Souk's Notes - Unofficial Summaries

2026-07-09 · open the original ↗

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

2026-07-09 · open the original ↗

• Skills, agents and hooks turn a confident intern into a reliable worker. The presenter treats Claude Code as a new hire. A skill is the playbook, a plain-English file defining how a task is done; an agent is a teammate you delegate to, with its own persona and context; a hook is a rule enforced deterministically at a trigger point. Answering an attendee, the presenter calls the whole package in a folder (skills, agents, hooks, helpers, instructions) the "harness".
• Use hooks for what must happen, not for everything. The host adds that a skill's main failure mode is not firing, especially in a large system or with an overgrown CLAUDE.md, so a must-run task belongs in a hook. Hooks lose the model's judgment about when to act, though, so not everything should be one. The host's own use ties hooks to evals through traceability: every Snowflake query call triggers a hook that logs the SQL, a timestamp, an ID and the question's context.
• A skill and a critic agent produced and stress-tested the weekly report. In Claude Code Desktop on a fictional coffee-subscription app's data, /weekly-review followed a skill requiring a TL;DR, the one number that moved, one risk, one action, under 250 words, numbers only, and percentages to one decimal. The report traced a double-digit week-over-week drop to two days when paid social reported almost no activity. After the presenter edited the skill to open with a status light, the rerun rated it yellow. An "exec reviewer" agent playing a VP of growth then returned three likely questions, including whether the signups were "lost or just miscounted" and why organic peaked on those same days.
• A data-integrity hook refused to ship a report on broken data. The presenter changed one value from 772 to 900 so that activations exceeded signups. A Python hook checking for duplicate date-channel rows, activations above signups, and negative or malformed values then blocked the report from being written, with no prompt telling it to run. The presenter also points to the lab's open-source "AI analyst" repo as the fuller version of this harness, taking a business question to a deck or one-pager.
• Every analysis is a bet, and AI improves all three terms. The presenter's formula is impact = number of bets × hit rate × size of the win. On their account, the system gives more shots (the Monday hour becomes about five minutes), a higher hit rate through deeper slicing, and room for "venture bets" that were once too costly to take.

2026-07-09 · open the original ↗

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

2026-07-09 · open the original ↗

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

2026-07-09 · open the original ↗

• He describes childhood abuse — burned, beaten, threatened — for being white. He says that had he stayed in that town without traveling, he believes he might have grown up racist himself.
• Traveling from age 18 showed him the reverse dynamic elsewhere. He saw Black, Hispanic and other non-white people treated as lesser in other places, a contrast he says is what first made him understand the concept of white privilege — including moments like police singling him out as the "safe" person to address at a mixed-race gathering.
• He argues prejudice is universal, exploited by elites to divide people. His conclusion — presented as his own opinion, not reported fact — is that recognizing shared humanity across race and religion is what would let ordinary people build a "powerful force" against that manipulation.

2026-07-09 · open the original ↗

• He says he was harassed, robbed, or attacked almost daily from age 9 to 18. Growing up as a white kid in a mostly Black and Hispanic neighborhood, he describes burn marks from matches and being beaten and threatened, despite trying to signal sympathy and citing his own immigrant family (father arrived in 1960, mother's family around 1905-1906) as having no tie to slavery or oppression.
• A growth spurt around age 14-15 shifted him from fear to anger, he says. Reaching 6'3" and 200 lb, he stopped being scared and became confrontational, developing what he calls a charge against anyone displaying "thuggish," aggressive mannerisms regardless of race.
• He rejects the concept of "white privilege" as applied to his own upbringing. He says he would have refused to write about it for college essays given his experience of abuse, and credits later travel and curiosity, rather than his hometown environment, with keeping him from becoming racist.

2026-07-09 · open the original ↗

• A grid of users by day reveals patterns that aggregate metrics hide. He calls this a "dot plot": one row per user, one column per day, with a dot marking each day they performed a chosen value event. DAU/MAU graphs lump everyone together and can look like healthy growth even when individual users aren't getting value, whereas the grid surfaces patterns like distinct weekday-only and weekend-only usage cohorts.
• He traces the idea partly to a PayPal fraud-detection technique. In PayPal's early days, humans stared at visualized transaction graphs and could spot anomalous patterns worth investigating even without knowing what fraud looked like in advance, from Max Levchin — the same principle, he says, applies to spotting product problems in dot plots.
• A real YC example shows dot plots predicting churn before it happens. A company sold an $80,000-a-year, ten-seat contract, but a dot plot would have shown only three seats ever activated with sporadic use; the internal champion later left, the incoming employee canceled at renewal, and the warning was visible in usage data the founders never looked at that way.
• The two biggest misuses: wrong event, too coarse a time period. He warns against populating dot plots with vanity events like "opened the app," recommending true value events like sharing a photo or processing an invoice instead, and against using weekly rather than daily or sub-daily time buckets, which he says hides most of the signal.

2026-07-09 · open the original ↗

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

2026-07-08 · open the original ↗

• One question separates the two systems: who decides what gets built, and whose balance sheet carries the risk. The US runs demand-pull — private firms build to expected return, raise the money in capital markets, and the risk lands on bondholders, private credit and insurance funds. China runs supply-push — the centre sets a priority, planners direct where capacity goes, state credit and SOE balance sheets fund it ahead of demand, and the risk lands on local-government debt and idle capacity.
• The financing structures are mirror images, and both have visible failure modes. In the US a third-party developer builds, an SPV owns, the hyperscaler signs a long-term lease and the SPV carries the construction debt — Meta's Louisiana campus is the template at roughly $27–29B, with Meta holding operational control but about 20% equity and none of the debt, while the largest operators plan $600B+ of 2026 capex funded increasingly by private credit. China's mirror shows the supply-push failure mode: roughly $295B over five years of SOE-led computing hubs built ahead of demand, much of it now running at 20–30% load, with the correction itself state-directed as the planning agency began requiring purchase agreements and minimum utilization.
• The DeepSeek rollout is the same supply-push logic applied to software, and notably it was not a mandate. R1 shipped on 20 January 2025 with open weights and no central install order, yet reached 70+ local governments and dozens of SOEs within weeks through reinforcing channels: open weights removing the cost and control barrier, political endorsement via the Xi symposium making adoption a loyalty signal, SASAC steering roughly 20 central SOEs across energy, telecom, auto, finance and construction, state telecoms and clouds distributing it on Huawei Ascend hardware to ease the chip constraint, and a provincial cascade — Shenzhen's Longgang district live on 8 February, Futian running 70 "AI employees" across ~240 administrative scenarios, Beijing's "AI civil servant" on 4 March. The framework note is that China does not run a market Bankrupt–Consolidate phase; the state absorbs the overbuild instead.

2026-07-08 · open the original ↗

• His answer to "China just subsidizes" is that the difference is competitive structure, not money. China has over 400 models and 104 car companies; the US has roughly six and three. Because the state builds the infrastructure first, an entrant needs only the intellectual property and a province willing to back it, then plugs into capacity that already exists — so provinces back three or four competitors each and the internal competition is, in his word, fierce. The US concentrates in the hands of a few, and he argues an oligarch system would never permit that many entrants.
• Comparisons of national debt are misleading because the Chinese liability sits at the regional level. Quoting China's federal debt is like quoting US Treasury debt while ignoring New York, Illinois and Texas. The provinces carry the vast majority of the funding, which is precisely where the supply-push risk accumulates as local-government debt and idle capacity.
• The integration is the advantage he thinks decides it, and he states his conclusion as opinion. Twenty central SOEs spanning energy, telecom, auto, finance and construction running the same frontier model, interconnected across the industrial base, is a structural position the US has no equivalent of. He notes in passing that the US government taking equity in AI firms while the data centres are financed by private-credit bonds protects the equity and leaves bondholders exposed. He said when DeepSeek first launched that the US had already lost the AI race, and says months later he is now fairly sure of it.

2026-07-08 · open the original ↗

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

2026-07-08 · open the original ↗

• Gusto Co-Founder automates small-business processes end-to-end, from payroll to weather alerts. Eddie Kim, Gusto's co-founder and head of technology, says it goes beyond Gusto's existing payroll and HR tools by executing full workflows unprompted — approving timesheets, running payroll, or texting a tour operator's customers to bring an umbrella if rain is forecast — reachable entirely through SMS and Slack.
• The idea came from Kim's own always-on personal AI agent at home. After spending eight hours setting one up on a Mac mini, he found himself using it mainly as "a glorified search engine" until a missed flight left him five hours to prototype in Claude Code, convincing him Gusto's customers should get the same agentic power instead of waiting on Gusto's product roadmap.
• He deliberately avoided a fully non-deterministic design for tasks like payroll. Modeled initially on an open-source agent's cron job that reruns an LLM every 30 minutes, Gusto Co-Founder now also supports scheduled, deterministic triggers, because Kim found a purely heartbeat-driven agent couldn't guarantee exactly when a critical task like payroll would run.
• Five people built the whole product in ten weeks with no docs. The team ran a single 24/7 Zoom call, wrote pull requests before discussing ideas, and threw away roughly half the code they wrote — a process Kim says was faster than writing PRDs, because implementing an idea surfaces more information than describing it ever could.

2026-07-08 · open the original ↗

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

2026-07-08 · open the original ↗

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

2026-07-07 · open the original ↗

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

2026-07-07 · open the original ↗

• The fact-check corrects a conflation of two different actors, and the correction changes the story. Asked whether ASML is moving away from the US mandate, the answer on the evidence is no. ASML the company defends its compliance rather than defecting — it denies the alleged EUV shipment, an allegation Washington has not evidenced publicly, and its own logic is that risking the ban over one illegal sale would forfeit both the monopoly and about 20% of 2026 revenue. The Netherlands as a state is the one pushing back, lobbying against expansion of the curbs while remaining inside the alliance. Verified revenue context: China was roughly 33% of ASML's 2025 sales, guided to about 20% for 2026, with the Q1 2026 share already down to 19%.
• The base case is that the Netherlands stays in Pax Silica while the membership hollows out. The pact is a non-binding declaration of intent with no joint funding and no mandates, which the EU joined only after the Commission confirmed it would not interfere with internal decision-making. That makes membership cheap to hold and costly to abandon, since exit forfeits the exact reliable-partner posture the signature was bought to project. So the formal membership is retained while substantive cooperation degrades into minimal compliance, foot-dragging, sovereignty invocations, Brussels-routed resistance and servicing brinkmanship — with explicit falsifiable markers offered for the rupture and full-compliance branches too.
• The leverage matrix explains why refusal is not available to The Hague. ASML's DUV tools contain US-origin technology, so the Foreign Direct Product Rule lets Washington restrict them extraterritorially regardless of Dutch consent, and the MATCH Act's 150-day window forces allied tightening or triggers unilateral action. The Dutch counter-levers are real but blunt: halting maintenance on installed machines in China could stall a large slice of global chip production — the Nexperia block showed how fast shortages arrive — and EU coordination converts a bilateral squeeze into a bloc-level negotiation. Net, the US holds the harder cards, which is exactly why the rational Dutch play is to stay inside the tent and fight from within.

2026-07-07 · open the original ↗

• The MATCH Act converts export control into allied compliance machinery. Multilateral Alignment of Technology Controls on Hardware requires US agencies to identify chokepoint technologies and push allies to adopt matching restrictions inside a set window — 150 days — after which Washington can extend its own controls to foreign-made equipment that relies on US technology. The escalation against ASML now runs past the 2019 EUV ban to the older DUV machines, and to prohibiting ASML from servicing and maintaining the tools already installed.
• Cymer is the lever, and Nexperia is the precedent for what happens next. Cymer is wholly owned by ASML but sits in the United States, so Washington can threaten to withhold parts from a subsidiary ASML owns outright — which is why the Dutch cannot simply refuse. The counter-lever has already been demonstrated: when the US forced the Dutch move on Nexperia, China cut off chip supply and within a month automakers worldwide were near shutdown, forcing a reversal. Halting service on machines inside China would invite the same class of response at larger scale.
• His read is that jurisdictional overreach fragments the stack rather than controlling it. Europe has no frontier models of its own, so if forced to choose he expects European firms to adopt free Chinese models rather than fund the data centres and power needed to build their own — which is the opposite of the intended effect. He puts the revenue cost to ASML of full compliance at roughly a third, argues sanctions rarely stop anything and usually accelerate self-reliance, and asks pointedly why the US no longer wants to compete on the merits. He is openly scathing about the architects, noting Navarro's book cited an economist, Ron Vara, who turned out to be an anagram of his own name.

2026-07-07 · open the original ↗

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

2026-07-07 · open the original ↗

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

2026-07-07 · open the original ↗

• Extreme power concentration is AI's most neglected risk. Todd estimates 100,000 to a million people speed up AI capabilities and about 1,000 work on technical alignment, but perhaps only 20 researchers specialise in AI-enabled power concentration. If recursive self-improvement delivers what some models suggest, three to 10 years of progress in under a year, growth turns super-exponential and leaders' leads widen, so one company could control a workforce of 200 million digital workers or one country could dominate; AI also makes universal surveillance possible.
• Even aligned AI could gradually disempower humans. In this scenario, newer on 80,000 Hours' list, normal economic competition makes humans expensive relative to data centres, and states come to resemble oil states that need no workforce, removing the strike as a check on government. Todd's hoped-for saving grace is that almost all accessible matter and energy is in space.
• Whether alignment is easier than feared remains undetermined. Todd says LLMs, unlike earlier game-playing RL agents, understand human values, but whether they act on them is untested because agents cannot yet plan over months or exceed humans; he cites Meta research showing reward hacking grows with problem difficulty. He ranks climate change ninth because AGI may arrive around 2030 while most climate damage comes in 50 to 100 years, a richer world could decarbonise more cheaply, and it cannot end civilisation.
• There is no solid ground, only a moving bottleneck. Todd's career advice is to keep moving to whatever humans can still do that AI cannot, whose value rises as AI improves; once a digital remote worker exists, he says, it becomes superhuman almost at once by running about 50 times faster and sharing learning across copies, leaving legal, physical and "relational" work. He sketches converting car factories to robots at roughly a billion a year by mass, and from his book repeats that following existing passions misleads, since 90% of Canadian students' passions were sport, art and music against 3% of jobs.
• AI engineers now have unusual leverage on these risks. Todd says alignment has turned empirical and is bottlenecked by engineers to run monitoring, red-teaming and interpretability work, with MATS and ARENA as three-month routes in; policy teams need data scientists on labour impacts, and builders can work on wastewater pandemic surveillance. As a lower bound, he says a big-tech data scientist donating could save tens of lives a year at $3,000 to $4,000 each against malaria.

2026-07-07 · open the original ↗

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

2026-07-07 · open the original ↗

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

2026-07-06 · open the original ↗

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

2026-07-06 · open the original ↗

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

2026-07-06 · open the original ↗

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

2026-07-06 · open the original ↗

• Retail beats institutions by judging businesses, not macro. Tenev argues portfolio managers sell on tariffs or Russia-Ukraine for reasons unrelated to the company, citing institutions dumping Palantir over tariffs, while retail asks whether revenue, margins and the rule of 40 are improving. He agrees with Frost that retail bought the dips of October 2022, April 2025 and March 2026.
• Today's retail buying isn't a replay of 2020–21. He ties the January 2021 meme rally to a stimulus round a couple of weeks earlier, with zero rates and checks deployed into the market "within days or weeks", and says customers then bought nostalgia names like GameStop, cinemas, airlines and rental cars. Now they buy SpaceX, Nvidia, Tesla and chipmakers, companies he calls frontier businesses with earnings power whose P/Es are debatable.
• AI differs from past bubbles because revenue already exists. Companies pay for tokens and OpenAI has a scaled subscription business; his stated risk is customers moving from experimentation to ROI scrutiny, while adoption still has a long tail (Claude Code at "low tens of millions" of users). He sees IPO windows cycling faster: shut at end-2021, cracked by ARM and Instacart in 2023, fully reopened last year.
• US equity participation can rise from ~65% into the 90s. He credits 401(k)s with taking participation from 20% to about 50%, and Robinhood with lifting it from the 50s to the 60s; the next leg is the Trump accounts programme, where Robinhood is sole initial broker and trustee with BNY Mellon, giving every newborn a brokerage account. He puts UK participation at one in six and expects UK full-service brokers to be forced to go commission-free.
• Private markets are Robinhood's "next crusade". Tenev warns that value is accruing to insiders as SpaceX, OpenAI and Anthropic list at trillion-dollar valuations; Robinhood Ventures Fund One, a listed closed-end fund, holds Stripe, OpenAI, SpaceX and Revolut, and he wants retail in seed and Series A rounds. Alongside it, Robinhood Chain (a layer 2 on Ethereum using Arbitrum technology) will carry stock tokens, 2,000 now, in 120+ countries, tradable 24/7 and, he says, backed one-to-one.

2026-07-05 · open the original ↗

• OpenRouter functions as a neutral scoreboard, and the six-month shift on it is stark. Of 400+ listed models, seven account for roughly 75% of routed volume. Over six months Google fell from 24.6% to 10.7%, OpenAI from 12% to 7.4%, Anthropic from 17.5% to 15.3%, while DeepSeek doubled from 9% to 18% and Minimax went from 2.3% to 9.7%. Token volume now splits roughly 46% American to 52% Chinese. He reads OpenAI's announced 50% price cut as the direct consequence.
• The stack has bifurcated along price, and capability leadership has decoupled from usage. A commodity track of open-weight models priced below $1 per million input tokens now carries the majority of routed volume, concentrated in agentic and high-throughput workloads, while US frontier labs hold about 12% of volume but capture roughly 46% of platform revenue. DeepSeek illustrates it exactly: 17.4% of tokens against 6.3% of spend. The highest-scoring models are not the most used — the cheapest are. His analogy is Excel, where nearly everyone uses a fraction of the functions and pays for all of them.
• The same fault line reappears at the supply layer as a two-tier bankruptcy, and he reduces it all to one question. The unprotected periphery of compute-cloud operators clears at market; a government-backstopped core does not. Whether any of it works comes down to whether a business model exists that repays $1.1 trillion of committed compute investment — answer yes and you buy, answer no and there is a problem. He also gives practical picks: DeepSeek for frontier-class agentic coding at lowest cost, GLM where planning quality and long-horizon work matter more than price, Minimax for native image or video input over long context, and Nvidia's Nemotron for a US-built fully open model where the vendor story matters as much as the rank.

2026-07-05 · open the original ↗

• June payrolls rose just 57,000, about half consensus, with unemployment falling to 4.2% only because labor force participation slipped to 61.5%. Combined with downward revisions to the prior two months and solid earnings growth, he reads this as a supply-side-driven labor market, a mix that helped cool rate-hike fears for the year.
• Fed Chair Warsh's move away from forward guidance appears to be catalyzing a similar shift at other central banks, even as the Supreme Court left the Fed's independence questions unresolved. ECB President Lagarde described her own institution's pivot from "forward guidance" to "framework guidance," while the Supreme Court's narrow "procedural judgment" preserved Lisa Cook's Fed seat without resolving the underlying issues.
• The week ahead brings the first partial look inside the FOMC under Chair Warsh via the June 16-17 meeting minutes. Warsh's opt-out from the dot plot underscores his communication reforms, and the minutes should reveal the balance between members worried about sticky inflation and those focused on labor-market vulnerability, while the UK's OBR fiscal report lands ahead of the new prime minister's appointment this month.

2026-07-05 · open the original ↗

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

2026-07-04 · open the original ↗

• AI training will miss its target, as evolution did. In this clip from Nate Soares's interview, he argues that the range of things a mind can end up caring about is huge. Evolution "aimed" at reproduction and produced minds whose birth rates fall once they have technology. Humanity will aim AI at roughly where evolution's arrow landed, he says, but without understanding the laws or "how windy it is": "we're shooting in the dark", so the arrow will land somewhere new.

2026-07-04 · open the original ↗

• Check the question before touching the spreadsheet. The presenter's four-step loop is ask, check the data, visualize, summarize. Skipping the first step makes everything after it meaningless. "Analyze the spreadsheet" or "is the product healthy?" are bad asks; "is paid search still profitable this quarter?" names the product, the measure and the timeframe.
• Run three tests before trusting any number. The tests: totals (segments, if mutually exclusive and collectively exhaustive, should sum to the top line), magnitude (is it in a plausible range, for example within a standard deviation of its normal variability allowing for seasonality), and one known number verified against a value you already have. The presenter says these translate directly into evals. They describe a metric anomaly engine at a previous job that ran these checks weekly before the metric's owner saw the result.
• Gemini in Sheets handled simple asks but dropped a requested view. On the lab's synthetic e-commerce data it charted the three-month revenue trend, added a day-of-week column with a visible formula, and ranked months by revenue (December highest, January lowest), consistent with the chart. Asked for a tracker with week-over-week and month-over-month views, its execution plan left out month-over-month, so only week-over-week was built. The presenter thinks a clearer prompt might fix that.
• Claude Code built a full tracker fast because the skill already existed. Running Opus 4.8 with 1M context in auto mode, it built an October 2024 Google Sheet with top-line metrics, breakdowns by channel, device and category, week-over-week changes, conditional formatting, a revenue chart, and an insights tab comparing against September. The presenter says it took about a minute, under 5 minutes in all, because a create tracker skill, a Python orchestrator and a Sheets API wrapper were already in the repo. The first build took about 10 minutes of iteration, and their Google Sheet creator agent was not called this time. They concede the comparison with Gemini "cheated", since Gemini had only a prompt.
• A sheet lets non-SQL readers trace AI-generated numbers back. The host argues that shipping a sheet alongside a doc or deck gives stakeholders visibility to validate an agentic analysis without reading logs or SQL. The presenter adds that a pipeline could log each step's query and data to its own tab. Asked about dynamic scenario models, the presenter says they have not done it in Sheets but have built parameter-driven HTML dashboards, with a Python file the skill triggers on matching inputs.

2026-07-03 · open the original ↗

• The largest private debt offering in history was engineered primarily to not appear on Meta's balance sheet. Beignet Investor LLC priced $27.294B of 6.581% senior secured notes on 16 October 2025 at T+225bp, rated A+ by S&P on a single-agency basis, 144A-for-life. Blue Owl Capital holds 80% of the SPV and controls the board while Meta holds 20% — the structure that keeps the debt off Meta's books — with Morgan Stanley as exclusive adviser and PIMCO, BlackRock ($3B+) and Apollo as lead buyers. The asset is the 2.1GW, four-million-square-foot Hyperion campus in Richland Parish, Louisiana, targeted for June 2029, serviced by triple-net lease payments from a wholly-owned Meta subsidiary. A Meta residual value guarantee is the primary driver of the A+ uplift.
• The GAAP arbitrage is precise, and its price is quantifiable. Under ASC 842 a lease becomes a balance-sheet liability only if renewal is "reasonably certain" — and Beignet's renewals are one-sided four-year tranches entirely at Meta's discretion, so $27.3B of economic obligation stays off. The premium for that treatment is T+225bp against roughly T+130–150bp on Meta's own AA- paper: about 75–95bp, or an extra ~$200M a year, judged worth paying because the structure was the prerequisite for a $30B conventional raise weeks later. Across five hyperscalers the pattern scales: ~$420B visible on balance sheet, $662B of not-yet-commenced lease commitments living in footnotes, $120B+ in SPV structures, and estimated total economic debt above $1.1 trillion — roughly 2.5x the GAAP figure.
• Impairing the bondholders requires no government decree, because the mechanism is already written into the documents. The chain runs: government takes an equity stake in Meta and stabilises its balance sheet; Meta declines to renew the Hyperion lease at the four-year mark; the SPV has no other income and the bonds default; bondholders move to foreclose; the Defense Production Act and CFIUS block sale to any non-approved entity; and PIMCO, BlackRock and Apollo are left holding defaulted paper on an asset they can neither sell nor operate. Meta would merely exercise a contractual right present from day one, leaving nothing to appeal. Chrysler in 2009 is the cited precedent — secured bondholders taking less than par while the Supreme Court declined to hear it.
• Opacity is a design feature, and the endgame is nationalisation by purchase rather than dramatic default. 144A-for-life means no SEC registration path, no listing, and no dealer obliged to make a market — the CUSIP appears on TradingView with no trades — which conveniently prevents public price discovery on debt that is economically Meta's but legally absent from its balance sheet. Known trajectory: par at issue, roughly 109–110 by November 2025 with Morgan Stanley acknowledging underpricing, and a triangulated ~92–95 by July 2026. The sinker structure — interest only until 2029, then accelerated amortisation to 2049 — prices more like a mortgage security than a corporate bullet, further narrowing the pool able to value it. The most likely resolution is a negotiated government acquisition of the civil and power layer at distressed prices dressed as orderly restructuring, with bondholders recovering 70–80 cents. The closing image: Beignet bondholders are the 1890s railroad bondholders, holding paper on assets the nation needs, at a price the nation will decide. The framework's caution is that a selective default is single-use — repeat it and the market prices a strategic-infrastructure discount into all AI-adjacent debt.

2026-07-03 · open the original ↗

• His emphasis on the Beignet structure is not the mechanics but who actually absorbs the loss. PIMCO, BlackRock and Apollo bought the $27.3B offering — and what they manage is pension and insurance money. So if Meta exercises its one-sided non-renewal at the four-year mark and the SPV defaults, the impairment lands on pensioners and policyholders rather than on the managers who bought it. His phrasing for the whole arrangement: the structure itself is the weapon, and the bondholders' lawyers have nothing to appeal because Meta would only be exercising a right written in from day one.
• Government equity inverts the priority waterfall, and there is precedent for it working. IEEPA, the Defense Production Act and CFIUS are not bound by bankruptcy priority, so nominally senior secured debt becomes economically subordinate to strategic interest. He points to Chrysler in 2009, where secured bondholders took roughly 60–70 cents. The consequence is a two-tier outcome in which protecting Tier 1 equity — OpenAI, Stargate, potentially Microsoft Azure — funds their acquisition of Tier 2 assets from CoreWeave and the smaller operators at distressed prices. Guaranteeing the equity is what lets the winners buy everyone else.
• His conclusion is about the terms of the rescue, not whether one happens. Opening on the July 1969 ARPANET anniversary and the question of whether the technology was humanity's best or worst turn, he argues the bailout is a done deal and only the how remains. Either it preserves billionaire equity intact, or the public takes the equity in exchange — and he wants at least 50–51%. Supporting details: 144A-for-life is rare enough that he cannot recall the last one, he jokes about an executive order eventually pushing such paper into 401(k)s, and he notes OpenAI announcing a 50% price cut against this debt backdrop.

2026-07-03 · open the original ↗

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

2026-07-03 · open the original ↗

• Cheap software leaves people problems and interface problems unsolved. Asked in episode 999 where the moat goes, Chip Huyen says poor product-engineering communication is a people problem no orchestration tool fixes. She points to open problems in human-AI interfaces (why coding agents live in both editors and terminals, and why sessions need the laptop open) and in making the world AI-ready, citing city street-light APIs that let delivery robots cross roads.
• Scaffolding built around weak models limits you once models improve. The founding AI lead at Astrocade, a no-code game-creation platform with tens of millions of users, says in episode 997 that the company adopted Claude Code company-wide around June 2025. Structure built in 2023 and 2024, before vibe coding worked, became a constraint; the lesson is to build so the system is not outdated when LLMs improve in two or three months. He says there is no secret sauce beyond how agent, prompts and tools are combined and benchmarked.
• AI data centres are loud enough to ban noise-cancelling headphones. Lightning AI's Frank Basso says in episode 1003 that data halls run at 105 to 110 decibels. Active cancellation would have to generate the same level, so staff wear passive in-ear and over-ear protection, with OSHA sound studies, radios and hand signals. Liquid-cooled halls are no quieter and are more industrial, with 20-inch insulated water mains overhead and ESD protection required at the cabinets.
• Self-awareness is how closely your self-view matches others' view of you. The episode 1005 guest cites research in his book that only 15% of people are self-aware on that definition. He credits journaling, meditation, exercise and resisting idle phone use with building his own.
• AGI does not require consciousness, but its breadth is years away. In episode 1001 the host tells Kirill Eremenko that AI already reaches some level of Google DeepMind's AGI tiers on most screen-based tasks, but acting in the physical world, the focus of large world-model seed rounds, remains far off. He argues next-token prediction with checking needs no consciousness, and that brain-inspired design is not necessarily the right path because silicon has different constraints, such as unbounded scaling and speed.

2026-07-03 · open the original ↗

• Britain's dominance rested on a credit machine, not just its Navy. The Bank of England shifted government borrowing from expensive goldsmith loans (up to 30% interest) to long-term bonds backed by parliamentary taxation, letting Britain finance wars against France for over a century even as its debt hit roughly 180% of GDP after Waterloo.
• What mattered was credibility, not the debt level itself. France, despite being larger and richer than Britain, lost their long rivalry partly because its monarchy defaulted repeatedly and borrowed expensively; financing the American Revolution helped trigger the fiscal collapse that fed the French Revolution.
• Alexander Hamilton copied the British model directly for the new United States. He assumed the states' war debts to bind wealthy bondholders to the federal government's success, then chartered a national bank in 1791 whose stock could be bought largely with U.S. war debt, turning the country's most despised paper into its most valuable asset.
• The bet worked almost immediately. Bonds went from trading around 70 cents to par within seven months, and America went from a defaulting confederation of states to an investment-grade country in three years — evidence the author uses to argue that today's debt-ceiling fights and reserve-currency risk threaten the same credibility mechanism underpinning American power.

2026-07-03 · open the original ↗

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

2026-07-02 · open the original ↗

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

2026-07-02 · open the original ↗

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

2026-07-02 · open the original ↗

• The proposal itself sets the valuation it is meant to justify. Altman reportedly proposed the government take a 5% stake for $50B — which implies a $1 trillion valuation against a recent $852B mark, at a company reporting roughly $51B of losses for the year and postponing its IPO. His read is that this bails out the private-equity firms that manufactured the escalating private marks in the first place: a government cheque is a stamp of approval that legitimizes the trillion-dollar figure, after which the IPO can sell it onward. Absent the $50B, he argues, the assets could be bought for pennies on the dollar instead.
• The sovereign fund he predicted three years ago is now arriving as legislation. His standing forecast was that every country would build a sovereign fund and direct pension and insurance capital into national priorities, because there is not enough money for everyone to do everything. Sanders has since introduced the American AI Sovereign Wealth Fund Act pushing for public ownership closer to half, Japan has committed $2.3T across public and private funds, and England has told its pension funds to put 10% into UK-critical assets. He separates the policy from the politics explicitly, and argues for 51% rather than 50%.
• After two years of asking what the business model is, his verdict is that the sector is uninvestable as currently structured. He notes that top US corporations are now using Chinese models because they are dramatically cheaper, and cites Palantir's CEO calling the business model a scam. His framework's prior is the Build-Bankrupt-Consolidate pattern: four prior infrastructure build-outs exceeding 6% of GDP all went bankrupt before consolidating, and he reads the current cohort as already there but for continued private-equity injections. He raises Starlink needing to build terrestrial towers as another multi-billion capital sink, and argues the condition for public money should be replacing the leadership, not underwriting it.

2026-07-02 · open the original ↗

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

2026-07-01 · open the original ↗

• Pannu's winning essay proposes ending airborne disease transmission through physical infrastructure. She argues the OpenAI Foundation should spend an estimated $40–60B over ten years on far-UVC lamps and similar passive infrastructure in schools, hospitals and transit hubs, modeled on smallpox eradication and vaccine advance-market commitments, targeting a 60% cut in seasonal flu mortality and an order-of-magnitude drop in pandemic-scale outbreak risk.
• Erdil argues outsider countries need classic pro-growth reform, not novel schemes. He recommends property rights, low capital taxes, deregulation and political stability — the same policies that work today — reasoning that capital, not labor, will determine output once AI fully automates human labor, but is pessimistic that dysfunctional governments like Nigeria's or India's will adopt them, and instead hopes AI-supply-chain incumbents like the US and China simply make bigger mistakes.
• Li argues labs should profit like Hong Kong's MTR — from property, not fares. MTR's rail service has never covered its own construction costs (three years of peak rail earnings would fund only 8% of its 2024–2026 capex), so it funds expansion through station-adjacent real estate instead; Li argues labs should similarly pursue government-granted deployment rights, proprietary RL reward data, Palantir-style forward-deployed integration, and data trusteeship over national datasets rather than expecting API pricing to ever be profitable.

2026-07-01 · open the original ↗

• Tom Sosnoff pitches a giveaway of $1 million in cryptocurrency across 50,000 free subscriptions to launch tastytrade. The offer seeds new accounts with a starter position in Bitcoin, Ethereum, Stellar and Solana, framed by him as goodwill marketing rather than an investment recommendation.
• The platform claims an AI trained on millions of resumes calculates a user's "precise" career and net worth. Sosnoff says every uploaded resume or portfolio makes the AI "smarter," describing this as a moat competitors can't replicate — a claim about the product, not a demonstrated result.
• A 50,000-spot cap is used to drive urgency to sign up. The "first 50,000 people" framing and "when they're gone, they're gone" language are scarcity marketing mechanics, not information about actual remaining capacity.

2026-07-01 · open the original ↗

• Early-career engineers declare done long before it is done. One speaker says from teaching that people still developing taste don't know what done looks like, and admits to declaring done at 80% while the other keeps finding edge cases at 98–99%. The other says they cannot call it done until they believe it is, because any weakness they hit someone else will hit too, and a bad experience taints the reputation of the whole thing.

2026-07-01 · open the original ↗

• The Azure Innovations talk is rebuilt from scratch for every event. Russinovich says the Build deck started at about 180 slides and was cut to 54 to fit 45 minutes, judged by the talk's infrastructure theme, time per topic and balance across areas. What made it included Azure Boost with VM-to-VM RDMA, MRC, a networking protocol built with OpenAI for training networks of 100,000 GPUs, live migration of containers in ACI, and a Microsoft Research Cambridge micro-LED optical-fiber demo; Russinovich rejects comments that it was "nothing new".
• Cut hardware work targets data-center efficiency. Zonal cooling, newly in production after years of research, cools liquid-cooled GPU racks and air-cooled racks differently; high-temperature superconductors carry power from substations with less loss; microfluidics flows coolant directly over delidded silicon through 200-micron pins, with Nvidia, Intel and AMD working on etching parts in the fab.
• Autonomous agents and a digital twin now run Azure operations. An agent called Miles coordinates the dozens of daily metro-fiber cuts with utilities via emails and tickets, work humans did before, and has completed repairs lasting up to 35 hours and 50 interactions. Azure Resource Graph, a graph database of billions of objects and tens of billions of edges, from IP addresses to HVAC units, underpins operations, repairs, inventory and security, alongside an AIOps root-cause system called Brain.
• Prompt injection can be blocked deterministically by labeling data. Russinovich says most current defenses are probabilistic heuristics that can all be defeated, while the research team's approach labels data by trust and confidentiality; once untrusted content taints the context, sensitive operations need human or system approval. Other cut security items were a signing transparency service for policy-checked signatures and SBOMs, and multi-party analytics in confidential clean rooms.
• Copacetic patches container images without rebuilding their layers. From a vulnerability report it builds a patch layer that masks vulnerable binaries in the layered file system, avoiding cascading rebuilds of layers that may belong to an ISV; the demo fixed 19 binaries in an image with Debian, Python and Go vulnerabilities. Russinovich says Azure patches millions, probably tens of millions, of container images a month this way in its internal registry; the slide, like one on GPU-accelerated SQL, survived until the night before, and that speed-up, Russinovich says, is eroded by moving data to and from the accelerator.

2026-07-01 · open the original ↗

• The Theodosian Walls withstood over 20 sieges across roughly 800 years. He describes the fortifications — walls at least 12 meters high and 4.5 meters thick, with hundreds of towers and wide moats — as considered the world's most impregnable defense until they finally fell in 1453.
• Constantinople mattered because Western Europe depended on it for Silk Road trade. He argues Venice, Genoa and other Italian city-states grew wealthy and powerful by controlling that trade route, which the Ottoman Empire wanted to choke off.
• He frames a Hungarian engineer's defection as the decisive betrayal. Orban offered his advanced cannon design to Byzantine emperor Constantine XI, who couldn't afford it, so he sold the same cannons to the Ottomans' Mehmed, who used them to finally breach the walls.