Souk's Notes - Unofficial Summaries

2026-06-03 · open the original ↗

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

2026-06-03 · open the original ↗

• A prompt file is a Markdown prompt template run by a slash shortcut. Instead of writing out a whole prompt, you type the shortcut and it expands into the prompt, with optional parameters for the words that change each run. Giard says this suits prompts run repeatedly, especially ones slow to compose or type.
• A plain prompt works once but must be edited every month. His use case is a social-media planning file: a prompt told Copilot to add text to social media.txt, repeating a template (a "July DD" date, an "I am grateful for" line with #gratitude, and #songoftheday) 31 times with DD replaced by a sequential number. It produced the entries in about 30 seconds, but each month he would have to change the month name and the day count by hand.
• Typing /prompts and choosing "new prompt file" creates the template. The default location is the .github/prompts folder, which VS Code creates if missing; the file is the prompt's name followed by .prompt.md (his is add-dates.prompt.md), with a name and description at the top and the prompt below.
• Parameters are dollar-sign input placeholders, each with a name and description. He replaced "31 times" with a number-of-days-in-the-month input and "July" with a month-name input. Running /add-dates then asked for each value in turn (July, 31) and produced July 1 through July 31 as before.
• Editing the Markdown file changes every future run. He added a joke-of-the-day line to the template and ran /add-dates for September, chosen because it has 30 days; the output used September, included the new line, and stopped at 30 entries, showing the parameters are dynamic.

2026-06-03 · open the original ↗

• The pattern day trading rule was eliminated effective June 4th. Scott Sheridan says he pursued this change with regulators since 2018; the rule had capped accounts under $25,000 to three stock or options day trades per rolling five-business-day period (futures and crypto were exempt).
• Violating the old rule carried escalating, severe penalties. A first violation used to freeze an account for three months, later a year; roughly three years ago it changed to a one-strike warning followed by a lifetime trading ban unless the account was raised above $25,000.
• The change removes trade-count limits but not risk limits. Under the old rule, traders sometimes had to hold a profitable position overnight because they'd used their three day trades; now accounts under $25,000 can trade as often as capital allows, though available buying power and leverage rules are unchanged.

2026-06-03 · open the original ↗

• Exchange data fees are an unregulated cost retail traders should scrutinize. He argues exchanges are for-profit and can price data however they want, that most brokerages absorb the cost but some don't, and advises negotiating fee removal or switching firms if a broker passes data charges through to you.

2026-06-03 · open the original ↗

• One speaker rejects the claim that people won't need to read code. They say they "micromanage because I'm sculpting": nothing can be specified fully enough for AI to reproduce everything they meant implicitly. The other replies that a full specification would just be a programming language; the first cites a ZoomIt video trimmer, where they only knew what they wanted after rejecting what the AI first made.
• Early-career developers get worse results by leaving gaps to chance. The other says dictation gives them a boost through more words, context and specificity, while early-career people leave "so many things up to the dice roll". The clip ends on the question of why anyone would then spend four years at college; it does not touch on checking references, the subject of the full episode with this title.

2026-06-03 · open the original ↗

• Mark's Ref Checker found hallucinated references across a top AI conference. He says it scanned the 5,300 ICLR papers and found hundreds of hallucinated references, and that he disclosed the 12 papers with five or more to the program chairs; he expects more after refining the flagging. He says a company had already sampled 300 submissions in December and found several dozen, after which ICLR desk-rejected papers, yet the practice continues.
• Most flagged authors admitted fault; some argued or asked for silence. Of 10 responses, nine admitted error (AI formatting the bibliography, a student compiling it) and said they would notify ICLR, and one withdrew its paper. One claimed the references were unpublished but could point to no pre-publication copy; another argued five was arbitrary since it was 4% of their references; a young lab asked not to be named, and Mark is publishing only anonymous statistics.
• Hanselman likens leaning on AI to sending a forklift to the gym. When he raises "system three" thinking, Mark names the paper's term, "cognitive surrender". Hanselman says he feels it in his vibe-coded Windows tools, yet argues against a LinkedIn critic that less friction lets him find out whether an idea has legs.
• An AI cannot infer requirements nobody gave it. A count bug on Hanselman's Tiny Tool Town (13 of 316 tools) prompts Mark, playing provocateur, to ask whether AI should have known; Mark concludes it can't, and Hanselman says a model that already knew every answer would leave the web stuck where it is. The site itself is static, taking submissions through GitHub issue forms triaged by an agentic workflow that acts alone only at 90% confidence.
• Mark uses AI to write prose from his own bullets. He says Copilot writes well over half of his LinkedIn posts from points he sets out in order, which Hanselman had not noticed; Hanselman asks whether he has any reason to be offended if he can't tell. Mark had asked Hanselman for an analogy of a tool useful only to an expert and they settled on a lathe, agreeing the harm from AI is slower than a lost finger: "a slow rot" of the brain.

2026-06-03 · open the original ↗

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

2026-06-03 · open the original ↗

• Positive drift justifies taking market risk for most people. One host defines it as the extra return a risk-taker expects over the risk-free rate, and argues everyone should act on it unless they are very conservative or on fixed income, while keeping some money in cash as an allocation choice. He extends it to careers: if you are not getting better and earning more over time, something is wrong.
• Viewer trade ideas favor selling premium against stretched names. Netflix fell from 107 before earnings to 82; the host prefers selling July 75 puts (~$1.80, delta ~25, outside the expected move) to the 80s. Other ideas: a July 725/730 SMH call spread for ~$1.30 with SMH ~634, and a $5-wide at-the-money South Korea ETF put spread capped at $2.50 because its options are illiquid.
• SOXS covered call is pitched as an asymmetric upside bet. With the 3x leveraged ETF at ~$4.98, buy stock and sell July 7 calls for ~$0.65, a ~$4.30 net cost, ~60% probability of profit, IVR 105 and a $2 expected move. The host calls it his favorite and is long; he concedes the fund decays over time and will never return to its $228 52-week high.
• Roll or adjust by IVR and days left, never by waiting. With IVR high and over 30 days left, adjust within the same cycle; under 24 days, roll to the next cycle and recenter (24–30 is a judgment call). If IVR drops below 30, consider closing undefined-risk trades, while defined-risk trades can be held longer. A covered call deep in the money should be closed and replaced with a short out-of-the-money put.
• Career growth comes from volunteering, not from saying no. The hosts list quick decisions, problem solving, being visible, going beyond your title, saying yes to challenges, treating the role as your own business and making colleagues stronger. They split on refusals: one says a single "no" means he never asks again, another that nervous people turn down work they could do.

2026-06-03 · open the original ↗

• He argues Ireland's historical refusal to join NATO stemmed from British mistreatment, not pro-Russian sympathy. He says Ireland declined a NATO invitation outright out of resentment toward Britain, but that Russia's invasion of Ukraine has pushed the Taoiseach to weigh raising military spending to NATO levels and possibly eventual membership, ending a 69-year neutral stance.
• He says Ireland's continued exports to Russia are legal and require no justification. He states there's no law barring Ireland from exporting goods to Russia, even while he says he wants Russia to fail in Ukraine.
• He proposes any sanctions on Rusal's Aughinish refinery should include compensation for lost trade. He argues that if sanctions are imposed, they should be contingent on letting Aughinish's workers or owner sell their refined alumina to other buyers at a higher price to offset the damage.

2026-06-03 · open the original ↗

• Russia depends on an Irish refinery for roughly 40% of its alumina, a precursor used in military aluminum. He names the source as Aughinish Alumina in Co. Limerick, owned by Russian firm Rusal United, which ships in bauxite from Guinea, refines it to alumina, then sends it to Siberia to be smelted into aluminum used in Russian drones and fighter jets.
• Taoiseach Micheál Martin has resisted sanctioning the refinery, citing its roughly 2,000 local jobs. The speaker notes this sits alongside Ireland condemning the 2022 invasion, accepting over 250,000 Ukrainian refugees, sending non-lethal military aid, and backing broad EU sanctions on Russia.
• He attributes Ireland's WWII neutrality (1939-1945) to unresolved grievances against Britain. He cites the crushing of the 1916 Easter Rebellion, the 1919-1921 war for independence, and Ireland's 1922 independence leaving six counties still under British control as the source of lasting distrust toward the UK.

2026-06-03 · open the original ↗

• The best AI-services markets share four specific traits. The speaker, presenting a YC playbook, argues founders should target work that is already outsourced (low trust), breakable into automatable steps (low judgment at the task level), genuinely hard (a high intelligence threshold), and regulated — since regulation raises the bar and the moat rather than just adding friction; examples cited include tax, audit, insurance, mortgages and healthcare.
• Founding teams need domain fluency, model fluency, and operational rigor — the product is the operation. He points to General Legal Team, an AI-native law firm YC backed, whose founders combine law-firm experience at Cooley and Fenwick with technical leadership from CaseText and have built shift-based staffing to cut cycle times, calling variance in output the "existential problem" that causes customers to fire a vendor faster than slowness or price ever would.
• New AI-services founders should deliberately cap their first pilot customers. He warns of an "early demand trap" where easy early sign-ups overwhelm delivery capacity before the product can scale, and advises pricing on outcomes or per-unit (per return, per claim, per loan) rather than cost-plus or straight-line undercutting, since services compete against the cost of labor, not other software.
• The core financial bet is "AI operating leverage" pushing services margins toward software's. He argues traditional services firms cap out near 30% margins while AI-driven ones can approach 50%+ as the product lowers cost of goods sold (models, hosting, humans), on markets two to three times bigger than software's; he also argues buying an existing services business to bolt AI onto rarely works, since legacy hiring and performance expectations don't change with it.

2026-06-03 · open the original ↗

• The goal is a meal in every town in the state. Laudati and his wife are eating their way through New Jersey, which has 564 official towns, against Massachusetts' 351, where a former coworker, Chris Bowen, and his wife started the idea. About three years in, they have eaten in 287 towns, including 87 visited before the challenge began. Some towns have no restaurant at all, and one is only a members-only golf course with about six homes.
• Video editing, not visiting, is the bottleneck. Each 60–90 second video takes him 10–12 hours, so he averages about one a month. Post 129 was the latest, with roughly 150 towns still unpublished; Giard calculates one a month would take 48 years. Since going public about a year ago, one early video drew about 70,000 views when they had 50 followers. They now have about 20,000 followers across Facebook and Instagram, and town videos usually reach around 100,000 views.
• An AI coding agent built the interactive map; he wrote no code. Laudati says it took two or three hours on a Friday night with GitHub Copilot, using Claude, starting in plan mode, where he rejected technologies he did not want. The result is a static React site on GitHub Pages using a mapping library he believes is Leaflet. Copilot made four or five tries before finding a source for town boundary polygons. He overrode its suggestion of Cosmos DB, which he says is expensive, in favour of a JSON file, and unprompted it flagged the oversized logo as the page's biggest performance cost.
• Natural language is the newest abstraction, but it needs programmer-level specificity. Laudati argues English has become "the ultimate programming abstraction" over machine code. He says results go "off the rails" unless prompts are precise. His example is the AI-built script that turns his long-kept Excel tracking sheet into the site's JSON, run by a GitHub Action whenever he pushes the sheet.
• AI should build deterministic tools, not generate the content. AI-generated images in early videos drew criticism from artistic commenters, so he cut back; he says 95% of what they publish is him talking over their own photos and footage. He fed a three-page write-up of his manual editing steps to Copilot, which produced a 10-phase plan. Of the seven tools he has scaffolded, the finished one is a command-line map-animation generator that makes each opening zoom in 15–20 seconds and can batch all 564 towns. He says the tools are on GitHub under what he thinks is the MIT license.

2026-06-03 · open the original ↗

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

2026-06-03 · open the original ↗

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

2026-06-02 · open the original ↗

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

2026-06-02 · open the original ↗

• He keeps most of his trading account in uninvested free cash. That's so money is ready for a spur-of-the-moment trade, avoiding money funds, T-bills, or Treasury ETFs. For the roughly 20–35% of the account he rarely touches, he does park it in a Treasury, Treasury ETF, or money fund to earn a risk-free rate just above or below 4%.

2026-06-02 · open the original ↗

• He argues the drone that entered Romania was an errant strike meant for Ukraine, not a deliberate NATO provocation. He says Russia's military lacks the strength to challenge NATO's 32 members directly and predicts Article 5 will not be invoked, but that NATO might respond with "harmless" air patrols along its entire Russian border.
• He predicts such patrols would strain Russia's military past its limits, since its forces are already committed to Ukraine. He cites Ukraine regaining about 100 square miles of territory in May alone, and argues that if Russia had to divert resources to counter NATO patrols, Ukrainian gains would accelerate further, forcing Russia to choose between the two fronts.
• He dismisses the nuclear-escalation threat as an empty pattern. He notes Russia has threatened to use nuclear weapons since February 2022 without following through despite repeated "red lines" being crossed, and questions whether Russia even knows how many of its own warheads would function or reach their targets.

2026-06-02 · open the original ↗

• The central bank's balance sheet trilemma is the frame: pick two of three. A small balance sheet, low volatility in short rates, and limited market intervention — any two are achievable, never all three. The Fed's balance sheet ran from about $800B to $9.9T and now sits near $6.704T, of which $4.4T is Treasuries and $1.9T mortgage-backed. The MBS is left to run off rather than sold, to avoid pushing mortgage rates up, so the $4.4T is what is actually in play. Because reserves back key bank liabilities, shrinking that buffer builds pressure in money markets and spreads, which is what degrades rate control.
• The confirmation hearings asked the wrong question, because Warsh's agenda is the balance sheet, not the policy rate. Every senator asked whether he would cut rates for the president. His actual position is that the balance sheet must come down substantially — he is notably unclear on how, since selling into the open market is not straightforward — on the theory that a smaller structural footprint restores the Fed's ability to drive short rates. The framework's estimate is that a roughly $4T balance sheet is equivalent to suppressing long-term rates by about 200 basis points, so unwinding it implies materially higher long yields.
• The duration mismatch is the risk he keeps returning to, and it now runs on both sides of the government. A large balance sheet finances long-duration holdings with overnight liabilities. Treasury has adopted the same posture fiscally, moving issuance from roughly 21% short-dated under Yellen to 84% under Bessent. Financing long-term liabilities with short-term ones is the one thing every bank in history has been warned against, and it loses eventually. Combined with Miran's stated willingness to force short rates down 250–300bp regardless of conditions, the implication is a deliberately steepened curve with the long end left to find its own level.

2026-06-02 · open the original ↗

• The Comet Q is a USB-C KVM powered by its target. The host shows the GL.iNet Comet Q KVM controlling a Linux mini PC, an iPad and an Android device; it joins local Wi-Fi, serves a browser interface, and supports Tailscale or ZeroTier for remote access. A second USB-C port passes power through to the controlled device.
• It is crowdfunding on Kickstarter at an early-bird price. It starts at $69 on the super-early-bird tier; the host gives no other pricing or shipping date.

2026-06-02 · open the original ↗

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

2026-06-02 · open the original ↗

• Astrocade's product is the games platform, not the AI. Kurenkov, founding AI lead at Astrocade (which the host says has raised $68 million), describes it as a "TikTok of games" where anyone vibe-codes, publishes and plays casual mobile games free, with tokens currently paid from VC money. He says it has over 20 million engaged users after growth in the last 6–8 months, that only a small share of users make games, and that this is why it fits his earlier view that AI works as a feature rather than a product, unlike chatbot wrappers; he cites Harvey as a domain-specific success despite Anthropic launching Claude for legal.
• There is no secret sauce, only how the standard parts are mixed. He says the stack is an agent, tools and an LLM, and the hard part is benchmarking harness and prompts when a request like a Tetris-merge-puzzle game has countless valid answers. His hard-won lesson is that scaffolding built in 2023–2024, before models could write whole games, became a limit once vibe coding worked in 2025, so systems must be built not to date when models improve every few months; Astrocade adopted Claude Code company-wide around June 2025.
• Consumer and user-generated content are harder but the moat is creators. He says B2C took over 2 years to pay off but is the only route to Instagram-scale hundreds of millions of users, and that the lasting moat is a skilled creator community (a Discord of about 50,000) supported by a $10 million fund paying creators per play, the Astro Academy cohort course, and follower graphs. He agrees with Sequoia's David Cahn, as quoted by the host, that many users are women aged 20 to 40, saying that reflects existing casual mobile gaming.
• Recommending games is harder than recommending video. Games often have no end point, so a minute of play may mean enjoyment or confusion, unlike a TikTok watched past 10 seconds; and small mechanical tweaks change whether someone likes a game, so genre is a weak similarity signal. He says existing recommender work largely does not cover this domain.
• Household chore robots are arriving far sooner than he expected. Kurenkov says humanoid startups such as 1X and Figure, applying LLM methods via vision-language-action models, have progressed more than he expected, cutting his estimate from a decade two years ago, though he adds demos deserve scepticism and a $50,000 dishwashing robot may not make economic sense. He calls himself an artificial-superintelligence sceptic because physics, computation and the need for experiments limit what AI can solve, says AI-powered scamming is among today's most harmful real risks, and argues AGI should be treated as a spectrum of generality that arguably began before GPT-3.

2026-06-02 · open the original ↗

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

2026-06-01 · open the original ↗

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

2026-06-01 · open the original ↗

• He argues Germany could rapidly build one of the world's most powerful militaries if it chose to, pointing to its post-WWI recovery within 20 years as precedent. He says the only reason it hasn't already is that the US and NATO have historically discouraged it, a stance he says is shifting because of Russia's war on Ukraine.
• Germany's artillery ammunition production rose roughly 15-fold, from 70,000 to 1.1 million rounds a year, closing in on Russia's 2.2 million. He predicts Germany could match or exceed Russian production within a year, backed by a GDP he cites at $5.3 trillion in 2026 versus Russia's $2.5 trillion.
• He predicts Germany will overtake Russia and France to become the world's second-largest arms exporter behind the US. He currently ranks Russia third and Germany fourth in arms exports, and highlights the Leopard 2 tank's 120mm gun (also used on the US Abrams) and the Type 212 submarine as flagship German exports.

2026-06-01 · open the original ↗

• Stocks now scalp better than options because markets are tighter and free. The hosts point to penny-wide markets in names like Apple versus 50-cent-to-dollar-wide option markets on expensive names, plus the disappearance of stock commissions at online brokerages, as what makes stock scalping attractive.
• The top 50-100 most liquid stocks can be traded nearly around the clock. Unlike options, which stop trading at the market close, these names can be traded after hours, though liquidity thins out compared to the regular session.
• Scalp targets should be a percentage of the daily expected move, not a fixed dollar figure. The rule of thumb given is aiming for roughly 10% of that move — $0.50 on a $5 expected move in Apple, or $3 on a $30 expected move in Micron — rather than chasing an arbitrary target.

2026-06-01 · open the original ↗

• The 1973 War Powers Resolution's core enforcement tool — a concurrent resolution letting Congress order troop withdrawal without presidential signature — was a "legislative veto" that the Constitution's presentment requirement made vulnerable from the start. Congress passed it, and overrode Nixon's veto, only because Nixon was politically weakened by Watergate at the time.
• The Supreme Court's 1983 decision in INS v. Chadha, written by Chief Justice Warren Burger, struck down the legislative veto as unconstitutional for bypassing presentment. This gutted the War Powers Resolution's enforcement teeth: Congress can now only force withdrawal via a joint resolution, itself subject to presidential veto and a two-thirds override.
• Both Obama and Trump have exploited the resulting gap using different tactics. Obama argued the 2011 Libya operation didn't meet the "hostilities" threshold requiring authorization since it was NATO-led with no U.S. ground troops; Trump has sent "hostilities terminated" letters to reset the 60-day clock on Iran while relying on a veto wall to block any resolution that manages to pass.
• The author argues Congress could still build an authorization requirement that doesn't depend on beating a presidential veto, but has chosen not to try. Absent that, or the Court revisiting Chadha, the piece names the bond market — Treasury yields near 5% pressuring the White House ahead of the midterms — as the more effective real-world check on presidential war-making.

2026-06-01 · open the original ↗

• A call he made three years ago is now being argued openly in Chinese state-tolerated media. NetEase — one of China's largest platforms, not a fringe outlet, operating in a strictly censored environment where publication implies approval — ran a piece in December 2025 headlined that if Russia collapses, 7.1 million square kilometres must not be lost, invoking the 1858 and 1860 treaties as unequal. The supporting argument is that Russia's GDP is smaller than a single Chinese province with fewer than 50,000 troops remaining in the Far East. His reading is that this is not a war scenario but an abandonment one: the economic penetration is already complete, with business entities across roughly a million square kilometres paying Chinese taxes and running Chinese systems, and the Urals as the only defensible boundary.
• The binding rare-earth constraint is human capital, not geology — and that is the harder gap to close. China operates more than 40 specialist rare-earth laboratories plus at least 11 universities enrolling over 500 students a year in rare-earth degree programmes, while the cited researcher could not identify a single undergraduate rare-earth degree anywhere outside China. Beijing has tightened access further: restricting contact between industry professionals and foreigners, ordering some technicians to surrender passports, and earlier barring rare-earth scientists from emailing, phoning, writing to or travelling to the West. A Beijing research centre developed refining capable of 50,000 metric tons a year of highly processed rare earths — five times Lynas's output — against China's ~90% share of processed rare earths and magnets.
• His generalizable point is that everything now resolves to time, and the estimates keep lengthening. Buying the equity of a miner does not produce metallurgists. The Gulf is put at five years to return, with some saying 2032; rare earths he puts at twenty for the US. When the framework launched three years ago it argued that things which break can no longer be fixed in days but in years — and the revision since is that the real numbers are ten and twenty. His instruction is to stop litigating how and why and concentrate on what comes next, because that is the tradeable part.

2026-06-01 · open the original ↗

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

2026-06-01 · open the original ↗

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

2026-06-01 · open the original ↗

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

2026-06-01 · open the original ↗

• She shelved her AAVE corpus because communities should own their language. She began the dataset from an undergraduate thesis arguing that African American Vernacular English is a tie binding African languages and English, and worked on it at Stanford, but stopped after a peer who worked on the since-removed AfroBERT model argued that giving powerful institutions the ability to mimic a group's language can be used against people who already feel unsafe. Her position is that she, as a self-described "guest in an AAVE home", can provide the bridge, while an AAVE language bot should be built by people of that community (she names American descendants of slaves) if they choose, and some have contacted her about doing so.

2026-05-31 · open the original ↗

• One distinction drives every difference in scale: 1973 was selective, 2026 is indiscriminate. The embargo was a political weapon aimed at the US, Netherlands and Portugal, so re-routing through non-embargoing suppliers worked. A chokepoint seizure affects every nation moving oil or LNG through the strait regardless of alignment. Supply removed runs ~4.5 mb/d (9% of global) against ~15.8–20 mb/d (20%); commodities hit go from oil alone to oil plus LNG, fertilizer, helium and refined products; bypass capacity is near-zero, with pipelines covering only ~2.6 of 20 mb/d; spare capacity was significant then and is effectively nil now; and strategic reserves did not exist in 1973 — the SPR was created in response — while today's are sized for a 1973-scale event.
• Bonds are the most consequentially different asset class, because the safe-haven logic inverted immediately. In 1973 flight-to-quality held. On 5 March 2026 the 10-year spiked to 4.13% and later peaked at 4.68% — investors sold Treasuries rather than buying them, because oil-driven inflation fear overrode safety demand from day one, and rate hikes rather than cuts were priced. A structurally new channel accompanies it: Japan holds $1.225 trillion of Treasuries and its own 10-year reached the highest level since 1999, so rising domestic yields pressure Japanese institutions to sell foreign holdings and push US yields higher independently of anything the Fed does.
• Housing diverges most sharply, and in the direction that surprises people. In 1973 nominal prices rose — new home prices up 7.5% during the recession and the average US home from $36,600 to $40,900 by Q1 1975 — because households store value in hard assets during inflation, while real prices stayed flat for a decade. In 2026 the market froze rather than crashed: the lock-in effect, owners holding 3–4% mortgages refusing to sell at 6%+, suppresses supply so prices hold while volume collapses. The result is illiquidity at elevated prices rather than clearing at a lower level — a 2020s structural feature with no 1973 precedent. The UK crashed in the 1970s on different mortgage structure and bank leverage; variable-rate economies carry that risk today.
• The timeline lesson is that the second shock, not the first, sets the clock — and the institutional legacy outlasts both. Normalization after 1973 took roughly ten years rather than five, because the 1979 Iranian Revolution reset the trajectory just as rates began easing; mortgage rates peaked at 18.63% in October 1981, eight years after the initial shock. Whether 2026 repeats that path depends entirely on whether a second shock lands. Meanwhile 1973 was a single-vector energy disruption while 2026 runs through food, energy, technology and security at once — roughly a third of global fertilizer trade and the helium feeding Korean and Taiwanese semiconductor manufacturing both transit Hormuz, a vector with no 1973 analogue. The IEA, the SPR and CAFE standards were all born from 1973; whatever is born from 2026 will still be reshaping markets in 2036.

2026-05-31 · open the original ↗

• He is comparing this to a shock he personally traded through, which is why the piece exists. Hired at Merrill in August 1975 and licensed in early 1976, the aftermath of the first embargo was the start of his career — he bought his first home at 11.25% and recalls the price at $22,000. That vantage is what makes him treat the ten-year arc rather than the acute shock as the real subject.
• His scale verdict is that 2026 is ten to twenty times worse than 1973 on both magnitude and vectors. Oil went from $3 to $12 in 1973 on 4.5 mb/d removed, 9% of supply, aimed at three countries. This time it is 15.8–20 mb/d and 20%, hitting every nation, across oil, LNG, fertilizer, helium and refined products simultaneously. He dismisses the pipeline workarounds as arithmetic — Oman, Saudi and prospective Turkish routes cover 2.6 million of 20 million barrels a day — notes OPEC's spare capacity is effectively out of the picture, and points out the SPR exists only because of 1973 and was sized for that event.
• His forward call is peace, followed by a price problem that peace itself creates. He expects an agreement because nobody destroys the world economy for appearances, but relays an oil-industry view that crude runs to $160–180 when China returns to the market — China having been largely absent, taking barrels by rail from Iran and Russia. His framing throughout is the one he repeats: stop litigating how and why, work on what's next. On duration he will only say normalization took ten years last time, and he expects at least that again.

2026-05-31 · open the original ↗

• Risk assets hit new records on hopes for a US-Iran deal even as fighting continued and Israel intensified attacks on Lebanon. The Nasdaq reached a 16.1% YTD return, the S&P 500 10.7%, and Japan's Nikkei surged 31.8% YTD in a single month, while the DXY stayed range-bound — a rally he calls equity-specific rather than dollar-driven.
• A widening growth divergence saw the US economy revised down while Canada and France slipped into contraction. Q1 US GDP growth was revised to 1.6% (domestic demand held at 2.4%), Canada entered a technical recession on a 0.1% annualized Q1 contraction, and France also contracted, even as booming US equities masked a widening Main Street-Wall Street gap of declining labor income shares and softening consumer sentiment.
• US PCE inflation marked 62 consecutive months above the Fed's target just as leadership transitions from Powell to incoming Chair Warsh, who inherits a divided FOMC. Headline PCE climbed to 3.8% and core to 3.3%, with sticky New York Fed inflation expectations compounding a policy challenge further constrained by US government debt near 100% of GDP.

2026-05-31 · open the original ↗

• Scaling has diminishing returns once the added data is bad. Henry describes her 2023 NeurIPS work (accepted to a workshop, which she calls a top 5% paper) as a challenge to the Kaplan et al. scaling idea from OpenAI that adding data, compute and parameters extends capability indefinitely. She borrows the economics of diminishing returns, likening it to a government printing money until it loses value.
• Undeduplicated data makes models forgetful and prone to overfitting. She argues that the largest training sets are so big they are nearly impossible to deduplicate, and that this makes models more forgetful, less intelligent and more overfit over time.- You can't compute your way out of bad data. In her account, a machine with 10,000 GPU clusters trained on bad data will do worse than less compute on less but better-quality data. She says the field is now coming round to this, pointing to 70 billion or 120 billion parameter models that consistently outperform models of one trillion parameters or more.

2026-05-30 · open the original ↗

• Trump backed down when Xi stood up to him. He characterizes Xi Jinping as effectively saying he will do whatever he wants with Taiwan and the US can do nothing about it, and Trump saying nothing in reply; the piece frames this as the schoolyard bully getting stood up to and folding.
• The reversal contradicts years of Republican China messaging. He argues Trump and the Republican Party made China the villain for years, yet after what he describes as a one-hour meeting with Xi, Trump came home willing to give China farmland and more — in his words, "suddenly he's half Chinese."

2026-05-30 · open the original ↗

• Waiting for a checkpoint to evaluate a model wastes too much time. Henry says that knowing whether a training run is working takes anywhere from 2 hours for a quick, small-model experiment to days for a large one. A big company like Microsoft could absorb that time and compute, but a startup with a demo due by Friday cannot.
• Monte Carlo simulation on early reward trajectories predicts checkpoint performance. Henry's approach, part of her company's RIGS petroleum-engineering foundation models, takes the reward computed during reinforcement learning (about an hour in) and runs Monte Carlo simulations to forecast where the model will be at the checkpoint. That lets her kill a run heading the wrong way, and run several ablation variants at once (say, weighting data towards equations or towards oil and gas) and keep the winner.
• A gap between forecast and checkpoint result exposes reward hacking. Models learn the least they must do to get a good reward. Her example is guessing the five most common equations to score on retrieval at three and five, which collapses once more equations are added. If the simulation predicts a weak result but the checkpoint looks excellent, she takes that as a sign of reward hacking.
• Exam-style evals miss how users actually ask questions. The model may do well on petroleum engineering exam questions (hydrostatic pressure for a given well and tool), but real users say "tell me about this well" or just "decline curve". Henry says she injects such questions into the training environment mid-run so she can adjust before the run ends. The host attributes the framing of four systematic failures that make current evaluation of agentic systems inadequate to a paper she recently released.

2026-05-29 · open the original ↗

• Panama is not a player in this dispute — it is the terrain, and the paper says so plainly. The actual contest is the US against China over who controls the infrastructure layer at the world's second most important maritime chokepoint, through which roughly $270B of annual trade and eventually 80% of Chinese oil imports route. Panama's "decisions" — the Supreme Court ruling, the port seizure, the operator selection — were forced moves executed through domestic institutions and dressed in the language of sovereign choice.
• The sequence shows the legal instrument doing the work and then poisoning the asset. Trump's December 2024 claim, CK Hutchison's $22.8B ports sale to a BlackRock-led consortium, the 30 January 2026 annulment of the Balboa and Cristóbal concessions as unconstitutional, seizure on 23 February with Maersk and MSC on interim contracts — then Chinese retaliation, and on 3 March the consortium closing the $23B deal without the Panama assets, which had become too legally encumbered to include cleanly. By March, 93 Panama-flagged vessels had been detained in China, a 300%+ jump from February, and six Latin American countries had issued a joint condemnation.
• China's playbook was effective tactically and self-defeating strategically. Vessel detentions were the most precise instrument, because the flag registry of roughly 8,000 vessels is sovereign revenue independent of canal tolls — so the pain lands on the treasury, not the operators, and by April Mulino had turned conciliatory. The $2B+ Hutchison arbitration is the long-duration lever, running independent of the diplomatic track and impossible for Panama to de-escalate alone. But the bluntness produced the six-country condemnation that handed Washington a clean coercion narrative across the hemisphere, and COSCO's demand for a majority stake in the BlackRock deal — spoiling tactic or miscalculation — nearly collapsed the transaction while making state intentions explicit.
• The durable finding is about ambiguity, and it produces a new candidate concept. A Toll State's premium is a function of the credible ambiguity of its alignment, not its geography. Panama retains the geography and has lost the ambiguity, having demonstrated that under simultaneous pressure it bends toward whoever holds greater direct military and financial leverage — and every future counterparty has updated its model. The paper proposes "Forced Chokepoint Exit": a Toll State compelled to transfer operational control of its chokepoint to a patron-aligned operator through its own courts or legislature, preserving the formal appearance of sovereignty while the multi-vector premium collapses anyway. It is explicitly marked CANDIDATE with one instance, pending a second under a standing two-instance rule, with Djibouti, Oman and Sri Lanka named as watch cases.

2026-05-29 · open the original ↗

• His organising frame is physical China against paper America, and Panama is the clean test case. Every move Panama made was paper: a Supreme Court finding of unconstitutionality, a seizure, a reassignment of operators. Every Chinese response was physical: detaining ships, withdrawing COSCO from Balboa, freezing all construction in the country, and refusing to build new vessels destined to fly the Panama flag. He argues the politics of who said what matters less than watching which layer each side can actually act on.
• The physical leverage he points to is structural rather than tactical. China holds seven of the world's top ten ports, charters most of the world's ships, and builds essentially all of them. Panama earns roughly $860M a year from its flag registry, so detaining Panama-flagged vessels — which may be Greek- or Maersk-owned and merely registered there — targets sovereign income rather than canal operations. And the shipbuilding refusal reaches years into the future, since anyone ordering a Chinese-built hull intended for the Panama registry simply cannot have one.
• He distinguishes two kinds of toll state, and Panama is the wrong kind. Some can defend themselves — his example is Poland as a node state that few would take on. Panama holds a genuine $270B-a-year chokepoint with no military or financial independence, so when both great powers squeezed simultaneously the ambiguity that gives a toll state its value simply evaporated. His scorecard is that the US won and China got nothing, because the detentions backfired into regional alignment against Beijing — but he flags that Wang Yi attending the New York meeting personally signals how seriously China is treating it.

2026-05-29 · open the original ↗

• Liquidity, not the underlying thesis, is his top screen for a trade. He says illiquid markets "eat you alive" on both entry and exit, and sets a minimum bar of a couple million shares of daily volume with bid-ask spreads only a few pennies wide — a nickel to a dime, or mid-price, in options — below which he'd reconsider trading that underlying at all.

2026-05-29 · open the original ↗

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

2026-05-29 · open the original ↗

• US sanctions relief on Russian oil is helping prices fall. He notes the administration temporarily lifted a number of Russian oil sanctions, letting Russian crude sell freely on the world market again, which he says is contributing to lower prices.
• He predicts Iran cannot sustain the shutdown much longer. Calling Iran "starving," he argues it cannot endure a full closure of trade for another five or six months, and expects a deal — via US or Iranian capitulation — to eventually reopen the Strait of Hormuz.
• He predicts countries will permanently reroute supply chains away from Hormuz. Even after reopening, he expects the altered, diversified energy supply chain to be a net positive for the world economy, though he still expects sharp price swings if hostilities resume.

2026-05-29 · open the original ↗

• Non-Gulf producers have sharply ramped output to cover the deficit. Guyana is near 1 million barrels/day (projected 1.7 million by 2030), Brazil rose from roughly 2 million barrels/day in 2025 to about 4.5 million by March 2026, and US exports climbed about 33%, from roughly 3.9 million to 5.2–5.3 million barrels/day.
• Saudi Arabia and the UAE built pipelines specifically to bypass the Strait of Hormuz. Saudi Arabia's 1,200 km East-West pipeline can move about 5 million barrels/day to its Yanbu port on the Red Sea, while the UAE's Fujairah pipeline moves about 1 million barrels/day, with a second one under construction (about 50% complete, targeting 2.5 million barrels/day).
• He predicts similar pipeline-building speed could be repeated in the Persian Gulf. He cites the US building the 1,400-mile "Big Inch" pipeline from Texas to New Jersey in under a year during World War II as precedent that infrastructure timelines can collapse from years to months.

2026-05-29 · open the original ↗

• Predicted $150–300 oil and a built-in recession have not materialized. He contrasts the muted impact of Iran's closure — which cuts off roughly 20 million barrels/day, about 20% of world supply — with the 1973 Saudi oil embargo and the 1979 Iranian Revolution, both of which caused severe global shortages.
• Strategic Petroleum Reserves are buying major economies time to reroute supply. He credits SPRs in the US, Europe, Canada, China, Japan and South Korea with cushioning the shock while imports are redirected; he cites China's claim that curbed imports could stretch its reserves through 2026.
• Venezuela's oil exports have surged since the US took over its industry after Maduro's removal. He notes Venezuela holds the world's largest crude reserves at 300 billion barrels, more than Saudi Arabia, and has hit a record of over 1.2 million barrels/day exported, its highest since 2018.

2026-05-29 · open the original ↗

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

2026-05-29 · open the original ↗

• YC's Harj rejected their EdTech idea, not their candidacy, and bet on the founders instead. Varun, GigaML's co-founder, says Harj told them "you guys are really good engineers, pick something else" without asking about their idea or traction, pushing them to abandon EdTech and pivot within a month of starting the batch based on their LLM research background rather than a specific pitch.
• A GitHub usage pattern, not a market thesis, drove the pivot into customer support. After their initial fine-tuning product proved a bad market — good only for cutting inference cost or serving security-sensitive buyers that are hard for engineers to sell to — Varun says they noticed customer support and coding were the two use cases growing fastest among their fine-tuning customers, and Zepto became their first customer once they switched.
• An eight-person team beat a roughly 400-person, well-funded rival for DoorDash's contract. Varun credits a YC-network "unfair advantage" (an introduction from Garry Tan) and a clean three-month pilot with no downtime, and says winning DoorDash proved to him that building a superior product beats building a sales team — a lesson he calls an "arbitrage" most competitors were missing.
• Coding agents let GigaML run with roughly one-seventh the engineers it would otherwise need. Varun estimates the company would need six to seven times its current engineering headcount without tools like Claude Code, and says the deeper benefit is avoiding context-switching and context transfer between people, not just cost; the company's interview process still has candidates write code with AI, then edit it without AI, to confirm they understand it.

2026-05-29 · open the original ↗

• Production agents need a control plane between apps and infrastructure. Bajaj says TrueFoundry, which the host says has raised over twenty million dollars, sits between the application layer and the model and infrastructure layer so one place can connect, observe and govern. He says enterprises first treated this as a thin LLM proxy they could build in-house. Moving agents to production showed they also needed observability, governance and a policy engine across many teams and cost centres. The host's research puts usage at a trillion tokens a day, which Bajaj accepts as "a big number", and names Nvidia and Siemens as clients.
• A model gateway centralizes keys, limits, fallbacks and guardrails. Developers get access to the gateway rather than direct API keys. Bajaj says the organization can set rate limits (100 calls an hour), budgets (no dev environment over $2,000 a day), regional fallback if an OpenAI region is slow, PII guardrails and prompt-injection blocking, with a full trace of every call.
• MCP gateways must map identities across each tool's own authentication. Unlike a model's single API key, MCP servers expose sensitive read-write data (Slack, Gmail), and each implements its own auth. His example is an agent reachable by outside consumers getting only whitelisted read-only tools from seven servers, exposed as one virtual server. Bajaj says tool sprawl breaks nothing in theory; the cost is in debugging and access management. Where needed tools aren't known ahead of time, tool groups carry descriptions so the model can pick them at runtime.
• Most enterprises have no kill switch for their agents. Bajaj calls the agent gateway "the hard part". Agents take actions, and at best companies can observe them but not stop them. Agents are now built across sales, ITSM and pro-code platforms, and he says the next gateway must put all of them behind one governance layer.
• Start with a few big agent workflows, not many small ones. Bajaj's "five workflow framework" says starting small leaves 500 low-value agents, none worth building governance plumbing for. Instead he advises picking one or two of the five business-critical workflows, building the plumbing for them, then scaling. He says customers who did this have run such agents for a year and a half and confirms, when the host asks, that individual agents create value in the hundreds of millions of dollars.

2026-05-29 · open the original ↗

• Index rule changes force passive money into SpaceX fast. S&P and Nasdaq loosened seasoning periods and profitability requirements right before SpaceX's IPO, meaning VOO, SPY, and QQQ must buy shares within months (S&P) or as soon as 7 trading days after listing (Nasdaq-100), regardless of fund manager judgment.
• The rule changes track index-provider revenue, not "representativeness." Nasdaq and S&P earn licensing fees based on assets tracking their indexes, so excluding a $1.5 trillion company (even one losing billions annually) would be commercially self-defeating, and the changes apply only to megacap stocks rather than uniformly.
• Faster inclusion lines up with when SpaceX insiders can first sell. The S-1 permits certain insider sales shortly after trading begins, so the accelerated index timeline functions as engineered exit liquidity, with retail 401(k) and IRA holders serving as the forced buyers absorbing the risk.

2026-05-29 · open the original ↗

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

2026-05-28 · open the original ↗

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

2026-05-28 · open the original ↗

• A beach club's Ubiquiti point-to-point network is laid out sensibly. The host reads one antenna as a wireless ISP feed, with a newer pill-shaped NanoStation below it, and NanoBeam and NanoStation links fanning out to outbuildings; he singles out a single-gang wall-mount enclosure with a pole for the NanoBeam as a clean mount.
• Indoor-rated gear is used outdoors, which the host would avoid. He spots an indoor antenna and a non-outdoor camera on the site, the antenna sheltered under a roof; he guesses an old NanoStation M5 was left in place beside its replacement. He would not have used the indoor gear outdoors himself, but calls the install a nice job overall.

2026-05-28 · open the original ↗

• Inference speed will become a capability ceiling, not just a cost lever. Presenting his "speculative speculative decoding" work, Stanford grad student Tanishk argues that once performance scales with amount of thinking, tokens-per-second is peak intelligence delivered; his method parallelizes normally-sequential drafting and verification by predicting verification outcomes about 80-90% of the time, winning on both latency and throughput sampling Llama 3 70B on 4 H100s.
• Factoring the action proposal from the dynamics model lets a robot policy adapt at runtime. Google DeepMind's Stannis presents Diffusion Model Predictive Control, where multi-step diffusion models for action proposals and dynamics reduce compounding errors and simplify planning; he shows a simulated walker with a broken ankle recovers most of its performance just by re-adapting the dynamics model, without retraining the action proposal.
• A single new regularizer lets a world model avoid representational collapse with one hyperparameter. Isaac Ward, from Yann LeCun's group (whose $1.03 billion raise centers on world models), presents "LeJEPA," which adds a "SIGReg" term enforcing Gaussian-distributed latent embeddings; he reports it runs roughly 50 times faster than competing world models on a single GPU under 24GB of memory with just 15 million parameters, and can flag its own prediction errors when conditions shift.
• Ensembling and distillation, not bigger single models, are the efficient path when data — not compute — is the constraint. Presenting a paper co-led with Suhas and Percy Liang plus another collaborator, Ku shows that regularized and ensembled pre-training recipes follow clean power laws with measurable best-possible-loss asymptotes; a joint recipe yields roughly a 5x data-efficiency win, and distilling an eight-model, 2.4-billion-parameter ensemble into one 300-million-parameter model retains about 83% of the gain.

2026-05-28 · open the original ↗

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

2026-05-28 · open the original ↗

• Liquidity, not asset class, is what separates a tradeable market. He sets the bar at a couple million shares of daily volume with penny-wide stock spreads and nickel-to-dime-wide option spreads, saying anything wider warrants stepping back from the trade entirely.
• Every brokerage sets its own risk-management rules, traders should know upfront. He warns that liquidation triggers, forced-flat times, and excess-risk fees are non-standardized across firms, since regulators only set minimums and brokerages layer on their own stricter thresholds.
• Support quality is a real differentiator he warns has eroded. He recalls staffing thinkorswim's original trade desk with ex-floor traders decades ago, and cautions that some firms now route customers to AI agents, overseas contractors, or charge for the right to speak to a live person.
• Margin rates on trading accounts are negotiable, often by several points. He notes retail investors quoted 10–11% can often talk firms down to 7–8%, or lower on larger balances, while free cash sits at near-zero interest unless moved into treasuries or a money fund.