Every one of those links needs an InP laser, so InP gates the wiring-together.
• His comparison of the two chokepoints names the trade-off precisely. Tungsten and WF6 are worse on price — up somewhere between 325% and 557% — but the substitution path runs 12 to 18 months. Indium and InP are up around 250% but are not meaningfully addressed until roughly 2029. "One's worse on price, the other one's worse on delivery." His conclusion is a policy one: the country needs an actual industrial policy that decides what it wants to build and then traces where every input comes from, and investors need to get far better at mapping chokepoints — who makes what, and how they make it.

2026-06-12 · open the original ↗

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

2026-06-12 · open the original ↗

• Japan's weakening demand for its own government bonds signals the end of its role as the world's price-insensitive buyer of foreign debt. A June 10th 30-year JGB auction drew the weakest demand since June 2025, with the bid-to-cover ratio falling to 2.94 from a 12-month average of 3.4, as domestic yields near 2.5-2.7% finally compete with returns abroad.
• As Japanese yields rise, Japan needs to buy less foreign debt, feeding what the author calls the "great repricing" of global bond yields. Japan holds about $1.2 trillion in U.S. Treasuries, and the life insurers and pension funds that have consistently bought foreign debt are starting to recalibrate toward higher yields available at home.
• A Bank of Japan rate hike, needed to fight inflation, threatens to unwind the global carry trade that lets Western investors borrow cheap yen to buy assets like U.S. equities. With Japan's debt-to-GDP near 230% and BOJ JGB holdings already down from 91% of GDP in mid-2024 to about 80% in January 2026, the piece frames the taper decision, not the rate move, as the key thing to watch.

2026-06-12 · open the original ↗

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

2026-06-12 · open the original ↗

• Protein language models are starting to show the same LLM-style scaling laws. A presenter says a new protein-model family from Biohub, ESM Cambrian, keeps improving log-linearly with more training compute and sequence data — unlike its predecessor, which plateaued — largely because the team scaled training data from 50 million to 2.8 billion sequences pulled from environmental DNA sequencing.
• Naive AI self-play collapses into generating deliberately convoluted, useless problems. A Stanford PhD student found that rewarding a "conjecturer" model for stumping a "solver" model just produces artificially messy math problems; his fix grounds new problems in ones the solver already failed on and adds a third model to judge relevance, letting a 7-billion-parameter model match a 670-billion-parameter baseline's solve rate after eight times more self-play compute.
• Lean formal verification marks a shift from "vibe coding" to "verifiable coding." A Caltech researcher says frontier labs have used the Lean proof language to solve IMO gold-medal problems and decades-old Erdős conjectures, and describes his own project formalizing neural-network operations — including proving flash attention mathematically equivalent to standard attention — as evidence AI-written code can eventually carry mathematical guarantees, not just pass tests.
• Treating agentic coding like a real-time strategy game tripled shipped pull requests. A startup CEO known for aggressive token spending says running many parallel Claude Code "worker" agents under one orchestrator, instructing them to push forward without stopping for feedback, and feeding a shared knowledge base after every task grew pull requests per engineer per month 3.5x, then another 60% once the whole team adopted the workflow.

2026-06-12 · open the original ↗

• Beat the speed of adoption, not the speed of research. Responding to a former Dutch professor of human-AI collaboration about teaching safe AI use, Kirill Eremenko, the show's founder, argues research moves like a bullet train while adoption in universities, enterprises and government moves like a highway traffic jam, slowed by legacy systems and change management. He says people anxious about AI only need to stay ahead of adoption, and that startups should avoid problems models will soon solve, citing hallucinations he estimates are down about 10-fold in two years and AI-text detectors as fighting the tide.
• The host would fund agents that shield people from manipulative feeds. Asked by Jepson Taylor what they would fix with $10 billion in two months, the host proposes an agent between vulnerable people and the feeds, scams and attention-seeking products they consume, steering them toward what makes them happy; Eremenko proposes AI-run charities that cut admin overhead from what he thinks is about 40% to about 5%. Taylor's own answer is closing what an engineer told him is the gap between the top 10% and bottom 90% of AI developers.
• Build what differentiates you and is slow to build; buy the rest. Answering a corporate listener on the "SaaS apocalypse", the host relays a 2x2 from episode 932's guest from Unframe: avoid projects that are costly and non-differentiating, build in-house what is differentiating but slow, buy what is quick and non-differentiating, and partner (he names Unframe and his own firm) where it is both quick and differentiating.
• The show's big growth lever was video, adopted far too late. Eremenko, who hosted from September 2016 to episode 431, says his one piece of hindsight is that he should have put episodes on YouTube from the start; the host says the show went video-first only 10 months before recording and grew from 20,000 to over 250,000 YouTube subscribers since. The host's own lesson, from Eremenko's feedback, was to be conversational and stop taking notes on air.
• In-person networking beats clicking "apply". The host argues weak new-grad hiring reflects degrees losing their "golden ticket" status and post-pandemic overhiring as much as AI, citing episode 994. His tips for entrants are a GitHub of projects actually done yourself and meeting people in person, because AI-tailored résumés let applicants send thousands of applications and flood the traditional pipeline; he also cites episode 977's NYU machine-learning course, taught by Kyunghyun Cho, that required students to use Gemini.

2026-06-12 · open the original ↗

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

2026-06-11 · open the original ↗

• Archera sells insurance that lets customers break cloud commitments. Khanna, co-founder and CEO, describes "insured cloud commitments": commit for 3 years and, if the last 24 months go unused, Archera pays back the difference; he says no one else offers it, and customers can cancel the insurance once confident. His analogy is buying the right to break a 3-year apartment lease after the first year.
• Commitments exist because clouds finance capex against guaranteed revenue. Like a builder whose lender underwrites leases rather than Airbnb income, providers borrow to build data centres (CoreWeave is his extreme example), so discounts can exceed 80% for moving from per-second use to a 3-year commitment. He says that in 2026 a commitment is often required just to get GPUs; Hanselman calls on-demand "the convenience store price of compute", and Khanna calls it the promotional price, where you prototype.
• Commitments became a hammer applied to every cloud problem. Reserved instances made sense in the early 2010s, he says, but from the 2017 SageMaker launch, which he worked on at AWS, they were used to ration scarce Nvidia capacity, then to deter multi-cloud and push modernisation, leaving "a spaghetti mess". At $80–90 million a year with one provider, he says, bills contain errors and cannot be treated as a source of truth.
• Underwriting relies on spend history, peer priors and diversification. Archera wants at least a year of real usage data, draws priors from over a petabyte of history across thousands of customers (offering less on Azure VMware Service, which tends to migrate to AKS within a year), and spreads its book across instance types, GPUs and customers. Only about 40% of its customers buy insurance; the rest use free billing and commitment tools.
• Small teams qualify, but savings matter most at larger spend. Khanna puts the meaningful threshold at $100,000–200,000 a year of cloud spend, with savings that could fund a headcount from $500,000-plus, though two-person startups on credits use it and a quote takes 5 minutes. He separates his rate optimisation, a finance question, from usage optimisation, an engineering one.

2026-06-11 · open the original ↗

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

2026-06-11 · open the original ↗

• Saudi Arabia, the UAE, Qatar and Kuwait are now using the same dark-fleet tricks as Russia, Iran and Venezuela to move oil. He describes tankers switching off transponders to slip through the Strait of Hormuz undetected by Iran, or turning them off just outside the Strait to do ship-to-ship transfers before the receiving vessel sells the cargo.
• He credits this evasion, alongside shifting trade flows, with helping keep oil under $200 a barrel. He pairs it with China and India cutting oil imports while the US, Canada, Venezuela, Mexico and Norway raise exports, though he calls the dark transfers slower and costlier than normal shipping.
• He predicts an "economic rollercoaster" within four to five months if the Strait of Hormuz stays disrupted. He argues the longer these workarounds continue, the less leverage Iran retains over global oil markets through control of the Strait.

2026-06-11 · open the original ↗

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

2026-06-11 · open the original ↗

• AI firms are floating token or energy taxes to preempt labor-displacement backlash. The hosts characterize this as AI companies signaling they'll need far fewer workers while offering to accept a token-usage or energy-usage tax as a concession.
• One host doubts AI will displace as much labor as feared, calling the tax question moot. The other pushes to table it anyway, arguing taxpayers are already subsidizing AI companies' heavy resource use since those firms pay little tax while building trillion-dollar market caps.
• They propose taxing firms that cut jobs via AI to offset lost payroll-tax revenue. The argument: if a company replaces a 100-person support desk with AI, a modest tax on that displacement would offset states' biggest revenue source, correcting a system where AI firms get all the upside and consumers/taxpayers get none.

2026-06-11 · open the original ↗

• Meesho's first product failed within three months because they never interviewed consumers. Vidit Aatrey, Meesho's co-founder and CEO, says their 2015 hyperlocal fashion-listing app collapsed once customers said it was "the worst of both worlds" — the limited selection of a local shop without the touch-and-feel of a mall — a mistake he calls "the most notorious mistake you can make."
• The real business emerged from noticing shops secretly ran WhatsApp storefronts. Sitting silently inside small shops rather than interviewing owners, the founders found sellers already used WhatsApp to message customers about new inventory, which led to a WhatsApp-commerce tool, and later to the discovery that the platform's true power users were online-only drop-shippers, mostly homemakers, needing supplier access rather than tooling.
• Meesho abandoned its profitable WhatsApp network in 2021 to launch an app. Aatrey says the shift from expensive mobile data to near-zero-cost data during COVID erased WhatsApp's core advantage, so the company knowingly risked alienating ten million reseller-run WhatsApp groups; the new app hit #1 on India's Android Play Store the day after its July 2021 launch and has stayed there every day since, growing to 100 million monthly users within five months.
• Meesho is building a voice-only agent for users overwhelmed by apps. Called Wani, it aims to remove reading, typing and clicking entirely — consumers speak preferences, hear responses, and give an address and OTP by voice — because Aatrey says rural users often don't understand what "add to cart" or a star rating means and fear that any button might deduct money.