Souk's Notes - Unofficial Summaries

2026-08-25 · open the original ↗

• Legora's founders bet on models improving rather than fine-tuning them for law. CEO Max Junestrand says the "tribal knowledge" circulating in 2022–2023 was that you had to fine-tune, citing Bloomberg's multi-million-dollar law-specific model; Legora instead built around the assumption that GPT-class models would keep improving, launching with a "ChatGPT, but compliant in Europe" pitch when OpenAI's product lacked EU data hosting.
• A six-month sales freeze, not a launch, turned the growth curve. After Benchmark invested $9.51 million and Redpoint led a Series A three weeks later — leaving a 10-person team sitting on $35 million — the founders halted selling to rebuild the platform around an October 2024 "product manifesto," reasoning that a lawyer who sees the system fail or lag rarely gives a second chance.
• Legora hires for growth trajectory, not resume pedigree. Junestrand says the company deliberately moved away from candidates with impressive starting points (what he calls a "Y-intercept" hire) toward people who "wanted to work insanely hard," pointing to its top seller — a 23-year-old with no sales background who has sold over $10 million of the product.
• Legora's next product frontier shifts agents from reactive to proactive. Rather than waiting on a prompt, agents are now wired to act on incoming triggers — routing or executing a contract the sales team receives, or organizing an entire data room into a due-diligence report unprompted — which Junestrand frames as the mechanism for one lawyer to produce the output of a ten-lawyer team.

2026-08-25 · open the original ↗

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

2026-08-25 · open the original ↗

• Data teams were building production systems without production tooling. Handy says that at RJMetrics, a prior-generation BI company where he ran marketing, a study of roughly 100 companies in 2014–2015 found early adopters moving to Redshift-based stacks yet hitting the same old problems. His conclusion was that always-on cloud dashboards made data work a production software system, which became the case for analytics engineering and dbt.
• dbt chose SQL and progressive complexity so analysts could own transformation. Handy picked SQL over Spark in 2016 because Spark was hard to set up and write for people from an analyst background, whom he values for spanning quantitative skill and business-domain knowledge. In dbt, transformation is a series of modest SELECT statements (ideally under about 100–120 lines each) forming a DAG, which dbt run persists as modelled tables in the database for BI tools or analytics agents to query.
• A semantic layer stops agents re-deriving metrics, expensively or wrongly. Handy frames it as an organisational problem: at a hypothetical Siemens-scale firm with 2,000 data engineers serving 300,000 employees, metric definitions are otherwise settled by convening meetings, and agents without stored definitions either burn many tokens or reach a plausible answer that isn't how the organisation measures. He places ontologies above it in a four-layer meaning hierarchy but thinks they suit only stable, verifiable, high-value domains such as genetic research or defence.
• Fusion brings type safety to SQL across database dialects. Handy argues there is no single SQL, since a syntactically valid statement can return different answers on two engines, which breeds lock-in. Fusion understands the code deeply enough to rewrite it for another engine and prove identical output, flags errors in dbt's VS Code extension as you type, and he says it gives coding agents tight, cheap test loops.
• Skill files compress dbt migrations from a year to weeks. The host cites Handy's LinkedIn post that a 12-kilobyte skill file can encode hundreds of hours of human experience; Handy says dbt Labs packaged the knowledge from its courses and certifications, which take a human about two solid weeks, into skills for agents. He says migrations of up to 10,000–50,000 tables at a big bank, once a year and multiple million dollars of consulting, can now take 6 weeks and tens of thousands of dollars, though AI does not yet show up in dbt's steady growth over about six years.

2026-08-25 · open the original ↗

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

2026-08-24 · open the original ↗

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

2026-08-24 · open the original ↗

• Section 338 is a dormant 1930 statute revived as a workaround after courts struck down the administration's IEEPA tariffs. It lets the president impose up to 50% duties by proclamation with no investigation, no sunset, and no defined standard for "discrimination"; its first live use hit Canada on 22 August with 50% duties on ~$20bn of autos, alcohol, and dairy, overriding USMCA.
• This isn't tariff policy — it's weaponised interdependence, applied through a layered stack of chokepoints. The Foreign Direct Product Rule, export controls, tariff coercion, bloc gatekeeping, and dollar-clearing access each reach further past the border than the last; allies respond by climbing a predictable ladder from hollow compliance to hedging to designing US content out entirely to bloc-shopping.
• The strategy trades durable dominance for short-term wins, and China's rare-earth chokepoint shows the game runs both directions. China controls ~90% of rare-earth processing and has already restricted gallium, germanium, graphite, and magnets; each side's coercion validates the other's case for de-risking, with the fragmentation cost to allies estimated at $400bn-$1tn.

2026-08-24 · open the original ↗

• Nvidia's scale is framed as capitalism eclipsing democratic power. Moving $313 billion in equity value in a single day without anyone reacting, the hosts argue, shows corporate concentration dwarfing other institutions; one raises the possibility such firms eventually get broken up.
• An autonomous drone using an Nvidia chip reportedly killed in Ukraine. One host cites a news report of Russia deploying a fully autonomous, no-human-intervention drone built on Nvidia hardware, arguing the company doesn't start wars but is nonetheless bound up in them.
• One host calls Nvidia the single stock that could pop this bubble. He says he's more worried about downside than upside and won't sell upside calls, even as analyst price targets cluster around $340 (about 41% above trading levels) — a level he suggests analysts favor partly because there's no upside to publicly forecasting lower.

2026-08-24 · open the original ↗

• Lossdog pitches three new tools for reading options-chain signals. "Relative IV" compares a ticker's volatility to the S&P 500, "unified volatility" combines four volatility signals into one read, and "ticker trade ideas" turns a bullish, bearish or neutral view into a trade built from options-chain data — all described as product features found under volatility analysis.

2026-08-24 · open the original ↗

• Waiting for confirmation is called a guaranteed way to buy late. One host says traders who wait for confirmation before entering "are always going to be the biggest losers," arguing successful trading requires being early, contrarian, and willing to catch a falling knife or sell into strength rather than wait for proof a move is real.
• Nvidia's roughly 5.5% earnings-implied move is called too small to trade. With options pricing a $313 billion swing for Wednesday's report, one host says he wouldn't touch the trade at that price, arguing a fair market would be closer to 16-18% wide; the other counters that market pricing can't be "wrong" since anyone who disagrees is free to buy the supposedly mispriced premium.
• Bitcoin's rally to near $80,000 splits the hosts on whether it holds. After bottoming in the low 60s, one expects a retreat back toward 60,000 once crypto-friendly policy optimism fades, while the other calls it a real bottom and says he'd buy any 25-30% pullback, framing digital assets as a genuine asset class rather than a short-covering bounce.
• The Treasury's bond-buyback increase is dismissed as too small to matter. Doubling weekly buybacks from $2 billion to $4 billion failed to hold down yields, which hit new highs anyway; against a nearly $30 trillion bond market, one host calls the intervention "the most ridiculous thing I've ever heard" and says "the markets win every time."

2026-08-24 · open the original ↗

• A day of training already yields grammatical, question-aware nonsense. The builder (Felix Rieseberg; the clip is from his Hanselminutes episode) trained a small model for one day on the latest version. Asked "what is the sun", it answered "the sun is a moon at the bottom of the earth". He calls that "just dumb", but notes it understood it was being asked a question it should answer.

2026-08-24 · open the original ↗

• Rosland Capital's Chapter 11 filing on July 2 left customers out roughly $85 million. That total breaks down as $49 million paid up front for coins and bars never delivered, nearly $12 million owed on metal customers returned under the buyback option, and over $23 million in unpaid marketing and TV debt, including a large sum owed to Fox News.
• No class action lawsuit is coming. The bankruptcy filing triggered an automatic legal stay barring any class suit while the case is active, and Rosland's customer agreements independently included arbitration clauses and class-action waivers that bar customers from joining one anyway.
• Affected customers must file Proof of Claim Form 410 with the bankruptcy court in California's Central District before the deadline, expected within the next few weeks. The speaker's closing warning: always take physical delivery of metal immediately rather than leaving it with a dealer, since a decades-old, well-regarded firm still went under owing millions.

2026-08-24 · open the original ↗

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

2026-08-24 · open the original ↗

• Compliance bought Canada nothing. Canada negotiated the deal directly with Trump, certified its goods under the required rules, and still got hit with 50% tariffs on about $20 billion of exports that were supposed to qualify — validating Carney's Davos warning that "compliance will buy safety... it won't."
• The administration routed around the compliance question entirely. After the Supreme Court ruled the emergency-powers statute didn't authorize tariffs, Trump's team reached for an obscure 1930s law that predates modern trade agreements, applying it regardless of whether Canadian firms had the required origin paperwork.
• Some U.S. grievances are real but don't justify the ask. Canadian boycotts cut U.S. spirits imports 81% ($582M) and auto imports 22% ($5.6B) over 18 months — yet the U.S. still demanded Canada limit its ability to strike trade deals with other countries, an asymmetric bind where "agreements bind you, not us."
• The lasting cost is a global risk premium, not the $20 billion itself. Other governments now have to price U.S. political risk the way EM traders price sovereign risk; middle powers are hedging toward the EU (Canada is SAFE's only non-European member) since China can't realistically absorb Canadian trade (73% of exports go to the U.S. vs. 4.4% to China).

2026-08-22 · open the original ↗

• Russia and China oppose an Iranian nuclear weapon because they fear the IRGC's apparent indifference to death, not its raw capability. He contrasts this with narcissistic dictators like Putin, Xi Jinping, and Kim Jong-un, for whom personal harm is the worst outcome, and says 2025's beatings by the US and Israel proved the IRGC would keep fighting even while hobbled.
• China's opposition is read as driven by economic interest: it wants a dollar-free trade network and stable Persian Gulf oil, both threatened by a nuclear-armed Iran near its neighbors, while Russia wants to remain the region's sole nuclear threat. Their public support for Iran's nuclear program is characterized as costless posturing against the US, since Israeli assassinations of Iranian nuclear scientists and the 2025 US bunker-busting strikes were expected to prevent real progress anyway.
• He recalls that Iran was close to a nuclear weapons program in the mid-1970s under US backing, with the CIA reportedly warning against teaching Iran uranium enrichment because a revolution was likely — a revolution that then occurred.

2026-08-22 · open the original ↗

• Resurfaced 2023 clips show then-Senator JD Vance comparing the dollar's reserve-currency role to a "resource curse." His argument: foreign central banks (China, Japan) hoarding dollars and Treasuries creates artificial demand that inflates the dollar's value, making US-made goods too expensive abroad and foreign imports cheap enough to drive American manufacturers out of business.
• The same clips show Vance directly challenging Fed Chair Jerome Powell in a banking committee hearing over whether the dollar's "sacred cow" status is hurting American business owners. The video resurfacing sparked a week of debate among economists, pundits, and media commentators split on whether Vance's view holds up.
• Critics and populists disagree on the fix, but the speaker argues one thing draws consensus: an artificially overvalued dollar is no longer sustainable given the $40 trillion national debt. This framing is the speaker's own view rather than a reported consensus, and is used to argue continued dollar dominance will keep pressuring the middle class and raise the value of gold and silver holdings.

2026-08-22 · open the original ↗

• The trade wall and the yuan-as-funding-currency are one story, not two. Measures that slow Chinese exports push unsold output back into a domestic economy that cannot absorb it, deepening deflation — which pins 10-year CGB yields under 2%, keeps the PBoC passive, and makes the yuan the cheapest large-economy money on Earth just as Deutsche Bank takes up RMB clearing and Panda issuance sets records.

• July data falsifies the overcapacity explanation: the deflation is demand-side. Jan–Jul FAI fell 6.7% y/y (July alone −12.9%, real estate −27.4%, infrastructure −14.7%), yet PPI still printed −3.5% after roughly 41 consecutive months of factory-gate deflation, interrupted only by the Iran-conflict oil spike (Mar–May 2026). Private investment −9.4% versus public −3.3% is the tell: with real rates facing owners at 4.5–6.5%, capital is on strike.

• Three walls built under three different legal families leave Beijing no door to negotiate through. US executive tariffs are fast but revocable (a 12.5% levy replaced the lapsed 10% in late July); EU trade-defence law is slow but durable (steel quotas cut 47% to 18.3Mt, out-of-quota duties 50% through 2031, EV CVDs 35.3%); Global South safeguards spread fastest because they are cheapest to invoke. Odds given: bifurcation ~50%, managed absorption ~30%, disorderly 15–25% devaluation ~20%.

• The September 24 summit is structurally a coupon date, and Phase 1 walls have not yet begun to bind. Xi's single White House day, skipping UNGA, closes a pre-agreed third iteration of rare-earths-for-tariff-relief before November's Busan expiry — the author's read being that Beijing is amortizing a depreciating chokepoint (DFARS 2027, Western stockpiles) in annual tranches. Meanwhile July exports rose 23.9%, semiconductors +122%, EU surplus a record $32.9B, masking the ratchet underneath.

2026-08-22 · open the original ↗

• Xi's one-day Washington trip is a signing, not a negotiation. Asked why Xi comes while the US escalates — Venezuela, strikes on Iran over oil shipped to China, pressure on ASML — the answer is that the term sheet already exists at the staff level. This is the third iteration of the same trade: rare-earth access and $17B/yr of agricultural purchases through 2028 for tariff relief, timed weeks before the Busan truce expires in November.

• China's deflation comes from a household and private-capital strike, not from too many factories. Fixed asset investment is −6.7% y/y (July −12.9%, real estate −27.4%, infrastructure −14.7%) and private investment −9.4% against public −3.3%, yet PPI still fell 3.5% in July after roughly 41 straight months of decline broken only by the Gulf war oil spike. Cutting capacity that hard without stopping the price fall settles the argument.

• The walls have not begun to bind, and devaluation is the last release valve. July exports rose 23.9% after June's 27% surge, semiconductors +122%, a record ~$33B EU surplus — tariff front-running and the AI boom are masking how far the walls have advanced. With a $1.2T surplus trapped inside a "Hotel California" capital account, money piles into deposits and government bonds, holding yields under 2%; he expects origin rules to close transshipment next and asks listeners whether the US should cut the deal at all.

2026-08-22 · open the original ↗

• US national debt crossing $40 trillion, doubling in just a decade, is forcing the Treasury toward an uneasy set of choices. The debt-servicing burden rose 15% this year alone with interest already consuming 20% of tax revenue; a doubled long-bond buyback briefly lowered yields before markets dismissed it as a "band-aid," and Bessent signaled further intervention tools plus unspecified fiscal consolidation ahead.
• The Canada-US trade breakdown and Middle East tensions pushed investors toward gold and Bitcoin as alternative stores of value. Washington's 50% tariffs on Canada prompted retaliation threats, while stalled US-Iran talks and Israeli strikes on Lebanon kept Brent at $94 and WTI at $86; gold surged above $4,600/oz and Bitcoin jumped 22% to $77,200.
• The week ahead hinges on Fed Chair Warsh's Jackson Hole address and Nvidia's earnings as tests of Fed reaction-function clarity and AI-boom durability. Warsh is expected to address his "reaction function" and potential reforms to Fed operations including the dot plot and meeting cadence, while investors look to Nvidia to validate enthusiasm for AI following a week of mixed retail earnings from Lowe's, Target and Walmart.

2026-08-22 · open the original ↗

• No spoken content. The 31-second short has no captions, and the whisper transcript of its audio is empty; its subject is in its title alone: "How much gold is left to mine in the world".

2026-08-21 · open the original ↗

• The Mountain Jews (Juhuro), settled around the Caucasus and Caspian Sea for roughly 1,500 years, are presented as a warrior people distinct from the persecuted stereotype of European Jewish history. He traces their origin to Jews who fled Israel and Judea after the Assyrian conquest of the 8th century BC and the fall of Judea in the 6th century BC, many settling in Persia.
• Persia is described as historically the greatest protector of the Jewish people, of any empire, over more than a millennium, including stationing Jewish communities as frontier guards against the Huns from the 5th century AD. Under Muslim rule after Persia's fall, Mountain Jews kept their lands and were treated as second-class citizens subject to the jizya tax, a status the speaker frames as still far better than Jewish treatment in Europe, where pogroms and bans on land and weapon ownership were common.
• In the 1800s the Mountain Jews allied with Russia's expansion into Persian territory to escape Muslim rule and its taxes, initially fighting alongside the Russians, but Russia turned more discriminatory toward them once its control solidified around the 1890s–1900s.

2026-08-21 · open the original ↗

• Sosnoff names firing people his least-favorite part of the job. In a "how well do you know each other" quiz between the two co-hosts, this was one of the first correct guesses; Sheridan's own first job out of college was a mail-order business selling sunglasses and CDs.
• Food, not cars, is named as Sosnoff's biggest money drain. Sheridan correctly guessed Sosnoff overspends most on food, describing him ordering several dishes at a restaurant "just for himself"; Sosnoff separately says his fantasy alternate career would have been a professional basketball player.
• The quiz ends in a near-tie, correct guesses on both sides. Sheridan guessed Sosnoff's go-to restaurant recommendation is Joe's, correctly beating his own instinct toward Ricardo's; earlier the pair tied 2-2 after both correctly named Sheridan's worst grade, in a course abbreviated "SI."

2026-08-21 · open the original ↗

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

2026-08-21 · open the original ↗

• Coding models for juniors should teach rather than do the work. Hanselman (in a clip from his Felix Rieseberg interview) wants an early-career model that says "I'd like to see you do it first" instead of writing the bubble sort, and wants universities to have access to it: "we should not become a subcognitive species… Just because it can do it doesn't mean it should." Rieseberg answers that his Language Model Builder is free and earns him nothing.

2026-08-21 · open the original ↗

• Treasury Secretary Scott Bessent let Japan borrow dollars against its Treasury holdings through the Fed's FIMA repo facility on July 31, avoiding a forced sell-off after the yen hit a 40-year low near 164 to the dollar. The speaker argues Senator Rand Paul's Fort Knox visit landed right as this bailout was arranged, serving as a calculated distraction from it — the speaker's own inference, not a stated fact.
• Days after the national debt crossed $40 trillion, the 30-year Treasury yield spiked to 5.34%, its highest level since 2007 — cited as evidence that foreign demand for US debt is weakening to the point of requiring emergency intervention.
• Bessent then announced the Treasury would double its long-term bond buybacks from $2 billion to $4 billion per operation, framed publicly as keeping rates stable. The speaker argues this instead artificially props up debt demand, forces more debt issuance to fund the buybacks, and dilutes the dollar's purchasing power through debt monetization — the speaker's own interpretation, presented as a reason to keep accumulating metal.

2026-08-21 · open the original ↗

• Senator Rand Paul's under-2-hour Fort Knox walkthrough on August 10 cannot substitute for the audit he has long championed. He posted that the reported 147 million ounces were "there," but a walkthrough skips unsealing compartments, weighing every bar, and running purity assays — steps the speaker says would take a team of experts months to actually complete.
• The speaker frames the visit as suspicious specifically because Paul is the primary sponsor of the Gold Reserve Transparency Act, so he knows better than most that a glance-and-tweet isn't an audit. This is presented as the speaker's own inference about motive, not a reported fact.
• The visit's real purpose, per the speaker, was to distract from a brewing Treasury crisis: the Japanese yen hit a 40-year low near 164 to the dollar in late July, threatening a forced sell-off of Japan's $1.1 trillion in US Treasury holdings that could have sent US bond yields soaring — averted only when Treasury Secretary Scott Bessent opened the Fed's FIMA repo facility to Japan on July 31.

2026-08-21 · open the original ↗

• The message is about pace, not provenance: patience beats speed. Over footage described as real pirate silver recovered from the Whydah shipwreck, the voiceover argues viewers "are not behind in life" and that building strength slowly is better than building quickly — a motivational framing rather than any claim about the coin's history or value.

2026-08-21 · open the original ↗

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

2026-08-21 · open the original ↗

• The real news isn't the buyback, it's that Treasury broke its own calendar. Bessent's buyback signal had a one-session half-life before yields rose again, but the author argues the bigger story is that Treasury released a full-quarter buyback schedule just two weeks ago and then abandoned it for "liquidity support" — a break from the "regular and predictable" debt management norm every administration has held since the 1970s, and a credibility hit at the exact moment the Fed under Warsh is trying to let markets set policy.

• Bessent frames both interventions as defense, not price-targeting, because naming a level would be an admission of manipulation. He confirmed joint yen-buying on August 3 and doubled long-end Treasury buybacks on August 19, calling the former a fix for "disorderly markets" and the latter "liquidity support" — never citing a yen level or yield target, since doing so would expose him to regulators and prosecutors rather than just market skepticism.

• The author's proposed test for manipulation is symmetry, and Bessent likely fails it. Would he intervene if the yen kept strengthening past levels that hurt Japan's Treasury holdings, or if yields fell sharply instead of rising ahead of the midterms? If not, the goal is a specific price, not orderly markets — notable given Bessent's own career was built on betting against officials who defend a line fundamentals won't support, including roughly $1 billion shorting sterling as part of Soros's team that broke the Bank of England in 1992.

2026-08-21 · open the original ↗

• Choosing a model only swaps which frozen weights answer. Drawing on a July Anthropic blog post by Lydia Hallie of the Claude Code team, the host explains that weights are read-only at inference, so context steers a prediction but never teaches the model, and a library absent from training is retained only for the request where its docs are pasted. At the time of recording Claude Code's choice was Sonnet, Opus or the newest and largest, Fable; bigger models know more and cost more per output token.
• Effort level is not a thinking-time slider. Effort shapes all output tokens (reasoning, tool calls and the text written to the user), which are billed at the same rate: high effort reads more files, verifies more and goes further before checking in, while low effort would rather ask for context. Anthropic's same-prompt comparison had high effort generating roughly seven times more tokens; the host stresses effort sets how far the model is willing to go, not how far it must, and that Anthropic watches for overthinking in training.
• Ask whether it didn't try hard enough or didn't know enough. Skipped files, unrun tests or an abandoned refactor call for more effort; a model that had the context, visibly tried and was still confidently wrong needs a larger model. Before touching either dial, fix a vague prompt or missing context, which the host calls the more common culprit; he also advises starting from the provider's default effort.
• Cheaper per token isn't always cheaper per task. On routine work both sizes succeed, so the larger model's price is wasted; on hard multi-step work the small model grinds near its ceiling while the large one reaches the bar in fewer steps. In Anthropic's testing, Fable finished long multi-step jobs that Opus and Sonnet couldn't reach at any effort level, which the host says argues for saving it for work that needs it.
• The whole industry has converged on these two dials. The host recounts OpenAI's GPT-5 router in ChatGPT, later replaced after user pushback by explicit modes and a reasoning-effort selector, and its GPT-5.6 tiers: Sol as flagship, Terra at half Sol's price, and Luna as fastest and cheapest. He reports OpenAI has confirmed a forthcoming model, Astra, meant to coordinate multiple agents on long-running problems, whose internal version reportedly cracked 10 open problems in mathematics and theoretical computer science, with no release date; Google moved from Gemini 2.5's raw thinking-token budget to Gemini 3's discrete thinking levels.

2026-08-21 · open the original ↗

• Prompting is not programming, because prompts are inherently ambiguous. Hanselman, the guest here on two California high-school students' podcast (cross-posted to his feed), says two people giving the same prompt get different code, while the same C for loop runs deterministically "until the end of time". His examples are the "get bread, and if they have eggs, get 12" joke and a one-shot "Minecraft clone" where "the word Minecraft did all the work".
• A spec is a lossy compression of an app. Like re-saving a JPEG or copying a mixtape, turning an app into a spec and prompting it back pushes every ambiguous part to the statistical mean, so the regenerated site "will look different, guaranteed"; a cloned repo builds identically. AI websites all look alike because they are "the statistical average website", and he says programmers must push away from that "gravity well" of mediocrity.
• Vibe coding is Ubering; early-career engineers must learn stick shift. His advice now, as against 2023: learn how the machine works, and separate computer science from software engineering, "the science of shipping" (the development life cycle, servers, security reviews). A vibe-coded app built on a road trip is a prototype, "but it ain't done". He rejects "vibe coding sucks" as too simple and describes a slider from "slop cannon vibe coding" to "AI augmented software engineering": vibe the prototype, then do proper work to reach production.
• The danger is unmaintained slop, not AI itself. Answering a host's analogy to diabetes damaging the body over decades, he calls bit rot the software equivalent, citing whether his Baby Smash app still builds. "AI doesn't take your job. Unchecked capitalism takes your job," he says, worrying about who cleans up insecure, rotting sites in 5 or 10 years.
• Reliability and focus matter more than being the smartest person. He calls adulthood "a big group project" where being pleasant and saying "yes, and" beats having the answers. His other advice: do what you say you will, use David Allen's four Ds (do it, drop it, delegate it, defer it), do anything under five minutes, and be five minutes early. Social media and doom-scrolling, he says, are "destroying you". He also urges learning how to learn and keeping a BS detector, since "half of the internet is trash".

2026-08-21 · open the original ↗

• Prompting is not programming, because prompts are inherently ambiguous. Hanselman, the guest here on two California high-school students' podcast (cross-posted to his feed), says two people giving the same prompt get different code, while the same C for loop runs deterministically "until the end of time". His examples are the "get bread, and if they have eggs, get 12" joke and a one-shot "Minecraft clone" where "the word Minecraft did all the work".
• A spec is a lossy compression of an app. Like re-saving a JPEG or copying a mixtape, turning an app into a spec and prompting it back pushes every ambiguous part to the statistical mean, so the regenerated site "will look different, guaranteed"; a cloned repo builds identically. AI websites all look alike because they are "the statistical average website", and he says programmers must push away from that "gravity well" of mediocrity.
• Vibe coding is Ubering; early-career engineers must learn stick shift. His advice now, as against 2023: learn how the machine works, and separate computer science from software engineering, "the science of shipping" (the development life cycle, servers, security reviews). A vibe-coded app built on a road trip is a prototype, "but it ain't done". He rejects "vibe coding sucks" as too simple and describes a slider from "slop cannon vibe coding" to "AI augmented software engineering": vibe the prototype, then do proper work to reach production.
• The danger is unmaintained slop, not AI itself. Answering a host's analogy to diabetes damaging the body over decades, he calls bit rot the software equivalent, citing whether his Baby Smash app still builds. "AI doesn't take your job. Unchecked capitalism takes your job," he says, worrying about who cleans up insecure, rotting sites in 5 or 10 years.
• Reliability and focus matter more than being the smartest person. He calls adulthood "a big group project" where being pleasant and saying "yes, and" beats having the answers. His other advice: do what you say you will, use David Allen's four Ds (do it, drop it, delegate it, defer it), do anything under five minutes, and be five minutes early. Social media and doom-scrolling, he says, are "destroying you". He also urges learning how to learn and keeping a BS detector, since "half of the internet is trash".

2026-08-20 · open the original ↗

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

2026-08-20 · open the original ↗

• Soares says the only winning move is never building it. A superintelligence that doesn't care about you is "game over", he argues, leaving no vulnerability for a last-minute rescue "like Tom Cruise" in a film. The clip introduces him as having spent over a decade on AI alignment.

2026-08-20 · open the original ↗

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

2026-08-20 · open the original ↗

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

2026-08-20 · open the original ↗

• Short-duration Treasury ETFs are the preferred parking spot for cash in a $100,000 trading account. Both hosts favor BIL or SGOV, calling them interchangeable, with one noting SGOV carries slightly lower cost; one adds an SPX box could pay about 50 basis points more but isn't practical at that account size.
• For non-trading cash, the preference shifts to short-term CDs and T-bills over money market funds. One host says he mostly uses three- and six-month CDs (sometimes 90 days) and T-bills in accounts outside his trading account, specifically avoiding money market funds.

2026-08-20 · open the original ↗

• Trading is about price, not textbook narratives about what "should" happen. Bonds failed to rally despite the 10-year yield spiking to 19-month highs amid escalating geopolitical tension, defying the textbook script. One host argues headlines are useful for financial-media conversation but irrelevant to trading, which comes down to reading price and volatility rather than predicting how news "should" move markets.
• Research places the best risk-adjusted premium-selling window at 21-45 days. The hosts say years of internal research show rolling short premium at 21 days out, rather than holding into the final stretch, delivers the best return for the risk taken; each month further out reduces risk by roughly 25-35%, since theta decay accelerates sharply only inside three weeks.
• Rising bond yields are having a surprisingly muted effect on chip stocks. Despite the 10-year hitting 4.75% and global bond yields spiking to multi-decade highs, Nvidia, Micron and Broadcom showed only scattered, contradictory daily moves rather than a clean sell-off tied to borrowing costs. Hosts dismiss the day-to-day divergence between chip names as prop-firm rotation rather than a real trend worth trading.
• One host calls this the best chance in decades to favor bonds. With yields at multi-year highs, he says he'd choose long bonds over stocks if forced to pick one, and floats trimming equity allocation for short-term bond exposure, though he stops short of making the trade himself, saying he's waiting for a cheaper entry price.

2026-08-20 · open the original ↗

• A soul file restores a stateless agent's identity on every call. Hanselman calls OpenClaw's soul.md "so clever" because each "are you there?" makes the agent "wake up from a dead sleep" and ask who it is. He likens it to Memento, whose protagonist tattoos every new piece of context onto himself.

2026-08-20 · open the original ↗

• The host calls the Pepwave adapter a fuller UniFi 5G Max. It costs $399, adds 5G WAN to any router over one ethernet cable, and he says it has more features and no memory surcharge; the Crosstalk Mobile store selling it offers free shipping and 10% off equipment paired with its 5G unlimited SIM.
• Hardware covers antennas, power and SIM options broadly. It has four SMA connectors for an external antenna, a 2.5 Gbps PoE-in port (the host powers his from a UniFi UCG-Fiber's PoE WAN), a 12 V 2 A adapter or direct 12/24 V DC, two SIM slots, eSIM, remote SIM, FusionSIM, a GPS puck, and a bundled 1 GB eSIM.
• Fit antennas before power, then demote the bundled eSIM. The host warns powering up first can cook the radio. Default IP is 192.168.50.1, login admin/admin; out of the box it ran on the included eSIM ("Webbing USA"), so he moved SIM A to priority one and the eSIM to three. As of recording firmware was 2.0.3.
• Read SINR, RSRP and RSRQ, and exclude congested bands. His thresholds: SINR 20+ great, 13–20 good; RSRP −85 dBm excellent, −100 okay, −110 bad; RSRQ −10 dB good, −15 rough, −20 really bad. UniFi's GUI exposes only RSRP; good RSRP with bad RSRQ signals tower congestion, which Peplink's custom band selection (e.g. dropping n41) can avoid.
• Add it to InControl 2 before enabling IP passthrough. Once registered by serial number it phones home and stays remotely manageable even with no local access. Passthrough (on VLAN 1 here) ends double NAT, giving the router a routable IPv4 address; with Crosstalk Mobile's static IP, the host says forwarded ports keep working on 5G backup without CGNAT.

2026-08-20 · open the original ↗

• Syria still holds leftover nuclear material from a program destroyed in 2007, but its government is currently cooperating with a UN probe overseeing it. The program traces back to a deterrence strategy against Israel begun by Hafez al-Assad after Israeli troops reached the outskirts of Damascus within a week of the Yom Kippur War, initially built around Soviet-backed chemical weapons and scud missiles rather than nuclear arms.
• Syria secretly began a nuclear program in 1997 with material and assistance chiefly from North Korea, starting construction of the al-Kibar site around 2002; Israel discovered the site in early 2006 and found a reactor but no fuel or key components yet in place. Israel brought the discovery to the United States in 2006, warning Syria was pursuing weapons of mass destruction capability.
• The Bush administration, still facing fallout from the discredited Iraq WMD intelligence and the unpopular Iraq invasion, declined to act militarily and pushed for a diplomatic approach instead. Israel struck unilaterally in 2007 with Operation Orchard, destroying Syria's nuclear facility.

2026-08-20 · open the original ↗

• The August 19 buyback expansion is a put option on dealer inventory, not new demand. Bessent doubled long-end (10–30yr) liquidity-support operations effective September 9: from $2B maximum twice a quarter to $4B minimum four times, while the total quarterly envelope stays at $38B. Day one, the 30-year fell ~9bp to 5.196% and the 10-year ~6bp to 4.647% as one-way shorts were squeezed.

• The leverage ratio is what traps the dealers and creates the doom loop. All 26 primary dealers must bid pro-rata (~$385mm each on a $10B auction), so auctions cannot fail — but unsold inventory consumes capital under a rule that treats Treasuries like junk loans. The April 1, 2026 rule lowered the enhanced requirement without carving out Treasuries, so after the late-June buyers' strike drove the 30-year to its highest yield since 2007, dealers bid more timidly, pushing yields higher still.

• The real cost is a quietly shortening U.S. debt maturity profile. Buybacks are funded by issuing bills, swapping locked-in 30-year funding for paper that must be rolled every few months — Operation Twist run out of the Treasury rather than the Fed, exploiting bottomless money-fund front-end demand. Structural duration oversupply is untouched. The escalation tell to watch: the $38B envelope rising while the bill share of issuance rises with it, which would mark de facto yield-curve control.

2026-08-20 · open the original ↗

• Treasury is refinancing the country from a 30-year fixed into a 6-month ARM. His central analogy: swapping a 30-year 5.2% mortgage for an adjustable that rolls every six months at 4%. Buybacks of 10–30yr paper are funded with short-term bills, so total debt is unchanged but duration is handed back to Treasury — cheaper today, with the rollover risk pulled forward.

• The move relieves pressure but changes nothing structural — and he says the tape already proved it. The quarterly envelope stays at $38B; only the long-end share doubles. He notes yields were back up the next day, higher than before the announcement, because the buyers' strike since June 2026 reflects a genuine shortage of duration demand. The value is the "put" — dealers now know they have a guaranteed $4B exit before they bid.

• He speculates the operation is an order of magnitude too small, and flags a foreign-dealer tail risk. Against roughly $128B of buybacks versus ~$400B of long-end supply, he argues something nearer $300B is what would actually be needed — a rough, loosely-stated figure. Separately he observes only eight of the 26 primary dealers are American-owned and asks, hypothetically, what happens if the European, Japanese, and Canadian arms stop bidding.

2026-08-20 · open the original ↗

• Data's supposed "zero terminal value" turned out backward, the host argues. Francois, who left his PhD to found Focal Systems, says VCs once wrote off data-labeling companies as commodities, but the category has generated over $100 billion in market cap in a decade; he argues each expert domain — an AI doctor, lawyer, or trader — likely supports only one dominant specialized data company, the way no one builds a separate Instacart for Apple versus Android.
• Snorkel's newest coding benchmark rates AI models on "taste," not just correctness. Vincent Chen says SeniorSWEBench, built with the original SWE-Bench team at Princeton, adds a "tasteful pass" metric for whether generated code matches how a senior engineer would actually write it — not just whether it merges — and found Fable, Opus, and Sonnet tied for first place as of the week of the talk.
• Inception's diffusion models generate text in parallel, not sequentially. Volo says the resulting Mercury 2 model tops 1,000 tokens per second, addressing the LLM latency bottleneck in cascaded voice pipelines. He credits the larger quality gain to Tau Forge, a synthetic-data system that builds realistic RL environments from real usage logs, which improved the Mercury 2.5 preview model's accuracy on held-out environments by over 23%, up from a 50% baseline, matching state-of-the-art open- and closed-source models.
• Cross-lingual transfer in pre-training is asymmetric, MIT researcher Shane found. In a multilingual pre-training study presented at ICLR, one language can help another without the reverse holding — Portuguese may aid Spanish without Spanish aiding Portuguese — and shared script mattered more for transfer than shared language family; the work also found Thai gets just 0.6% of the token volume English gets in the Madlad-400 training corpus.

2026-08-20 · open the original ↗

• The app teaches language models by having you build one. Felix Rieseberg leads engineering on Anthropic's Claude apps (Claude, Cowork, Code Desktop) and built Language Model Builder as a free side project. He says its interactive textbook and playgrounds take someone with no engineering background to a "very stupid" model in an afternoon. It is Mac-only on Apple silicon because he built it in Swift on MLX; Hanselman presses for a Windows version, and Rieseberg says an agent could probably port it. Hanselman likens it to Ben Eater's build-your-own 6502 and to learning stick shift: it "changes your relationship with the vehicle".
• It stops just before the curve where human understanding breaks. Rieseberg capped the textbook at about 90 minutes ("no artificial additives"), with only enough history to explain temperature and the 2017 "Attention Is All You Need" paper. He settled on a roughly 10K tokenizer because larger ones took too long on his M4 Max. A one-day run produced "the sun is a moon at the bottom of the earth", which he notes at least treats the question as one to answer. Hanselman frames the scale as GPT-2, "just before the bendy part of the hockey stick".
• Seeing the paths not taken demystifies the word guesser. The app has a model blueprint visualizer and an "x-ray view" of the alternative tokens at each position, and can import a small base model such as Qwen to sample from. Hanselman describes his own heat-map demo, in which "beach" is about 75% and "park" 20%. Rieseberg says knowing what a transformer block and weights are lets people join debates on moral patienthood, intelligence, and why researchers speak of "growing" models.
• Fine-tuning steers a small model cheaply; helpfulness is trained in. Pre-training takes at least a day in the app, but about 800 fine-tuning steps are "more than enough" to steer a small model. Examples include math, rewriting, a "sassy mean bot" and an "admittedly terrible" group-chat import. Hanselman stresses that a baked model doesn't keep learning and that conversation state "is an illusion just made by context". Rieseberg adds that much personality is already trained in, which is why OpenClaw's soul.md works.
• Larger models may come to recognise when they're asked to cheat. Rieseberg hopes for this and describes three stages: a model that does whatever it is given, one that follows hidden instructions (Golden Gate Claude; teachers' white-text "eggplant" traps in syllabi), and one with enough "humanity and morality" to see it is being asked to help a student cheat. Hanselman connects this to a preceptorship he is building with Russinovich at Microsoft: coding models that make juniors do the work, because "we should not become a subcognitive species".

2026-08-20 · open the original ↗

• Win rate is defended as meaningful, not just average win-loss size. Even though a single outsized loss can erase nine winning trades, one host argues win rate itself teaches traders "how to win" as a process, calling it the foundation of the firm's premium-selling philosophy over decades rather than a statistic that average trade size alone should override.
• Gold's rally to $4,500 is called cyclical trading, not a real hedge. After tumbling from $5,500 to $4,000, gold has bounced back over 10%, but hosts reject framing it as inflation protection or flight to quality, calling it ordinary commodity range-trading; with implied volatility rank down to roughly 30 from a peak near 130, neither wants much exposure at current levels.
• Crypto down 25-47% year-to-date splits the hosts on whether it's "dead." One calls Bitcoin, Ethereum, Solana and XRP a product with limited upside as regulation reshapes its original use case; the other frames digital assets as a technology-based asset class still mid-cycle, comparing it to commodities that spend long stretches in the doldrums before big moves, and says patience gets rewarded.
• Buying options into earnings is called a loser's game long-term. Hosts say internal research shows roughly 80% of stocks land inside their options-implied earnings move, meaning premium buyers usually lose to the volatility crush even when right on direction; their own rule is never to buy premium into earnings and to stay outside the expected move, sometimes twice as far.

2026-08-19 · open the original ↗

• The app teaches language models by having you build one. Felix Rieseberg leads engineering on Anthropic's Claude apps (Claude, Cowork, Code Desktop) and built Language Model Builder as a free side project. He says its interactive textbook and playgrounds take someone with no engineering background to a "very stupid" model in an afternoon. It is Mac-only on Apple silicon because he built it in Swift on MLX; Hanselman presses for a Windows version, and Rieseberg says an agent could probably port it. Hanselman likens it to Ben Eater's build-your-own 6502 and to learning stick shift: it "changes your relationship with the vehicle".
• It stops just before the curve where human understanding breaks. Rieseberg capped the textbook at about 90 minutes ("no artificial additives"), with only enough history to explain temperature and the 2017 "Attention Is All You Need" paper. He settled on a roughly 10K tokenizer because larger ones took too long on his M4 Max. A one-day run produced "the sun is a moon at the bottom of the earth", which he notes at least treats the question as one to answer. Hanselman frames the scale as GPT-2, "just before the bendy part of the hockey stick".
• Seeing the paths not taken demystifies the word guesser. The app has a model blueprint visualizer and an "x-ray view" of the alternative tokens at each position, and can import a small base model such as Qwen to sample from. Hanselman describes his own heat-map demo, in which "beach" is about 75% and "park" 20%. Rieseberg says knowing what a transformer block and weights are lets people join debates on moral patienthood, intelligence, and why researchers speak of "growing" models.
• Fine-tuning steers a small model cheaply; helpfulness is trained in. Pre-training takes at least a day in the app, but about 800 fine-tuning steps are "more than enough" to steer a small model. Examples include math, rewriting, a "sassy mean bot" and an "admittedly terrible" group-chat import. Hanselman stresses that a baked model doesn't keep learning and that conversation state "is an illusion just made by context". Rieseberg adds that much personality is already trained in, which is why OpenClaw's soul.md works.
• Larger models may come to recognise when they're asked to cheat. Rieseberg hopes for this and describes three stages: a model that does whatever it is given, one that follows hidden instructions (Golden Gate Claude; teachers' white-text "eggplant" traps in syllabi), and one with enough "humanity and morality" to see it is being asked to help a student cheat. Hanselman connects this to a preceptorship he is building with Russinovich at Microsoft: coding models that make juniors do the work, because "we should not become a subcognitive species".

2026-08-19 · open the original ↗

• Three-standard-deviation overnight moves are called a routine cost of selling premium. One host says such rare outlier moves — expected roughly 1% of the time — have hit him twice already this year, in silver and Micron, and calls them a near-certainty for any active premium seller rather than a fluke.
• One earnings cycle is the limit for riding out a bad trade. He says he'll tolerate one earnings report going against a position, adding delta as a partial hedge, but by the second earnings report he holds himself to closing it regardless of outcome — discipline he contrasts with positions he's otherwise kept rolling for up to two years.
• Rare blowups are framed as normal, not evidence of a broken strategy. Having gone some years without a three-standard-deviation loss and other years with two or three, he treats the frequency as inherently random rather than a signal to change approach.

2026-08-19 · open the original ↗

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

2026-08-19 · open the original ↗

• His former nuclear-site colleagues sought his help when the war began. Maynard Holliday, then at the Pentagon as assistant secretary for critical technology, says former Chernobyl colleagues messaged him on LinkedIn that Russians had damaged their equipment. He put them in touch with a Department of Energy nuclear safety working group, which got them computers to monitor the new sarcophagus.
• Occupying forces damaged the new containment structure over the reactor. He reports that a Russian drone holed the hangar-like structure over unit four, which he says the European Bank for Reconstruction completed "probably less than a decade ago". He adds that Russia had seized Zaporizhzhia, Europe's biggest nuclear plant, and shelled from it. Before the war the Chernobyl team had asked him for help designing gantry robots to take apart the original sarcophagus and size-reduce the fuel.

2026-08-19 · open the original ↗

• The Treasury buyback increase is confirmation of a long-end buyer strike, not a liquidity tweak. Treasury doubled buybacks from $2B to $4B a month starting September 9th, targeting the 10-20 and 20-30 year sectors; 10-year yields fell ~6bp and 30-year ~9bp on the news. He reads the official language about "strong sponsorship" in long-dated operations as an admission that dealers have been selling paper back since late June.

• AI hyperscaler issuance is crowding out the sovereign, reversing the usual direction. Roughly $400-474B of hyperscaler corporate bonds this year compete for the same institutional dollar as Treasury supply, at 6.4% (Alphabet) and 7.5% (Meta) versus ~5.2% on Treasuries. Banks are offering back roughly $40B/month against a $15B cap — about 2.7x oversubscribed — and $128B of annual buybacks covers only ~32 cents on the dollar of new supply. A pressure-relief valve, not a fix.

• This is reactive yield-curve control with a Japanese end state, and the precedent cuts against it. He calls it mini-QE that lowers long yields without expanding money supply, but pushes duration to the short end — $15-18T in bills now, possibly $20T by year-end, half the debt rolling annually. He cites academic work that the 2000-2002 buyback program ultimately raised yields 95bp, and flags a possible BOJ quarter-point hike in September; the 2024 hike, he claims, erased $6T in global equity value.

2026-08-19 · open the original ↗

• The hyperscaler crowding-out thesis is now Treasury-acknowledged fact. Treasury doubled long-end buybacks from $2B to at least $4B per operation across the 10–20 and 20–30 year sectors, effective September 9 through the November 4 refunding; the 10-year fell 6bp to 4.647% and the 30-year 9bp to 5.196%. The official "strong sponsorship... significant volume of high-quality offers" language is read plainly as institutions queuing to hand long bonds back — a buyers' strike since late June, evidenced by $40B of offers against a $15B cap (2.7x).

• Bessent's two levers are arithmetically short of the problem. The $440B shift into short-term issuance plus ~$128B/year of maximum buybacks totals $568B of long-end relief, but $474B of annual hyperscaler corporate supply already exceeds the issuance shift alone, and buybacks cover only 32 cents on the dollar against $400B/yr. Offsetting it would require roughly $300–400B/year — 3x the new maximum — at which point it stops being liquidity support and becomes yield curve control, the policy Japan ran from 2016–2024 until the BOJ owned over 50% of JGBs.

• "Mini QE" gets the rate effect right and the monetary effect wrong. Fed QE creates reserves; Treasury buybacks swap long debt for bills without expanding the money supply, so inflation risk is low in isolation while rollover exposure at the short end rises. The underappreciated risk is that 2000–2002 buyback studies found repurchases pushed yields up by an average 95bp, since removing off-the-run supply forces new on-the-run issues to clear at higher yields — Bessent is betting 2026 market structure differs enough that this does not repeat.

• The self-defeat loop and a three-way September trigger are the tradeable parts. If buybacks succeed and the 10-year reaches 4.3–4.4%, yen-funded carry narrows to ~330bp and migrates to EM assets (baht, real, peso) and to long-dated AI paper at 6.4–6.6%, costing Treasury a marginal buyer and deepening the strike; if they fail, Japan's $1.14–1.19T Treasury position faces forced liquidation. September stacks a likely BOJ hike to 1.25%, possible hyperscaler credit stress (CoreWeave), and the buyback start — versus 2024, when one 25bp hike wiped ~$6T in equity value. The level to watch is USDJPY 155, a floor manufactured by the $85B August 3 intervention; through it, 152 and full carry unwind.

2026-08-19 · open the original ↗

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

2026-08-19 · open the original ↗

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

2026-08-19 · open the original ↗

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

2026-08-18 · open the original ↗

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

2026-08-18 · open the original ↗

• Fixed regulatory red lines don't survive AI's pace of change, he argues. Kratsios points to the EU AI Act, finalized before ChatGPT existed and now binding models it never anticipated, and to the Biden administration's hard compute-disclosure threshold, as examples of rules that "do not stand the test of time" once set in statute.
• A single national AI standard is the administration's top ask of Congress. Kratsios says large companies like Google can "hire an army of lawyers" to navigate different rules in California, Maryland, and Texas, but a state-by-state patchwork would lock out smaller startups — a federal preemption push the president raised when the AI Action Plan was released last July.
• Rumors of a coming crackdown on open-weight models are false, he says. After a founder letter (organized in part by Y Combinator) warned against a rumored executive order restricting open weights, Kratsios says the administration's position is unchanged from "chapter one" of the AI Action Plan he co-authored with David Sacks and Secretary Rubio: a commitment to a "vibrant closed and open source ecosystem."
• A new science report envisions AI agents running autonomous "cloud labs." Written in response to a Trump letter modeled on FDR's 1945 charge to Vannevar Bush, the report describes AI posting bounties for experiments and settling results on a ledger, paired with concrete process changes — fast grants decided in under a month, every major agency filing an action plan within 90 days — as federal funding has fallen to roughly 30% of U.S. R&D, down from about 70% in 1950.

2026-08-18 · open the original ↗

• Comic-book scientists of African descent were his early role models. Maynard Holliday cites Tony Stark, who "has no superpower but his mind", and T'Challa, who in the comics is Wakanda's chief scientist rather than Shuri as in the current MCU. He says seeing people of African descent in positions of power, authority and knowledge made an impression on him as a boy.

2026-08-18 · open the original ↗

• Hosts split on whether new highs amid extreme valuations are a "melt-up." With over 85% of reporting companies beating earnings estimates yet stock reactions running only 50/50, one host calls the rally a genuine, if mild, melt-up; the other dismisses the term as a "dumbass label" people reach for once a rally runs longer than their own comfort level.
• Record $1.5 trillion margin debt is called scary but not tradeable. Margin debt is up roughly 51% year-over-year and sits near 2.2% of total equity value, a level the hosts call a real red flag in hindsight but not systemic risk, since options and futures traders don't borrow this way; they say the number offers nothing actionable in the moment.
• A shift to call-skew in index options is called the bigger red flag. One-month skew has hit its lowest level since mid-2024 as investors abandon downside hedges for upside calls, a reversal from the historical put-skew pattern. One host says this scares him more than record margin debt, comparing bullish call-chasing to interest rates going negative as a signal something is off.
• Michael Saylor's pivot away from an all-in Bitcoin stance is called vindication. After years of calling MicroStrategy's Bitcoin-treasury model a "scam," one host notes the company recently sold billions in Bitcoin near a recent low and raised capital to diversify into non-Bitcoin investments, despite Saylor's past pledge to never sell — evidence, he says, that Bitcoin had a single point of failure.

2026-08-18 · open the original ↗

• "Never let a winner turn into a loser" is called a ridiculous rule. One host argues that since most trades hover near 50-delta, they routinely pass through being a loser before becoming a winner and vice versa, making the rule near-impossible to honor consistently and unsound as a general principle.
• One host defends taking profits early, at 25-50% of a target. He says once a trade hits a set profit level — even just a modest 10-20 cent gain — he exits rather than holding for the full 75-100% potential, accepting that some early exits leave money on the table.
• Never adding capital to a losing trade is offered as the better rule. Rather than fixating on protecting a winner from reversing, the hosts suggest the more useful discipline is refusing to throw additional money at a position that's already going against you.

2026-08-18 · open the original ↗

• Hyperscalers are crowding out the US Treasury — the reverse of the textbook mechanism. The Big Five sold $159B of bonds in the first five months of 2026, up 47% y/y against a $28B/yr average from 2020–2024. S&P counts $225B including Nvidia and puts the full-year pace at $400B; Nomura reckons big-tech borrowing alone equals ~25% of Treasury net issuance to private investors, five times its 2025 share. BofA attributes +25–35bp on the 10-year to this supply competition.
• Bessent's short-end issuance strategy was offset dollar-for-dollar and then some. Barclays estimated the deliberate shift toward bills curbed net Treasury note and bond supply by roughly $440B in 2026 — but hyperscalers added $474B of largely long-dated corporate supply, leaving more duration in the market than before the strategy was implemented. Morningstar shows the rotation is structural, with IG bond funds at a three-year-high 30% corporate allocation.
• The tipping-point signal is the collapse in oversubscription, and the first default is named. Demand coverage on hyperscaler deals fell from nearly 5x in February 2026 to below 2x by July; under ~1.5x, primary dealers absorb the slack onto their own balance sheets. The likely first failure is a Tier 2 GPU cloud operator — CoreWeave, given Microsoft is 60%+ of revenue. Phase 2 implies +135–285bp spread repricing and $40–80B of forced index selling within days.
• The yen carry trade is the funding architecture behind the AI bond wave, not a parallel risk. Carry investors borrow JPY at 1.0% and buy the same paper — Alphabet 30Y at 6.4%, Meta data center bonds above 7.5% — a 550–650bp spread. With BOJ at a 1995-high 1.0%, JGB 10Y at 2.84%, and short-yen positioning near 2007 extremes (124,575 contracts, $9.5B), a September hike landing with AI credit stress is a compound trigger. Treasury flows then conflict: domestic flight-to-quality buying against Japanese repatriation selling, with no historical precedent.

2026-08-18 · open the original ↗

• He concedes his own crowding-out call ran in the opposite direction. The Five Factor thesis predicted sovereign issuance would squeeze corporate borrowers; instead the FT's August 18 front page reports long-term government borrowing costs at multi-decade highs driven partly by AI issuance. His framing: "half right and half wrong" — the squeeze is real, but hyperscalers are doing it to the Treasury rather than the other way round.
• Passive index mechanics are the transmission channel, not credit analysis. Tech borrowers now outweigh banks in several major IG indices — a first — and roughly 30% of investment-grade ETF holdings are hyperscaler paper. A single fallen-angel downgrade forces every passive IG fund to sell the same day, unwinding positions built by $159B of issuance in five months and producing an estimated $40–80B of forced selling.
• Phase 3 hands Treasury a narrow refinancing window before the oligopoly re-forms. When the $400B/yr corporate pipeline collapses, flight-to-quality rallies the 30-year and lets Bessent term out debt cheaply — but only for six to eight months before $2T deficits reassert. The endgame is two or three government-backed issuers and JPMorgan's estimated $375B of AI infrastructure debt from 2026–2030, with crowding-out resuming permanently at larger scale.

2026-08-18 · open the original ↗

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

2026-08-18 · open the original ↗

• The setup starts three pegs at three corners of a square. The legal move is a jump-over: a peg leaps over an adjacent peg to land an equal distance beyond it on the opposite side, and this can be repeated indefinitely.
• Jumps aren't limited to straight lines — diagonal jumps are also legal. Two pegs can hop over each other repeatedly along a straight line, but a peg can also jump diagonally over another, opening up different angles of movement across the board.
• The question: can a peg ever reach the square's fourth, empty corner? If it's possible, the puzzle asks for the smallest number of moves required, with a proof that no faster route exists.
• If it's impossible, that too has to be proven, not just asserted. The puzzle offers no hint as to which outcome holds — reaching the corner or ruling it out both demand a rigorous argument, in keeping with the video's puzzle-of-the-month format.

2026-08-18 · open the original ↗

• AI raised the technical floor; taste is now the ceiling. Vergadia argues that because models are trained on average internet data, what they produce is at best average, so value lies in a person's own perspective and lived experience. When hiring she says she no longer looks for skill with a particular tool, coding included, but for whether candidates have a perspective of their own and can generate ideas.
• AI-assisted work earns attention when lived experience is visible in it. She discloses on her blogs that she uses AI for research, but says readers come for the examples drawn from her conversations with CTOs and other executives. Her sketch tool is Gemini's Nano Banana fine-tuned on some 300 of her hand-designed sketches, yet she says a bare prompt still yields slop; a storytelling skill (pick one word for the concept, then an everyday image of it) runs first and needs her approval before anything is drawn.
• Build a skill by writing the task as a human would do it. Vergadia breaks blog writing into explicit subtasks, each with its own prompt: research restricted to high-quality academic and institutional sources, a human-in-the-loop check where an agent she runs in Hermes texts her the research doc (or audio), then analogy and storytelling, then her format of three developer-focused images and one practical example drawn from her. She says this gets output about 80% of the way, against slop from "write a blog on this topic".
• Most of an AI budget belongs to upskilling, not tools. Her 10-20-70 rule puts 10% on tools, 20% on execution and 70% on education and upskilling. She cites a just-released Stanford impact report putting real production AI use cases at about 8–9%, with the rest experimentation, and blames buying tools without forming habits, which leads to "token maxing" on trivial tasks. Showing staff real use cases from their own jobs and building internal communities follows a J-curve of about 6–8 months before productivity appears; the host adds an Economist analysis suggesting organisations would need to spend an order of magnitude more on transformation than on tools.
• Enterprise AI is won on trust and storytelling, not products. Vergadia says that at Microsoft, brought in on GitHub Copilot's go-to-market as Cursor was gaining ground, she built a team of outcomes-based engineers for Fortune 100 CTOs before "forward-deployed engineer" was a term. She argues that as products and résumés converge, product storytelling and career storytelling differentiate, the focus of her own new firm and of a tech-storytelling book she announced on the show.

2026-08-18 · open the original ↗

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

2026-08-17 · open the original ↗

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

2026-08-17 · open the original ↗

• Even the most sophisticated models mostly recommend insecure code. Hanselman cites the guest's own figure that 60% of AI-generated code is insecure by default and asks why models should be trusted to find the problems they cause. The guest (Varun Badhwar; the clip is from his Endor Labs interview) answers that on the benchmarks they run, 70 to 80% of code recommended by even the most sophisticated models is insecure by default.