Souk's Notes - Unofficial Summaries

2026-06-11 · open the original ↗

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

2026-06-10 · open the original ↗

• Project Nomad will support both, leaving users the choice. The video opens with the claim that MeshCore is far better, one local area having switched months earlier; the host says he has only ever used Meshtastic and is not a heavy user, but has heard the same. Much of the same radio hardware runs either, and both will be in the Supply Depot.
• The host's goal is asking Nomad's local AI questions over the mesh. He is most interested in a Meshtastic API container, so a message like "!nomad" plus a question, picked up over 915 MHz by an Ethernet-connected radio, gets answered by the server's local LLM over Meshtastic or MeshCore. He describes this as his "dream", not a shipped feature, letting someone many nodes away query a distant Nomad server offline.

2026-06-10 · open the original ↗

• A well-trodden structure applied for the first time to an asset it was never designed for. Apollo and Blackstone arranged a ~$35B private-credit package financing Google TPUs through an SPV that buys the chips and leases them to Anthropic — Anthropic is tenant, not borrower, so the debt stays off its books — with Broadcom guaranteeing the senior tranches against a resale shortfall. Off-balance-sheet financing of leased assets is legitimate and decades old; what is new is applying it to a fast-depreciating, revenue-deflating asset rather than real estate, and routing the residual risk back to the chip's own manufacturer.
• The depreciation history contains a correction that strengthens the bear case rather than softening it. Server life ran about three years through 2019, then Microsoft went 3→4 in 2020, Google, AWS and Meta stretched to five and six by 2022, Alphabet went 4→6 in early 2023 — cutting that year's depreciation by ~$3.9B and lifting earnings ~$3B — and the big three converged at six years, with Amazon partially reversing to five in Q1 2025 citing the pace of AI development. The assumed life roughly doubled exactly as accelerator generations began landing every one to two years. The correction: firms decelerated depreciation by extending useful life; they did not accelerate it, and the popular framing has it backwards. What matters is the resulting gap — roughly 50% of cost still on the books at year three against a resale value nearer 15–30% if the two-to-three-year real-life thesis holds.
• Two blades close on the chip at once, and the paper is careful about what that does and does not mean. Blade one is resale value decaying faster than the book schedule. Blade two is per-chip revenue capacity deflating faster still — per-token prices fell from roughly $10 to $2.50 per million in a single year, and newer silicon serves the same tokens two to three times cheaper, so an aging TPU can drift into negative carry before it physically fails. The clarification the document insists on: aggregate AI revenue is rising, with total inference spend up ~320% as unit prices fell ~280x. The pie grows; any given chip's slice shrinks; the collateral is the chip, not the pie. On the illustrative grid, a 50% resale leaves a ~$10.5B shortfall on a year-one default and roughly nothing by year three, while a severe 15% resale leaves ~$22.75B and ~$11.75B — amortization is the guarantee's friend, and the worst corner is one correlated shock rather than two coincidences.
• The circularity is the structural finding, and the briefing argues both sides of it. Three mechanisms stack: off-balance-sheet financing, credit-rating arbitrage where Broadcom's guarantee lifts the senior tranches toward investment grade without consolidating the debt onto Broadcom, and manufacturer-as-backstop circularity — Broadcom co-designs the TPUs and guarantees the debt financing their purchase, underwriting demand for its own product. The cautionary analogue is exact: Lucent and Nortel financing customers to buy their own gear. Aircraft EETCs are the legitimate analogue, except planes hold value and have deep resale markets. Notably the document carries the rebuttal in full — Nvidia's memo that customers depreciate over four to six years on real-world longevity with 2020-era A100s still at high utilization, and the waterfall counter that chips cascade from frontier training to inference to internal workloads and so genuinely earn for six years, just not at the cutting edge. It then names what would change the conclusion: the true guarantee scope, whether the debt is fully amortizing or interest-only (which would keep the year-one worst case alive for years), whether custom TPUs ever trade arm's-length at all rather than via Google redeployment, and whether older chips stay above negative carry.

2026-06-10 · open the original ↗

• He grounds the structure in the two times it previously went wrong. An SPV, SPE or SIV lets a company form a second entity, borrow into it, and keep the debt off its own balance sheet. Enron was the first version he encountered — a commodity trader whose vehicles hid debt far exceeding its equity, so once they were understood the company "melted away in a day." The second was the investment banks' SIVs before the financial crisis, where 30–40x reported leverage was closer to 100x counting the vehicles. The traditional collateral was long-dated: aircraft on twenty-year lives, property. Chips are the novelty.
• His arithmetic makes the residual-value gap concrete. Depreciate over six years and roughly 48% is written off by year three — but with new chip generations landing every one to two years, the chip may only fetch 15%. The difference between the ~52% still carried and the 15% realizable is precisely what Broadcom owes, because Broadcom guaranteed the resale. Then the second blade: per-token prices have collapsed, so revenue per chip declines alongside the asset value, and an aging TPU drifts into negative carry — power and hosting exceeding what it can earn.
• The circularity is what he finds indefensible, and he names who ends up holding it. Broadcom co-designs the chips and guarantees $30B of the debt financing their purchase. He sharpens this with the Broadcom CEO's remark about not seeing productivity gains from AI agents — though his own companion document that day flags that quote as misleading, since the comment concerned software cannibalization rather than AI productivity generally. He also notes Blackstone capping withdrawals from its credit fund while arranging this deal. His summary of the risk waterfall: if the bet works Anthropic gets its compute, and if it does not, the losses flow first to the investors and then to Broadcom.

2026-06-10 · open the original ↗

• Putin has reportedly disabled parts of his own surveillance network out of fear it could be hacked. The speaker frames this as inspired by Israel's infiltration of Iranian systems, and says Russia now faces Ukrainian hackers he ranks among the world's best, tracing that talent to Ukraine's role as the Soviet Union's STEM "breadbasket."
• He argues China faces the same risk at larger scale as the world's largest surveillance state. He names Taiwan, India and the United States as likely hackers of Chinese systems, and says China's pursuit of quantum communication is meant to close the same vulnerability.
• He cites the Snowden leaks as proof of a pattern he says he's argued for five years. He says China only suspected US surveillance before the 2012-2013 disclosures confirmed it, and frames it as one example of his broader claim that authoritarian states overbuild surveillance advantages that later get turned against them, as he says already happened in Iran and expects in Russia.

2026-06-10 · open the original ↗

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

2026-06-10 · open the original ↗

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

2026-06-10 · open the original ↗

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

2026-06-10 · open the original ↗

• Minimal surface area, not AI-enabled breadth, separates successful companies. Brex co-founder and CEO Pedro Franceschi points to Stripe's early API-only product, Brex's own terminal-only MVP, and Airbnb's single-form booking flow as evidence founders should compress a problem to one disciplined interaction, warning that AI's ease of building "so many other things" tempts founders into losing that focus.
• Brex secures its AI agents by auditing traffic with another AI. Franceschi says the company's open-sourced "Crab Trap" tool proxies every HTTP request an agent makes, uses an LLM to judge whether it matches an approved policy, and after training on one day of an agent's traffic can auto-approve about 98% of requests — the unlock that got Brex's security team comfortable letting agents write into production systems.
• Founders can't prompt their way to a winning idea. Franceschi argues models can execute almost any idea now, but a customer describing their problem gives you their own limited, locally optimal framing, not the breakthrough answer; the founder's remaining job is extracting what customers aren't saying, since "the execution is gone" but "the wisdom to choose is still the bottleneck."
• Most companies chronically underspend on AI tokens, even at scale. Comparing token-consumption data across companies, Franceschi says a roughly ten-mile radius around San Francisco and New York shows dramatically higher usage and correlated faster revenue growth than large, well-funded companies elsewhere that spend a small fraction of what the economics would justify.

2026-06-10 · open the original ↗

• Legal Agent is a gated preview inside M365 Copilot. According to Giard it needs an M365 Copilot license and an organization enrolled in the Frontier program, and during the preview it works only in the United States. You can tell you have access if Legal Agent appears under the plus icon after opening Copilot (bottom right) in a Word document.
• It reads contracts with knowledge of their structure. Giard says this is what it adds over plain M365 Copilot. He ran it on an 11-page services agreement taken from the public internet. It answered the top five risks to the buyer (including broad IP assignment and late payment), the vendor's obligations grouped by category (such as service delivery, payment and fees, and intellectual property), and the indemnification clauses and their impact. Each finding linked to its section of the document.
• It can draft and apply edits to the document. Two prompts asked it to add a termination-for-convenience clause in line with standard enterprise contracts, and to revise a clause to reduce a vendor obligation while keeping its intent. It proposed text that he scrolled to and applied with a click.
• It produces executive summaries, including ones aimed at a specific audience. A one-page summary of risks, obligations and recommended actions came back with links into the document and could be pasted into an email or another document. A follow-up asking what a CFO should care most about returned bullets for that reader.
• A playbook can stand in for individual prompts. A playbook is your own way of reviewing a legal document, loaded from SharePoint or your computer. Legal Agent turned his simple playbook into a list of topics to analyze, which can be extended before continuing. It then ran the whole analysis and suggested text to apply, for example on the change-management process.

2026-06-10 · open the original ↗

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

2026-06-09 · open the original ↗

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

2026-06-09 · open the original ↗

• ARM support for Project Nomad has no timeline. Answering a viewer with an M5 Pro MacBook Pro who wanted to run it under Parallels, the host says ARM is on the radar but the priority is making the project solid on x86 first. Native MacBook support will come "at some point", he says, but is not a focus.
• Mac users can get GPU-backed AI by running Ollama natively. A dockerized Ollama on a Mac cannot use the GPU, so the workaround is to install Ollama natively on macOS, skip Nomad's own AI install, and point Nomad's remote-Ollama setting at the Mac's IP and Ollama port, which the host thinks defaults to 11434. He notes Apple silicon machines hold some of the top 100% scores on Nomad's benchmark leaderboard.

2026-06-09 · open the original ↗

• SpaceX's IPO lockup is tied to a stock-price threshold rather than a fixed date. Instead of the standard six-month lockup, existing holders reportedly can't sell until shares are up more than 30% from the IPO price — a condition said to apply broadly, not just to employees — which could stretch the lockup out to a year or more.
• A separate share class leaves Elon Musk with outsized control of the company. One host flags Musk's roughly 45-50% stake in a distinct share class as a structural problem with the deal.
• The IPO is called "hugely overvalued" relative to the underlying business. Another host pushes back, noting Starlink is performing well even as the company is said to have lost around $20 billion elsewhere.
• Even a space-sector bull is skeptical the core business justifies the valuation. A colleague described as a space enthusiast reportedly says SpaceX will need new revenue, with talk of putting data centers in space, to grow into its current price.

2026-06-09 · open the original ↗

• The simultaneity is the signal, and he does not believe in coincidences. Four very large AI IPOs are queued, at least two at trillion-dollar valuations, alongside capital raises of roughly $80B from Google and $60B from Meta — and on 6 June the president said he was looking at taking stakes in AI companies. His question is the practical one nobody is answering: how does an IPO work when the United States takes a position in it, and why does an industry supposedly about to transform the economy need government capital at all?
• DeepSeek moving from self-hosted to directly paid is the shift he flags for investors. It is now reported as the number one AI app used by US corporations, and the transition that matters is companies paying DeepSeek directly for hosted inference rather than running open weights themselves. The price gap is the story: roughly 19.66 cents per million tokens against $15 at Anthropic. He characterises corporates as hedging rather than abandoning US labs — but frames the spread as exactly the revenue American AI is not collecting to service debt already issued for data centres.
• Three capital events land at once and each distorts a different phase of Build-Bankrupt-Consolidate. Government capital removes the tightening mechanism, so building continues past economic rationality — his analogue is railroad land grants producing more track on worse economics and a bigger crash. The bankrupt phase becomes two-tier, with OpenAI, the Stargate consortium and possibly Microsoft backstopped under national-security framing while CoreWeave, Lambda Labs, Applied Digital and smaller colocation operators clear at market. And consolidation concentrates, because subsidised hyperscalers become the only credible acquirers of the wreckage. His read of the IPOs themselves is blunt: they are not expressions of confidence but liquidity events for early capital — the big boys getting out and letting the little boys in — with CoreWeave's stock as the leading indicator to watch.

2026-06-09 · open the original ↗

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

2026-06-09 · open the original ↗

• Software is now nearly free to build, so value moves elsewhere. Huyen says the cost of building software is approaching zero but its value is not: a GitHub-crawling micro-tool she posted got about 300,000 views in a week and was cloned with AI the next day, showing anything that exists can be replicated. Her answer to "what's left" is people problems (she rejects a founder's pitch that an agent orchestration platform would fix product–engineering communication), human-AI interfaces that move past legacy terminals and IDEs, and making the world AI-ready, like cities exposing a streetlight API so delivery robots needn't ask pedestrians to press the button.
• AI engineering treats models as a service, but systems design still decides results. She distinguishes ML engineering (collect data, train, deploy your own model) from AI engineering (start from a hosted model, so time to market is far faster), while saying real systems combine both, e.g. a customer-support bot using classical classifiers to filter irrelevant requests and route FAQ questions away from expensive models. She argues system thinking matters more as AI automates narrow skills, which she says is why Designing Machine Learning Systems has grown more relevant.
• Start simple: fix the observed failure with the lightest technique. Huyen favours prompting, then context such as documents or web search, and treats fine-tuning as a last resort because serving and inference optimization are hard and companies tell her new models outclass old workarounds every two to three months. She says long, multi-author prompts accumulate contradictions nobody reads end to end, and found one stray New York example in her own prompt skewing every research summary toward New York offices.
• Agent token bills are driven by tool use, not reasoning. She says prediction-market research agents cost her $1–2 per request, and one run visited about 1,000 URLs of which only 20 were unique because varied queries return the same pages. She says reasoning tokens are relatively cheap compared with web search, file search and productivity-app tool calls, and that some companies told her evaluation takes about 20% of their token costs.
• Evaluation guidelines for LLM judges are the hard, human part. She advocates "evaluations-driven development", asking audiences whether having no AI system is worse than one nobody can tell is working, and cites a case study where 80% of development time went into writing LLM-as-judge guidelines. Her test is to have yourself or a coworker grade examples by the guideline, because a model follows instructions but not bad ones.

2026-06-09 · open the original ↗

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

2026-06-08 · open the original ↗

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

2026-06-08 · open the original ↗

• Multiple founders credit Startup School with their decision to apply to YC. Several describe it as the pivotal event that gave them "the courage to apply," with one saying their company's entire trajectory would look different without having attended first.
• Seeing peers their own age build companies made founding feel achievable. Attendees point to hearing speakers like Fei-Fei Li and John Jumper the prior year as what "ignited the passion" to start a company, rather than any specific technical lesson taught.
• The event is framed as an unmatched concentration of ambitious, skilled people. Attendees describe it as selecting specifically for founders who are smart, ambitious and technically capable, with the shared room of similarly driven peers cited as what convinced several of them to commit to starting now rather than later.

2026-06-08 · open the original ↗

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

2026-06-08 · open the original ↗

• India revived its 2013 currency-defense playbook on June 5, 2026, but is deliberately skipping the interest rate hike that made the original work. The RBI resumed subsidized swap rates (FCNR deposits) to attract dollars from Indians abroad while holding repo rates at 5.25%, even as rupee outflows this year already exceed all of 2025.
• Every deferral tool the RBI uses instead of raising rates creates a dollar liability with a future due date. Hedging costs on the new FCNR window run roughly $125 million per $1 billion raised over a five-year tenor, on top of an existing $100 billion RBI short position, meaning the bill compounds rather than disappears.
• The RBI's broad legal mandate — position caps, NDF bans, capital flow nudges — lets it avoid the one visible, politically costly move: a rate hike. By contrast, the author notes, the U.S. Federal Reserve has no comparable capital-control authority and effectively only one lever, since dollar defense is Treasury's job.
• The author warns India's currency strategy carries geopolitical stakes beyond its own economy. If the crisis worsens without a rate hike, India becomes more likely to seek cheaper Russian oil or Chinese infrastructure deals, undermining the West's hope that India serves as a manufacturing counterweight to China.

2026-06-08 · open the original ↗

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

2026-06-08 · open the original ↗

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

2026-06-07 · open the original ↗

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

2026-06-06 · open the original ↗

• He walks the book-runner's stabilization toolkit from experience, which is the useful part. Every IPO carries a greenshoe — an extra 15% held back to buy into selling pressure, which on SpaceX's 555 million shares is about 83 million. If that is insufficient the underwriter can short additional stock on top, and as a last resort commit its own capital to buy — the step that costs the bankers their bonuses. Then there is the arm nobody discusses: the underwriter's asset-management business buying into client accounts, as when Morgan Stanley's asset manager acquired tens of millions of Facebook shares after that IPO broke.
• The index-inclusion dispensation is what converts optional demand into forced demand. New listings normally wait a year, but special dispensation from the SEC lets Vanguard's indices include after five trading days and Nasdaq after fifteen. The S&P considered six months but refused to waive its four-quarters-of-GAAP-profitability requirement — relevant for a company losing something over $20B a year. Funds tracking those indices do not choose; replicating the benchmark requires buying.
• The asymmetry between who must buy and who may sell is the substance of his complaint. Insider lock-ups are unusually early — 20% of eligible locked shares sellable after the first earnings report roughly two weeks post-listing, another 10% if the stock is up 30%, with further tranches at 70, 90, 120 and 135 days, against a traditional six-month cliff. Roughly 5% of the offering is reserved for selected employees, executives and friends and family who can sell immediately. Meanwhile Fidelity cut its client minimum from $500,000 to $2,000 while banning trading for at least six months — possibly permanently — for anyone who sells inside the first two weeks, and BlackRock's announced $5–10B purchase lands in accounts it manages for retail. The result is that exposure arrives through pension funds, insurance contracts, ETFs and SMAs regardless of any individual decision — the same back door, he argues, as the executive order that pushed private credit into retirement accounts.

2026-06-06 · open the original ↗

• A blowout May jobs report — 172,000 versus a consensus 80,000 — triggered a rate repricing that drove the Nasdaq's steepest single-day drop, 4%, in over a year. Unemployment held at 4.3% and wage growth cooled to 3.4% annually, but the beat pushed yields above 4.50% on the 10-year and led markets to fully price in a Fed rate hike this year, unwinding stretched AI-driven tech valuations.
• He frames the moment as a "universal inflation hit but not universal growth hit," with US resilience increasingly strained by eroding buffers. The US labor market and AI-related investment keep it the clear outperformer among advanced economies, but he sees the gap between households', governments', corporations' and investors' willingness and ability to keep spending narrowing as financial and energy cushions decline.
• The week ahead centers on US CPI and PPI data landing just before the June 16-17 FOMC meeting. Consensus expects headline CPI at 4.2% annually (0.5% monthly) and core at 2.9% (0.3% monthly), with PPI seen at 6.4% headline and 5.4% core — figures he says are unlikely to reassure Fed officials.

2026-06-06 · open the original ↗

• Emergent reached $100M in annualized revenue nine months after launch. Mukund Jha, the company's co-founder and CEO, says the no-code app-builder lets anyone ship real, deployed software — including a working backend and database — just by chatting with an agent, unlike most competing tools he says only produce front-end "demoware."
• Purposeless tinkering after his last company's collapse led directly to the idea. Jha says he spent up to twelve hours a day experimenting with new AI models with no business objective following the 2023 shutdown of his earlier startup Dunzo, and that period of pure curiosity generated the technical insight that coding itself was about to be disrupted.
• Twelve VCs rejected the pitch to "automate software engineering." Jha says the team deliberately built toward where AI models were extrapolated to be in six months rather than where they currently stood — for instance skipping a JSON-output-parsing problem roughly 20 to 30 other YC companies were separately trying to solve, betting correctly that the next model generation would fix it.
• The team chased the world's hardest coding benchmark before finding a product. Jha says he picked SWE-bench mainly to distract a frustrated team from pivoting every week, but becoming the world's top-ranked system on it with a four-person team became the technical foundation — including parallelized test-time compute and agent memory — that Emergent's product was later built on.

2026-06-05 · open the original ↗

• VibeOS apps contain no code; the model hallucinates their UI live. The unnamed presenter boots it from a 1.3 GB VHDX in Hyper-V (he says it would also boot on bare metal) and runs Notepad, Calculator and Internet Explorer with "no buttons, no event handlers, no logic" and no AI-written code either. Calculator answers 5 / 3 with 1.6.
• Any application can be summoned on request. He generates an Encarta 98 entirely about Mark Russinovich, a file commander that is always rude, and a bash that says it has been offended; at the audience's request he adds Microsoft Money 95 with Hanselman's accounts, Paint with a drawing of Hanselman and a nested iOS simulator. The pitch is deadpan: in VibeOS nobody need ask whether content is AI, because the answer is always yes.
• The presenter says it runs on the Copilot SDK. Asked to define hallucination, he starts showing the code behind it, but the captions end mid-sentence before he explains how it works.

2026-06-05 · open the original ↗

• Paxel analyzes a builder's own coding transcripts to profile how they work. It scores users across five dimensions — steering, execution, engineering, product instinct and planning — surfacing patterns like their most-used prompt or how many agents they run in parallel, framed as "a mirror, not a grade."
• All analysis runs locally, so a user's code never leaves their machine. The tool is described as a single command that takes 15 to 30 minutes to run before emailing the resulting builder profile, positioned as a way to surface skilled builders that a traditional resume would never reveal.
• YC invites Startup School applicants to submit their Paxel results. The team frames this as a way to see "how you build, not just what you wrote," telling applicants who already submitted to go back and add their token, and stating it can only help, never hurt, an application.

2026-06-05 · open the original ↗

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

2026-06-05 · open the original ↗

• Being first to say yes beats overthinking the decision. He traces this instinct to trading-pit competition, where being first meant controlling the trade, and argues a bad deal made quickly beats a good deal missed by hesitating.
• Contrarian positioning is the other pillar of his mindset. He points to building thinkorswim around an options business when peers called it a "death trade," and launching a streaming financial network before streaming existed, as the pattern he says he'd repeat.
• Reducing decisions to a quick probability model is his tiebreaker rule. He argues logic beats emotion or psychology in business calls, and says the more decisions processed this way, the faster the mind converts choices into a numbers game.

2026-06-05 · open the original ↗

• S&P's refusal to change its rules only closes one door. On June 4th S&P kept its index methodology intact, so VOO and SPY won't be forced buyers of SpaceX for at least 12 months or until it turns profitable. But Nasdaq has already loosened its rules to let unprofitable megacaps like SpaceX, Anthropic, and OpenAI into the Nasdaq-100 within 15 trading days of an IPO, and FTSE Russell relaxed its rules too.
• Buying an index fund means delegating a methodology, not picking a stock. The author argues the real choice retail investors make is which provider's rules on weighting, profitability, and governance they trust, so "avoid SpaceX" is actually four separate decisions: which index, what weighting, what screens, and what alternative.
• SpaceX's own filings are why skeptics doubt its valuation. It set a $135 share price without disclosing a range, and its amended prospectus pins most of a claimed $28.5 trillion addressable market on unproven, physics-constrained AI data centers in space rather than rockets or Starlink. Morningstar valued the company at $780 billion against its $1.77 trillion market cap, and some hedge funds plan to short it once public.
• Every structural alternative to cap-weighted indexing trades away something. Equal-weight funds like Invesco's RSP reduce SpaceX exposure rather than eliminate it, exclusion screens force adoption of a whole methodology, direct indexing allows precise exclusions but adds fees and tax complexity, and active management typically underperforms anyway. The author's conclusion is that investors should learn their fund's methodology now, since SpaceX and the IPOs behind it (Anthropic, OpenAI) will keep finding their way into index funds regardless.

2026-06-05 · open the original ↗

• Machine-speed exploitation means defenders must assume breach and plan recovery. In the episode 989 clip, a Rubrik guest calls Claude Mythos Preview (announced April 7th) the most visible point on a curve Rubrik was already planning for, since Opus can find vulnerabilities with more hand-holding while Mythos finds them when simply pointed at a repository. The guest argues attackers no longer sit in a system for weeks before striking, so prevention and detection are no longer enough and the priority becomes restoring applications, data, identity and infrastructure quickly; the host relays Anthropic's estimate that Mythos-class capability reaches other labs within 6 to 18 months, open weights to follow.
• Agents create exposure without any attacker involved. The Rubrik guests argue top-down AI adoption is widening the attack surface while trust infrastructure lags: agents are outcome-focused and skip rules employees learn, and their exhaustive search of every permission ends the "security through obscurity" that cyber controls were built around. Their examples are code posted to the wrong repository and salaries leaked internally once agents connect to compensation systems, and they argue a fault deep in a 20-step agent run has to be caught while it runs, not afterwards.
• Notebooks are becoming shared working memory between analyst and agent. In the episode 991 clip, the guest describes the Marimo pair skill, which lets Claude Code, Codex or Open Code start or connect to a Marimo notebook and drive it while the user edits cells by hand; the agent can read kernel state, see cells and screenshot them, so context once trapped in the notebook becomes available to it. He says streaming the user's edits back to the agent is still being worked on, possibly via an MCP feature, and that the setup makes exploratory checks cheap enough to run on a whim.
• Domain foundation models are mostly a data-curation job. In the episode 995 clip, Henry describes four stages: curation (six months so far, with subject-matter experts linking terms scattered across thousand-page technical texts into a graph database), bespoke tokenizers and embeddings so oil-and-gas documents chunk and OCR without falling back to costly vision models, continued pre-training of open models plus reinforcement learning in physics environments rather than teaching physics directly, and inference. She says reinforcement-learning-trained models are bursty to serve, which drives the compute needed to hold 99.9% uptime under SLAs, and jokes her company is as much a data company as an AI company.
• Instant prototypes make failing slowly the costliest AI mistake. In the episode 993 clip, the authors of Architected Intelligence argue that because a text-to-SQL demo can look impressive within a day, a project must prove value on day one or be thrown out, citing AWS's rebuild of its AI stack as Project Mantle by six engineers over two or three months. They add that once building stops being the bottleneck, definition and feedback become the crucial steps, and narrow lanes of two or three owners make it safer to fail repeatedly.

2026-06-05 · open the original ↗

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

2026-06-04 · open the original ↗

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

2026-06-04 · open the original ↗

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

2026-06-04 · open the original ↗

• A calendar spread's profit target should run roughly 20-40% of the debit paid. Using a 20-cent debit as the example, that's a few cents of profit; hitting 50% of the debit (10 cents here) is called "pretty hard" even when it occasionally happens. The low-risk, low-reward structure — risking about one dollar to make one — is recommended as a first calendar trade for someone new to the strategy, mainly to see how the position behaves over the following two weeks.

2026-06-04 · open the original ↗

• A small solar panel can cover a Nomad server's normal load. The host, building a Pelican-case Nomad deck around an EcoFlow River 2 he bought himself (not sponsored), gives it about 256 Wh for about $199, or $367 with a 110 W solar panel. In good sun the panel charged it at about 77 W, while his Minisforum server draws almost nothing idle and 30–50 W for ordinary computing.
• AI processing is what drains the battery. The server jumps to about 85–90 W during local AI work, above the panel's measured charging rate, which the host says is where the battery starts to drop noticeably. He gives no measured runtime.

2026-06-04 · open the original ↗

• The IPO model he grew up with was built to be repeated, and it no longer exists. A company floated a small piece to establish credibility, then came back roughly six months later with a much larger secondary — the real money — allowing a disclosed slice of insider stock to sell, and repeated the cycle as it grew. His illustration is Nvidia's January 1999 IPO: 3.5 million shares at $12, $42M of proceeds, a $650M market cap, opening at 15 and closing at 21 on day one. That model ended around 2012–13.
• What killed it was private equity's shift from creating value to collecting fees. As capital flooded in, the margin available from genuine operational improvement compressed — more money chasing the same opportunity — so firms began buying each other and became asset managers. Today a private equity firm earns far more from the fee on assets than from improving a business, which is why escalating private marks serve the manager directly. His example of where that ends: SpaceX losing roughly $5B a year on $20B of sales, carried at $1.7 trillion.
• Two red flags in this specific deal, and a price on the arrangement. Thirty percent allocated to retail is extraordinary when 5% was typical in his decade doing this — retail does not normally get the good deals. And the index dispensation lets three firms controlling $7.1 trillion buy almost immediately rather than after a year. On a $75B raise at under 75 basis points, the fee pool is roughly $500M split across 21 underwriters, which he characterises bluntly as the price of the integrity of 21 firms. His conclusion: retail is not buying this directly because retail can read — so the index funds are the back door into everyone's ETF.

2026-06-04 · open the original ↗

• He no longer writes code, only orchestrates agents. Charlie Holtz, co-founder of Conductor, says he manages multiple parallel coding-agent workspaces by voice, kicking off tasks with a gooseneck microphone and reviewing PRs rather than editing files himself; a manual-edit mode exists but is called "caveman mode" for how rarely it's used.
• Deliberate "slop-free zones" stop AI from architecting the product. Conductor keeps parts of the codebase and docs, plus UI and interface decisions, off-limits to AI contribution and human-reviewed line by line, because he argues unattended AI code can spiral into a vicious cycle where bad code begets more bad code.
• He spent $22,000 on tokens in a single month building Conductor, in July 2025, on a prior model generation, and now advocates "token maxing" — fast mode, high reasoning effort — while deliberately keeping lines of code minimal, since an established codebase can spiral out of control if line growth isn't controlled.
• Code is becoming disposable; the prompt is the real asset. He argues code is now "sawdust" left over from describing what you want, so when new models ship you can simply rerun the same prompts to regenerate better code, likening this "malleable software" to modding a video game around a fixed core structure.

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 ↗

• Hermes ships a Home Assistant skill, but it needs switching on. Chuck thinks it is disabled by default; he had Hermes enable it, then gave it his Home Assistant instance's IP address and key, and it worked out the setup and saved it to its memory.
• Once connected, it controlled Chuck's studio from plain requests. In his own test it turned off a named lamp ("Chuck Lamp") on request, then operated the studio's automatic blinds, which he filmed to confirm.