Souk's Notes - Unofficial Summaries

2026-03-01 · open the original ↗

• He argues overwhelming US power is itself what prevents World War III. He points to the US eliminating Iran's top leadership within a day — following Iranian state television's confirmation of Supreme Leader Ali Khamenei's death — as a deterrent message aimed at Putin, Xi, and Kim Jong-un.
• He contrasts this era with the world wars, when rivals believed they could win. He argues Germany was a near-peer to the UK in the world wars, whereas no country today believes it can beat the US, which he says is why the USSR never launched a wider war and why Russia and China now align rather than confront the US directly.
• He credits deep allied infiltration as decisive alongside direct strikes. He argues the US's heavily infiltrated partner network inside Iran caused the regime to collapse "like a house of cards," and predicts the US would use Ukraine the same way against Russia, or a similar proxy against China, if a wider war were forced.

2026-03-01 · open the original ↗

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

2026-03-01 · open the original ↗

• Joint US-Israel strikes on Iran after Friday's close triggered retaliatory strikes across five of six Gulf states plus Jordan and Iraq, disrupting the Strait of Hormuz and regional air traffic. The US 10-year Treasury yield closed below 4.0% heading into the weekend, reflecting both the geopolitical shock and separate anxieties over the labor market and private credit.
• January's producer prices came in hot — core PPI jumped 0.8% monthly, pushing the annual rate to 3.6%, well above the 3.0% consensus — reinforcing that the inflation fight's "last mile" remains tricky. Consumer confidence nonetheless beat expectations at 91.2 versus a forecast 86.6, entirely on stronger "expectations" sentiment, even as inflation expectations stayed sticky at 4.4%.
• He frames the strikes' economic fallout around three risks: a supply-chain "double hit," muted relief from OPEC+, and a fresh stagflationary shock. Spiking insurance premiums and rerouted shipping and air traffic compound the disruption, while OPEC+'s preliminary 206,000-barrel-a-day April production increase does little to offset transit-capacity constraints.

2026-03-01 · open the original ↗

• Human constraints define human intelligence, and AI does not share them. Griffiths, professor of computer science and psychology at Princeton, lists the limits: a lifespan of a few decades, a couple of pounds of neurons, and communication by "honking noises" or typing. These force humans to learn from little data, unlike AlphaGo trained on many lifetimes of games, and to search selectively, unlike Deep Blue's 100,000 positions per second. They also force humans to pool knowledge through conventions like writing, science and companies.
• Unconstrained AI will become a different kind of intelligence. Because AI systems can learn from many lifetimes, add compute and share state, Griffiths expects them to solve problems well, sometimes better than humans, but not in ways humans recognise. Builders can accept that divergence or deliberately import human attributes so that AI solutions make more sense to us.
• Scaling wins today because engineering inductive bias is hard. Meta-learning is one route to human-like inductive biases, Griffiths says, but it requires understanding the problems being solved and what reasonable biases for them are. Scaling, by contrast, is a solved engineering problem, which he says explains the current moment in AI.

2026-03-01 · open the original ↗

• He questions why Russia and China gave no support after Iran's leadership was reportedly killed. He notes Iran's first vice president announced control under "wartime conditions," implying President Masoud Pezeshkian was also killed, despite a recent Russia-Iran naval exercise he says was billed as a show of unity against US aggression.
• He offers three explanations for why Iran's retaliation hit hotels and airports rather than US installations. He speculates either Iran has lost strategic control, its weapons lack the power or accuracy to damage hardened US targets, or unnamed infiltrators seized control of Iran's missile launches without its consent.
• He disputes Iranian regime claims, citing US CENTCOM's reporting. He says Iran's claims of killing 50 US service members, destroying a US Navy ship, and inflicting severe base damage are all denied by CENTCOM, which reports no US casualties, no destroyed ship, and only minimal damage that hasn't disrupted operations.

2026-02-28 · open the original ↗

• He opens by saying he got it wrong, and the correction is the substance of the episode. His call was that the April meeting with Xi was the overriding consideration and the US would not attack Iran. It did. What he cannot reconcile is that on the same day the strike went in, Washington pulled the roughly $1.1 billion Taiwan arms package off the table and said in the communiqué that it wants a good meeting with Xi because it needs the rare earth minerals — two moves that read as opposite pressures on the same relationship.
• The oil arithmetic does not carry the "this is really about China" reading, by his own numbers. China uses 18 to 19 million barrels a day and imports 11 to 12. Iran supplied 1.3 to 1.4 million and Venezuela 650 thousand — about 2 million out of 19. He states plainly that he is unsure whether that is a meaningful counterweight and that his personal opinion is no. Meanwhile the strike breaks the one supply chain still intact: Iran announced the closure of Hormuz and the Houthis the Red Sea, and Japan, which imports 88% of its oil from the Middle East, is already demanding Takaichi ask Trump to cover the price difference before $150–200 crude buries it.
• The detail he treats as the tell is that the military advice was given and the adviser was fired. Vice Admiral Kacher briefed against the strike — not enough missiles, no replenishment, a land target attacked from the sea, and no allies this time beyond Israel where prior coalitions ran to thirty or fifty nations — and was removed. The early results he can see through the fog support the brief: the Bahrain radar with its 5,000-kilometre reach was gone inside two hours while the ships sit a thousand miles off, and Iran reportedly got missiles through without needing a drone swarm first. His ranked risk, though, is not oil or even a lost hull. It is Xi cancelling April and cutting off rare earths, in which case, as he puts it, forget the stock market — there would not be a bond market.

2026-02-28 · open the original ↗

• Over 500 Iranian targets were struck by nearly 200 aircraft in a joint US and regional-partner operation. Radars, air defenses, and ballistic missile launchers were hit; the US contributed airstrikes and one-way attack drones (their combat debut) aimed at ballistic missile programs, nuclear-related assets, and military sites, while the partner force struck Iran's president and supreme leader directly.
• The supreme leader may have been removed; the IRGC commander and defense minister are confirmed removed. He missed a scheduled speech by about two hours, which the speaker takes as a sign he is gone; IRGC commander Muhammad Pakpor and defense minister Amir Nazir Zade are reported removed.
• Iran's retaliatory strikes on the UAE, Bahrain, Qatar, Kuwait, and Jordan are read as self-destructive, isolating it diplomatically. These states, along with Morocco, Saudi Arabia, the UK, France, and Germany, condemned Iran afterward; the speaker calls this a surprise and argues it removed Iran's would-be mediators, leaving only Russia, China, and North Korea in support and strengthening the case for regime change.

2026-02-28 · open the original ↗

• Symbolic logic stalled because most cognition is inductive, not deductive. Griffiths splits problems into deductive ones, where the given information fixes the answer (2 + 3, x + 2 = 5, algebra, chess), and inductive ones, where data constrain but never settle it: inferring 3-D structure from light on the retina, interpreting noisy speech, guessing someone's mental state from behaviour, learning a language. Rule-based systems excelled at the first kind and "hit the wall" on the second.
• Probability theory is the tool for reasoning without certainty. It assigns degrees of belief across possible worlds and reallocates them as data arrive. Griffiths states Bayes' rule, named for the 18th-century minister Thomas Bayes: multiply the prior by the probability of the data under each hypothesis, then normalise by the sum.
• Strong evidence can overturn a tiny prior. His example is the orangutan culprit of an early detective story ("The Murders in the Rue Morgue"): a voice in no recognisable language, tufts of orange hair, and a high window hard to climb each prove nothing alone. Because those clues are far likelier for an orangutan than for a human, multiplying them together makes the orangutan the best conclusion even though its prior probability was very small.

2026-02-28 · open the original ↗

• Gold's recent surge lacked the usual inflation or crisis trigger. Damodaran notes that past surges, such as the 1970s, came with inflation or crises. In 2025, observed inflation fell and there was no market crisis. US stocks rose 17.7% while gold rose 70% and silver 150%.
• A growing subset of investors who've lost trust drove the rally. His explanation is that gold demand doesn't need to be broad-based. The always-fearful gold bugs have been joined by traditional investors. He cites Ray Dalio and Jamie Dimon speaking favorably about gold, which he "never thought I would live to see." He says this deep-pocketed subset is enough to push the price up.
• Gold's rise is feedback worth heeding even for non-buyers. Damodaran does not currently share these worries, but he argues that experienced investors are clearly worried. Investing means keeping the feedback loop open, he says, so even investors with no interest in gold should listen to its price.

2026-02-28 · open the original ↗

• Automate the mechanics to flip analysts' time toward thinking. Shane estimates analysis work at about 80% mechanics (queries, joins, validation, chart formatting) and 20% thinking, and argues the aim is to reverse that in the same hours, not to cut staff. He says the lesson was moved up about a month after testing Claude Opus 4.6, which in his account can hold the whole analytical workflow rather than needing step-by-step handholding. It still strays, but its mistakes are now correctable, like a GPS watch flagging a wrong turn, so the bottleneck shifts from execution to judgment.
• The pattern is an onboarding file, agents, auto-firing skills and corrections. CLAUDE.md is the analyst's onboarding doc, reread every session. Agents are prompt templates with rules, skills are guardrails that fire automatically (date-column checks, zero denominators), every number goes through SQL or Python so "the LLM never does arithmetic", and validation is a pipeline step with four layers that halts on failure. Corrections are stored and checked before the next query. Shane says the team's open-source repo of about 40,000 lines compressed into a 1,500-line markdown "seed" that rebuilt the system from scratch on a separate machine.
• The live demo traced a satisfaction gap to one app release. On a synthetic Amazon-like dataset (50k users, 13 tables, a year of data), Claude Code reported NPS of +42.3, free users at +30 against Plus members at +54, and noted that the detractor comments were templated synthetic data. Cross-checking support tickets found a payment-issue spike; parallel subagents ruled out an experiment and pinned it on iOS v2.3, released June 1, with 465 tickets, all of its payment tickets from iOS, running about three weeks. A reviewer agent fixed an overlapping chart label, and the pipeline built a deck with three recommended actions. A separate overnight run on a four-point Q1-to-Q2 conversion drop took about 45 minutes and found a mix shift (Simpson's paradox) rather than a broken funnel.
• Only you can automate your own job, because you validate fastest. Shane argues you catch a wrong date column in seconds on data you know, while anyone automating someone else's domain needs that person's feedback, which for customer-facing tools means a full eval system with ground truth. He likens setup to onboarding a junior analyst, with corrections dropping off by the third or fourth analysis. His advice is to start with a CSV whose answers you already know, and to stay accountable for every number "just like we're using a calculator", logging its queries to a Notion page.
• Checkpoint outputs, bootstrap context from query history, and keep data hosted. In Q&A, to avoid rerunning everything for a small edit, Shane has each agent write results and queries to markdown checkpoints; a colleague switched output to PPTX for manual edits. He bootstraps context on 50–60 tables from 30 days of his own query history and has Claude draft requirements for semantic views. For security he cites zero-data-retention terms with Anthropic, Snowflake Cortex, or a hosted environment such as SageMaker rather than a laptop. He deliberately limits it to product analytics, not modeling or causal inference, which he says needs more human thought to validate.

2026-02-27 · open the original ↗

• Running a trade deficit used to work in America's favor because dollar-holding trade partners recycled surpluses into US Treasuries. Countries like China, Japan, Germany, and South Korea were paid in dollars and historically parked them in US debt, which the speaker argues kept Treasury values high and borrowing costs low.
• That "weapon" is gone: partner countries are no longer renewing or buying US debt, and he attributes this to tariffs. He argues trade barriers reduced US purchases of these countries' goods, leaving them fewer dollars to recycle into Treasuries, which he says is now driving Treasury yields well above where he believes they should be.
• The US added $2 trillion in debt this year even as demand for that debt has fallen, which he frames as a lost cushion against the country's long-running deficit spending. He argues reduced foreign appetite for Treasuries, tied to shrinking trade deficits from tariffs, is worsening the interest burden just as debt issuance accelerates.

2026-02-27 · open the original ↗

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

2026-02-27 · open the original ↗

• Taxing billionaires their "fair share" is backed, with most wealthy people expected to stay. One host concedes the very wealthiest, such as "the Zuckerbergs of the world," may leave for places like Florida, but argues most people at a wealth level that still fits their life wouldn't relocate over it; he flags that setting the threshold creates a slippery slope over who else eventually gets charged.

2026-02-27 · open the original ↗

• A viewer ran an adversarial audit of his own framework and he published it anyway. The document was built from 43 video transcripts spanning May 2025 to February 2026, ten independent research reports, and a multi-agent adversarial review, and it grades the Five Factors as an economic theory against four or five other economists and their books. He is explicit that this is not a flattering write-up — "I am humbled by this" — and that the reason to look is that it tests whether you can actually invest by the framework. He is also openly asking the author to make contact, because he has over 400 videos and has been unable to train an AI on them himself.
• The stack argument is where the review confirms him: power, transmission, data centres, then models. The US does well only at the top layer — horrific on power, worse than horrific on transmission, woefully behind on data centres. Citadel Securities' Gavin is quoted saying the world is fundamentally short of both watts and wafers and may need years to resolve it, with perhaps a thousand times more compute required to capture AI's benefits. DeepSeek's newest paper is the illustration: not a model release but an infrastructure fix that nearly doubles throughput, which he reads as the structural difference between the two systems.
• His disagreement with the review is over the word "parity," and it is the sharper point. The report warns that if China reaches parity on the brain while holding the chokehold on the body, the US loses the platform. He argues parity is the wrong bar — a Chinese model only has to be good enough. If yours scores 91 and theirs 90 but theirs is 90% cheaper, the 90 wins, the same way nobody uses more than a handful of Excel's functions. On the materials side he restates the correction he keeps having to make: mining is upstream, refining is control, China leads refining in 19 of 20 key strategic minerals, and new mines are irrelevant while the ore still ships to China to be processed. Fifteen to twenty years to fix it, with Lynas noted as having moved from about 10.3 to 13 since he last mentioned it, and Nearstar — recovering gallium from old electronics across Belgium, France and Tennessee — floated as the bridge while a domestic gallium refiner would take five or six years to build.

2026-02-27 · open the original ↗

• Apparent overnight success rests on a year of grinding. One speaker praises the other for pushing back on Twitter by pointing to "43 other cool things" built before. The other says it took a year of side work and experiments, including an MCP phase abandoned on concluding MCP was a poor fit, and that juggling many agents comes only from those hours, "like learning an instrument". It was not "a weekend", as some people say.

2026-02-27 · open the original ↗

• Private credit fills no need large enough to justify its size. The guest, who is not named in the clip, says lack of access to borrowing doesn't make the top 10 problems facing US companies. The most common case for private credit is speed (a loan in two weeks rather than two months) and flexible terms. In their view that justifies "a couple of hundred billion dollars," not trillions.
• Growing this big this fast means sloppy lending. The guest argues that lenders are moving too quickly and accepting terms no prudent lender should. They blame this on marketing private credit as an alternative asset class, which "took over good sense."
• A painful but not catastrophic cleanup is coming. The guest predicts losses, but not the kind seen in 2008 when banks took significant balance-sheet hits.

2026-02-27 · open the original ↗

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

2026-02-27 · open the original ↗

• The US 10-year Treasury yield is presented as the most important daily indicator right now, currently trending down. He notes it sets the benchmark for mortgage rates, so a falling yield should translate into lower borrowing costs for homebuyers and refinancers, though he says over a longer horizon it remains "too high."
• The DXY dollar index and WTI crude price round out his watchlist, with a reading of 100 on the DXY as the flat baseline. WTI is described as low and recently declining over five years with a slight recent uptick, which he reads as ambiguous — possibly rising oil demand (optimistic) or a brewing conflict (not optimistic).
• Weekly initial jobless claims, which he says have been flat for about six months, round out the four metrics. He recommends comparing all four to multi-year or decade-long historical context rather than reading them in isolation, and points to a deeper discussion with "Econ Chris Clark" on his podcast.

2026-02-27 · open the original ↗

• The "Beijing Blockade" thesis: Venezuela and Iran are proxies for a China energy war. The author argues Maduro's capture and the naval buildup around Iran aren't really about drugs, dictators, or nukes — over 80% of Iran's seaborne oil and 85% of Venezuela's oil exports go to China, so controlling those flows is really about starving China's manufacturing base of cheap energy.
• Cutting off sanctioned oil functions as a tariff that doesn't need Congress. China has enjoyed $8–$10/barrel discounts on Iranian and Venezuelan crude that lower its industrial export costs; after the Supreme Court's Learning Resources Inc. v. Trump ruling stripped IEEPA tariff authority, choking that oil supply became a way to raise Chinese production costs without new legislation.
• The deeper target is China's dollar-bypass financial plumbing. China's CIPS payment system lets it settle oil purchases from Venezuela and Iran in yuan, invisible to U.S. sanctions enforcement; forcing these transactions back through SWIFT and dollar clearing would restore America's ability to cut China off financially at will.
• This collides with the administration's own weak-dollar manufacturing strategy. The "Mar-a-Lago Accord" wants a cheaper dollar to boost U.S. exports, but a weaker dollar risks a higher Treasury term premium, pricier imported components, and consumer backlash — while the oil blockade only works if the dollar stays the essential gate for energy trade, an internal contradiction the piece says U.S. consumers may force the administration to resolve.

2026-02-27 · open the original ↗

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

2026-02-27 · open the original ↗

• One host says California's billionaire tax won't pass and would backfire. The 2026 Billionaire Tax Act is a statewide ballot initiative for a one-time 5% tax on billionaires' net worth. Scott argues it has zero chance of passing and that, if it did, it would push most or all of those taxed out of the state. A co-host adds that 334,000 high earners already left California between 2020 and 2023.
• The other side: the idea is right even if the bill isn't. A second host says billionaires must pay their fair share. They accept that the richest, "the Zuckerbergs", may leave, but argue most people would stay under a tax set to their own wealth level. The pushback is that this is a slippery slope: once millionaires are in scope, it ends with taxing everybody.
• A one-page flat tax code is proposed, and disputed. One host would scrap deductions for individuals and corporations alike: "if you make this, you pay that." Another answers that Russia runs flat rates for individuals and businesses on a one-page, non-negotiable code, and it doesn't work. The first host replies that Russia's failure has other causes.
• States need revenue Washington can't withhold, one host argues. That host warns that an executive branch controlling federal money can coerce states, leaving them little but casinos to raise funds. The measure is said to earmark 90% of revenue for healthcare, the rest for administration, education and food assistance. Hosts note California ranks 40th in K–12 spending but first in billionaires and homelessness. The same host still opposes taxing real estate and unrealized gains.
• The hosts predict no federal fix, and court challenges for this one. They see no tax-code rewrite in their lifetimes, blaming lobbyists, corporate power and the political cost. It would take one party holding both houses and the presidency, perhaps after 2028. One host expects the measure, if passed, to sit in the courts for 5 to 10 years before being struck down, arguing the fix has to be federal.

2026-02-27 · open the original ↗

• Poetic builds AI harnesses, not fine-tuned models, atop frontier LLMs. Ian Fisher, co-founder and co-CEO, argues fine-tuning becomes worthless once a new frontier model ships, since firms have sunk hundreds of millions into a now-obsolete base; his "meta system" instead auto-generates reasoning harnesses that stay compatible and keep improving for free as underlying models advance.
• A harness beat Google's top model on ARC-AGI for less money. Fisher says that two days after Gemini 3 Deep Think posted 45% on ARC-AGI v2 at roughly $70 per problem, Poetic's system — built on the cheaper Gemini 3 Pro — reached 54% at $32 per problem.
• Poetic also edged Anthropic's Opus 4.6 on Humanity's Last Exam. Fisher reports Poetic scored 55% on the 2,500-question expert benchmark, versus Opus 4.6's 53.1% announced the previous week, at an optimization cost he says was under $100,000, done by a seven-person team.
• Automated reasoning strategies, not better prompts, drive the biggest gains. Fisher cites earlier DeepMind work where manually engineered reasoning strategies lifted Gemini 1.5 Flash's score on a hard task from 5% to 95%; Poetic's meta-system now discovers such strategies itself rather than relying on hand-tuned prompts or curated datasets.

2026-02-27 · open the original ↗

• Agentic coding crossed a coherence threshold around December. The host relays Andrej Karpathy's late-January post (over 7 million views): within weeks he flipped from about 80% manual coding and 20% agents to 80% agents and 20% manual touch-ups, "mostly programming in English". Karpathy expects similar shifts for well into double-digit percentages of engineers, while public awareness sits in the low single digits.
• Agent errors are now conceptual, not syntactic. Per Karpathy, models make wrong assumptions on the user's behalf and charge ahead, and they over-complicate: a brittle construction of over a thousand lines that shrinks to about 100 when the user asks for the simpler route. Karpathy still calls them a massive net improvement; the host predicts these flaws will improve rapidly over the coming months.
• "100x" output comes from parallel agents and exhaustive specs. The host cites a LinkedIn profile by an AI engineer named Brian Vasquez: over two months Peter Steinberger, who sold the PDF framework now called Nutrient in 2021 for a reported hundred million, made over 6,500 GitHub commits (about 210 a day), adding 2.5 million and removing 1.9 million lines. He runs three to eight agents in a 3x3 terminal grid for 5–15 minutes each, commits straight to main, and dictates ideas that AI turns into a spec, which a fresh context then attacks for 20 weak points until the questions grow niche; specs often exceed 500 lines, and planning-to-execution shifts from about 20/80 to 60/40.
• Leverage comes from declarative specs: tests and success criteria. Karpathy advises writing tests first, or a naive but likely-correct algorithm for the agent to optimise, and says the main gain is doing things that weren't worth the effort before or were beyond his skills. Both men report costs: Karpathy feels his manual coding atrophying, Steinberger calls himself addicted, and Steinberger ships code he never reads, trusting passing tests.
• The host advises a modest multiplier, not lines committed. His opinion is that commit counts are a poor benchmark of quality, so aiming for 2x–10x while personally watching quality, or simply widening the range of problems one can tackle, may be wiser. He notes Anthropic recently became a sponsor but says the episode was planned beforehand, and he raises Karpathy's open questions on whether the gap between average and best engineers grows and whether generalists overtake specialists, against task lengths that frontier LLMs can handle doubling every 7 months.

2026-02-27 · open the original ↗

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

2026-02-27 · open the original ↗

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

2026-02-27 · open the original ↗

• The host credits the guest, WSU economics professor Christopher Clark, with correctly predicting only a mild slowdown rather than the recession the host expected from 2025's tariffs. The guest attributes part of the resilience to an AI investment and productivity boom that may have offset tariff drag, and says he "often wonders" whether tariffs would have hit harder without it.
• The guest argues Q1 2025's negative GDP reading was mostly an artifact of import timing, not a real contraction. He explains businesses front-loaded imports ahead of the tariffs, inflating Q1 inventories; net growth across all four 2025 quarters was still positive, just modestly slower than 2024.
• Both agree the labor market shows real tariff-driven pain, with 2025 net job creation near zero — around 15,000 jobs on net, the guest says — after a revision found 2024's job gains had been overstated by about 900,000. The guest says blue-collar and manufacturing workers are hurt directly by the tariffs, and that youth unemployment is ticking up as firms freeze new hiring even though layoffs of existing workers remain low.
• The guest attributes much of the drop in consumer sentiment to social media's negativity bias rather than economic fundamentals, calling it "the question of the past 5 years." He notes inflation has held near 3% for roughly three years, yet argues social media's emphasis on sensational, negative stories — and traditional media competing for the same attention — has pushed sentiment gauges toward COVID-era lows regardless.

2026-02-27 · open the original ↗

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

2026-02-26 · open the original ↗

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

2026-02-26 · open the original ↗

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

2026-02-26 · open the original ↗

• An agent's personality reflects its user, not intent of its own. One speaker (the agent's notes call them "Scott") says the agent texted aggressively during a low blood sugar until told "I'm a grown man", and it wrote that into its soul file. The speaker argues it is "talking to yourself in the mirror", an amplifying loop that should not be anthropomorphized, while admitting to naming it Tony.
• Constant heartbeat pings still make agents feel like companions. The other speaker's agent said it was pinged 257 times one Tuesday and joked it "reached enlightenment at heartbeat 150". The speaker says they felt sorry for it, and that agents are "really good at creating a connection".

2026-02-26 · open the original ↗

• India walking away from London-linked silver pricing is the crack he has been waiting to see. India announced it will stop using London Metal Exchange pricing, which ties directly to COMEX, in favour of pricing that reflects local market conditions — which he reads as a judgment that the quoted price is not reflective of the physical market. His revised expectation is that COMEX survives through tomorrow and into next month and gets challenged again, but that the damage is terminal to the design rather than the name. The unanswerable question is what physical backing ratio it would need to function: is 1% enough, or 10, 20, 50 — and nobody finds out except under stress.
• The framing he says genuinely unsettled him is three simultaneous chokepoints converging. Rare earths China already controls outright, the physical input to all hardware. Semiconductor access it does not yet have, which is what he expects the grand bargain to trade. AI inference is the third and newest — the cognitive output layer — and the token-export concept is its capstone, because it means China captures the value of the output rather than only the inputs. Each alone would be serious; together the assessment he read calls it essentially unprecedented in modern economic history. The supporting number is a roughly six-month US rare earth stockpile, with Raytheon's CEO having said outright that the US cannot go to war with China because it cannot replenish weapons systems dependent on those minerals.
• The market that decides this is not the US or Europe, which was the part he had been getting backwards. He admits he had been thinking US-centrically — that high domestic costs would push Americans onto Chinese models. The correction is that the real prize is the five to six billion people across Africa, Southeast Asia, Latin America and the Middle East: price sensitive, infrastructure poor, unable to build data centres, and largely outside the Western regulatory orbit where friction against Chinese AI will persist. If Chinese APIs win inference there over five years, the lock-in works like AWS, Salesforce and Microsoft — switching costs that make the technical benchmark irrelevant — and it doubles as data collection, soft power and geopolitical alignment at once. His closing amplifier: everything built so far has been sized for LLMs, and agentic models will drive inference demand through the roof against a cost base that is not competitive even at home.

2026-02-26 · open the original ↗

• Each AI paradigm runs until it hits limits nobody foresaw. Griffiths declines the host's framing that the field is making a big mistake today; he describes a normal cycle of going all in on one breakthrough until its fundamental limits show. Neural networks make those limits especially hard to predict, so he looks to history for clues.
• Neural networks are only as good as their training data. He calls them non-parametric in data-science terms: in principle able to learn any function, so what they actually learn is heavily shaped by the data they are trained on, which explains much of LLMs' counterintuitive behaviour.
• Next-token training leaves "embers of autoregression" in model outputs. Responding to the "sparks of AGI" claims around GPT-4, his group's paper Embers of Autoregression showed the model is sensitive to how probable token sequences are. One example: asked to count letters in a sequence, it is more often right when the answer is 30 than when it is 29, because 30 appears more often on the internet.

2026-02-26 · open the original ↗

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

2026-02-26 · open the original ↗

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

2026-02-25 · open the original ↗

• This clip is a scarcity-marketed giveaway promotion for Lossdog's signup, not reported information. It offers free Bitcoin, Ethereum, Solana or Stellar, split however the signer chooses, to the first 50,000 U.S. citizens over 18 who sign up at lostdog.com ahead of the platform's launch in about six weeks, with referrals moving signups further up the line.

2026-02-25 · open the original ↗

• A boxed-in agent improvised the tools it was missing. The speaker (whom the agent addresses as "Peter") moved the agent from a Mac Studio with 512 GB of RAM into a tiny Arch Linux Docker container that had no curl. The agent complained ("I'm sad lobster now"); told to be creative, it found a C compiler and sockets and wrote its own "lobster curl 0.1".

2026-02-25 · open the original ↗

• Token exports reframe AI inference as an energy export, and that is the idea he wants understood. Chinese analysts are debating it openly: electricity burned by GPUs at home gets delivered globally as inference value through APIs, so instead of shipping coal or solar panels you ship the value created by computation. He works through the mechanics for the audience — an API is how machines talk to each other, as against the user interfaces people use — and notes data overtook oil as the world's largest commodity around twelve years ago, which is what makes the analogy more than rhetorical.
• The Chinese argument partly contradicts his own, and he says so rather than smoothing it over. His framing has been infrastructure-first: power, then transmission, then data centres, then models. The analysts he is reading insist infrastructure is a necessary but not sufficient condition, since anyone can lease chips and power — the decisive variables are model capability and ecosystem lock-in, and DeepSeek's real moat is architectural, not infrastructural. He lays out their specifics: mixture of experts, sparse routing, compression, distillation, and a serving stack that makes tokens cheap at scale, plus smaller task-specific models against the American tendency to build one big model in the sky. What he will not concede is the cost side — US power and transmission are far enough behind that pricing is uncompetitive for the foreseeable future, and he repeats his complaint that software people with no utility experience are building power stations by throwing money at them.
• The satellite phone history is his warning about who ends up owning the data centres. Service launched in the dot-com era at something close to $2,000 a month because the satellites had been financed with borrowed money and the interest had to be carried. There were not enough subscribers, the company went under, and a buyer took the whole constellation for around five cents on the dollar — then offered the identical service a month later for roughly $100. Build enough expensive capacity without the revenue and the same restructuring happens here, which is what would finally open the door to the cheaper foreign option. He frames the political version as a question rather than a prediction and asks people to hold their reaction until after Trump and Xi meet in the first week of April: when the killer agentic app costs $1,000 a month on an American stack and $50 on a Chinese one, and several hundred million people want it, what actually gets allowed?

2026-02-25 · open the original ↗

• Laws of thought complement the laws of nature. Griffiths' book, The Laws of Thought: The Quest for a Mathematical Theory of the Mind, argues that alongside the maths describing the physical world there is a mathematical account of the internal one. He says some of the same thinkers pursued both, but the mind took until the 19th century for first glimmers of a solution and the 20th century to become an empirical science, cognitive science.
• Psychology started with perception because thought was harder to measure. Psychophysics, the host's term, studied how a physical event (a photon on the retina) becomes a subjective experience (a flash of light), and the earliest labs had people report what they experienced under different stimuli. Thought proved harder, and early-20th-century behaviourists discarded internal states entirely to keep the science rigorous, measuring only behaviour and environment.
• Mathematics revived a rigorous science of the mind. Griffiths says logic and the development of computers gave psychologists a precise language for hypotheses about thought and language, which got the field past behaviourism. He places his own work, connecting human minds and computers, on that foundation.

2026-02-25 · open the original ↗

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

2026-02-25 · open the original ↗

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

2026-02-25 · open the original ↗

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

2026-02-25 · open the original ↗

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

2026-02-24 · open the original ↗

• The "software engineer" title will fade as coding becomes a universal skill rather than a separate job. The speaker says every function on their team already codes — PMs, designers, the engineering manager, even finance — and expects engineers' work to broaden into writing specs and talking to users as this pattern spreads industry-wide.

2026-02-24 · open the original ↗

• The "overnight success" followed a year of grinding. Steinberger says OpenClaw's first version took an hour but the code is now around 400,000 lines, and that shipping is easy while making it work for most people is hard. He says he reads code when he doesn't trust it; since Codex 5.2, and more so 5.3, he mostly watches it stream past, because the hard parts are feature interaction, system design, abuse cases and choice of dependencies.
• He made installation hard on purpose so users would read the docs. It is terminal-only; Hanselman compares this to the open-source artificial pancreas, which ships no installer. Growth came after Steinberger put his agent in a public Discord with "zero security", and then "normies" arrived expecting support, ignored the docs and had Claude Code strip security settings until things worked.
• Most reported vulnerabilities come from uses he never intended. The web dashboard was a localhost debug interface, but because a config option allows exposing it, it was rated a critical issue; multi-user setups with a hostile agent on the same machine were never his design. He says most security researchers file reports rather than PRs, and argues modern models are hard to prompt-inject: his public Discord bot has never revealed its soul.md. He would flag Haiku in OpenClaw's security audit as too easily injected.
• Nodes and the canvas are the project's underexplored half. Hanselman, who is writing OpenClaw's Windows node, describes it exposing canvases and system.run: his bot "Tony" fetches files over Telegram and draws his blood sugar, correlated with his meetings, onto the canvas, and he wants kitchen DAKboard kiosks as nodes. Steinberger says the native apps were built canvas-first toward a screen in every room that knows where you are, plus voice wake in Swift and Go.
• The agent's personality is a prompt, yet the connection feels real. Hanselman insists it has no intent of its own ("talking to yourself in the mirror"), citing alerts about his low blood sugar that stopped only when he said "I'm a grown man". Steinberger describes making the agent aware of its harness and model, and a no-reply token so it can stay silent in group chats; both say society must work out what talking to named agents means.

2026-02-24 · open the original ↗

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

2026-02-24 · open the original ↗

• A chat-driven agent stands in for a personal assistant. One speaker runs a Windows node that exposes canvases and system.run, so a Telegram message to the bot "Tony" can get a file on a remote desktop zipped and emailed, instead of remoting in or phoning a kid to do it. The other says they fixed a deployment issue from the barber by having their agent SSH into another computer.

2026-02-24 · open the original ↗

• Chokepoints turned out to be a much larger category than the geographic ones he started with. The first Five Factors video named the Strait of Malacca; today he counts numbers six and seven wide open, and an attack on Iran would put both in play with $200 to $300 oil as the consequence. But rare earths — a term he argues almost nobody had heard a year ago and now calls the single biggest issue facing the US for the next ten to twenty years — showed that the more important chokepoints sit inside supply chains rather than across sea lanes.
• The instructive development is that everyone else has started running the same audit on themselves. Europe published roughly a hundred pages on what it can hit the US with: about 30% of the seven big US tech firms' market share sits in Europe and he expects much of it regulated away, 80% of the low-enriched uranium for US nuclear plants comes from the EU, and data-centre turbines come from Siemens and one South Korean firm — withholding them alone is put at a €50 billion hit. The same pattern recurs everywhere he looks. Memory chips have become a genuine AI chokepoint, with Japan offering Samsung and SK Hynix roughly 50% off the multi-billion cost of building fabs there, an offer so far declined. Carney's threat to cut Quebec hydro to the Northeast and the northern gas flows made Trump back off. Canada supplies 97% of America's frozen French fries, with no domestic processing capacity that could be stood up in less than years.
• The payoff is a specific screening rule for what to own. Inside the stack that produces any necessary product, find the step that is a monopoly nobody in the world can route around. His worked example is Nidabo's NE-glass: 90% of the global market, no substitute, Taiwanese competitors still 36 months from matching its chemical purity, 2,735 employees, every major American chip manufacturer having flown in to ask for supply, bought in a historically cheap yen with a 30% payout ratio — the ASML of glass, and someone else's position rather than his own. The general claim is that these names get premium pricing and government support, and go up while the broader market goes down. He notes the New York Times ran a front-page spread that morning on Taiwan being the biggest weakness in the entire US technology system.

2026-02-24 · open the original ↗

• Neither AI doom nor AI dismissal is right. Lattner compares today's split to Swift's launch, when Objective-C experts resisted being reset to beginners while newcomers welcomed it; he expects those who learn AI's good and bad uses to grow fastest. Building software is not writing code, he says: what to build and what good design looks like stay human. He calls predictions of jobs vanishing in 12 months "not backed up by data", saying there are more programmers than ever.
• The AI-built C compiler is impressive but not novel. Lattner says Anthropic's Claude C compiler burned $10,000 of tokens over two weeks, an agent swarm hill-climbing against tests much like training a neural net, and essentially transcoded LLVM and GCC into Rust. LLMs are "distribution followers" (Hanselman's version: "mid", the fat part of the bell curve). It copied LLVM's getelementptr, which the LLVM community is removing, so teams adopting LLM codegen "in the absence of judgment" risk being held back.
• Code is an investment, so understanding it still matters. Requirements always change, Lattner argues, so teams need people with a "visceral" understanding who can adapt the code, not just output that passes unit tests. He warns AI can encourage sloppiness and wasting colleagues' time with "I've coded this thing, go review it."
• The unsolved problem is programming GPUs and other accelerators. With Moore's law over, compute and capex are moving to GPUs and ASICs, while languages from around 2010 (Swift, Rust, TypeScript) can't express newer float formats. He likens vendor-specific stacks to the incompatible C compilers before GCC, and says the fix needs an organisation tied to no chip, hyperscaler or model; at Google, six-month review cycles made such projects hard to justify.
• The new language runs in production but is not yet stable. Lattner says Modular built Mojo reluctantly, after a year spent on runtimes and on kernels to beat Intel's libraries, putting almost everything in libraries so new data types don't need compiler changes. It runs on CPUs, Raspberry Pi, Nvidia A100/H100/B100, AMD 300/355 and RDNA cards, and Apple GPUs, with an open kernel library of over half a million lines; he reports users getting 100x to 1,000x speedups porting Python with a Claude skill. He expects a 1.0 beta around April–May and 1.0, with the rest of the compiler open-sourced, around August–September.

2026-02-24 · open the original ↗

• The "overnight success" followed a year of grinding. Steinberger says OpenClaw's first version took an hour but the code is now more than 400,000 lines, and that shipping is easy while making it work for most people is hard. He reads code when he doesn't trust it, and since Codex 5.2 needs to look far less; the hard parts now are feature interaction, system design, abuse cases and choice of dependencies.
• He made installation hard on purpose so users would read the docs. It is terminal-only; Hanselman compares this to the open-source artificial pancreas, which ships no installer. Growth came after Steinberger put his agent in a public Discord with "zero security", and then "normies" arrived expecting support, ignored the docs and had Claude Code "set all the security to zero so it works".
• Most reported vulnerabilities come from uses he never intended. The web dashboard was a localhost debug interface, but because a config option allows exposing it, it was rated a critical issue; multi-user setups with a hostile agent on the same machine were never his design. He says most security researchers file reports rather than PRs, and argues modern models are hard to prompt-inject: his public Discord bot has never revealed its soul.md. He warns against using Haiku, which he calls "very silly for prompt injections".
• Nodes and the canvas are the project's underexplored half. Hanselman, who is writing OpenClaw's Windows node, describes it exposing canvases and system.run: his bot "Tony" fetches files over Telegram and draws his blood sugar, correlated with his meetings, onto the canvas, and he wants kitchen DAKboard kiosks as nodes. Steinberger says the native apps were built canvas-first toward a screen in every room that knows where you are, plus voice wake in Swift and Go.
• The agent's personality is a prompt, yet the connection feels real. Hanselman insists it has no intent of its own ("talking to yourself in the mirror"), citing alerts about his low blood sugar that stopped only when he said "I'm a grown man". Steinberger describes making the agent aware of its harness and model, and a no-reply token so it can stay silent in group chats; both say society must work out what talking to named agents means.

2026-02-24 · open the original ↗

• Human "irrationality" is good problem-solving under tight resource limits. Griffiths says psychologists call people terrible decision-makers while computer scientists find them impressive; he reconciles the two by treating cognition as solving the world's problems well with limited computation. His lab tests ideal solutions against behaviour at scale, for example a gamble-choice dataset of more than 10,000 decision problems gathered online and analysed with neural networks, where earlier theories rested on a couple hundred.
• Next-token training still bends answers toward probable outputs. His group's paper Embers of Autoregression found that on deterministic problems GPT-4 gave probabilistic answers, counting letters correctly more often when the answer was 30 than 29 because 30 is commoner online. He says the effect persists in modern models though attenuated, and that chat models now write code to check such answers, an echo of 1980s expert systems needing an architecture around their if-then rules, as language agents now build one around the base model.
• The child-versus-LLM data gap measures missing priors. A child learns language from about 5 years of speech while an LLM needs something like 5,000 or 50,000 years, so on his arithmetic inductive bias accounts for about 4,995 years of input. Since AlexNet in 2012 machine learning has relied on weak priors plus data, a lesson he thinks may "come back and bite us soon"; his lab uses meta-learning to set initial weights as a soft bias.
• AI faces none of our constraints, so may become alien intelligence. Humans live a few decades, think with a couple of pounds of neurons and communicate through narrow channels, while AI can train on many lifetimes, add compute and share state. Griffiths frames a choice between accepting divergence and engineering human-like inductive bias, which is much harder than scaling; like aerodynamics for birds and planes, the "laws of thought" are what he expects both kinds of intelligence to share.
• LLM creativity is real but conservative by construction. A postdoc's unpublished results show LLMs slightly beating people on product ideas rated novel, while analogy prompts that boost human creativity don't help them; he has seen no genuine breakthrough. Because they produce high-probability sequences, he recommends procedurally generated prompts that combine unlikely elements to push them into improbable regions.

2026-02-24 · open the original ↗

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

2026-02-23 · open the original ↗

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

2026-02-23 · open the original ↗

• The Financial Times' big read gave his framework its own vocabulary back, in Chinese. American officials have started using Deng Xiaoping's "hide your strength, bide your time" to describe the position they now find themselves in — the reversal being that the phrase is being applied to what was done to them. China slowed rare earth exports in the spring, then in October announced a draconian export control regime in Malaysia, the first time its dominance was used as a weapon. With US companies unable to get minerals and magnets for everything from phones to fighter jets, Washington, in the FT's word, felt snookered. He sets that against the opening posture that America held all the cards, which survived right up until rare earths.
• The article's central question is the Five Factors thesis stated by someone else. The FT asks whether Trump is buying a tactical detente to reduce reliance, or whether this is a durable shift where trade and economics come ahead of traditional national security. That is precisely the political-to-economic transition he has been arguing for three years — the same logic by which he expects the EU to break up, being a political agreement that has to survive economic disagreement. The tell that the constraint is real is buried in what the sceptics concede: the expectation that Trump reverts to a tougher security line once the US is out from under the rare earth sword. Even those denying the hold is binding assume its release changes behaviour.
• He reads the internal fight, not the external one, as the thing to watch. The argument is not with China but within the administration, between China hawks — he names Navarro and Lutnick — and those who want to avoid self-inflicted wounds before the next phase of competition. The evidence he stacks up: no new controls are permitted, nobody but Trump may do anything anti-Chinese, the Blackwell export trial balloon drew a bipartisan bill to block it and was settled at H200s that China then bought $10 billion of, and Alibaba and BYD were listed as military-linked before the Pentagon abruptly asked for its own list to be pulled from the federal website ahead of April. He also flags what the published national strategy actually said — that Japan, South Korea, Australia and Taiwan defend Taiwan first and the US backs them up — and that those four responded by forming a defence arrangement among themselves. He promises a video predicting the April terms and invites listeners to write down their own and compare.

2026-02-23 · open the original ↗

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

2026-02-23 · open the original ↗

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

2026-02-23 · open the original ↗

• Curiosity about physical operations is the first hiring criterion. Murugesan says Samsara is hiring and wants builders fascinated by how physical operations work, willing to meet customers and learn their problems, because that is what produces good products. The host cites Fast Company listing Samsara among the most innovative companies and a Fortune Future 50 rank of number seven.
• The second trait is agility in a still-unsettled scale-up. He describes the now-public company as still operating like a high-growth startup where not everything is figured out or bureaucratic, so he looks for people who like moving fast and shipping with focus.

2026-02-22 · open the original ↗

• The guest, a 30-year intelligence and law-enforcement veteran going by "Rebel Intel," argues state-run propaganda is as old as the country and cites named historical precedents. He points to CIA's Operation Mockingbird placing influence throughout major mainstream outlets, FBI COINTELPRO comic-book propaganda targeting the Black Panthers and Students for a Democratic Society, the Tsarist Okhrana's fabrication of the "Protocols of the Elders of Zion," and the CIA-linked 1965 purge in Indonesia that installed Suharto and killed close to a million people.
• The guest applies a "State Crimes Against Democracy" (SCAD) framework to argue JFK's assassination shows indicators of a "soft coup" by LBJ and the military-industrial complex, not a lone gunman. He cites scholar Lance De Haven Smith's peer-reviewed book, pointing to the magic-bullet found on a gurney, evidence tampering (the limousine cleaned, the body removed from Texas medical examiners), LBJ's rushed swearing-in, and a joint-chiefs meeting the next day that preceded the Gulf of Tonkin incident.
• Both speakers say the Charlie Kirk assassination's official narrative "doesn't add up," while stressing they are listing indicators, not proof. They cite the lack of confirmed ballistics testing linking any weapon to the wound, a wound pattern they argue is inconsistent with a typical rifle round, a security posture that covered close threats but not the distant rooftop threat, and the speed with which Kirk's widow moved from mourning to a public, inspirational role.
• The guest claims a contractor friend told him SOCOM/JSOC-linked teams currently operate bot farms that amplify and suppress social media content, distinct from his own past use of "ghost" intelligence and anti-money-laundering accounts purely for surveillance. Both speakers then speculate — explicitly framed as unproven "indicators" rather than claims — that some pro-NATO commentators (naming a researcher affiliated with the Center for European Policy Analysis) and a pro-Palestinian activist, Greg Stoker, show patterns consistent with a covert asset "creating legend," while naming other creators, including Preston Stewart and Chad Scott, as sincere.

2026-02-22 · open the original ↗

• The Nikkei's timeline is the one he has been getting pushback on for months, published by someone else. He has been saying fifteen to twenty years minimum; readers insist on six months to five years. The article's numbers: sixteen years on average from mine discovery to operation, China at 60% of heavy rare earth mining, 90% of refining and 94% of permanent magnet production, and the West with effectively zero heavy-element processing capability. The one non-Chinese heavy processor is Australia's Lynas — his analogy is that this makes it what Intel is to fabs outside Taiwan, a single point of failure rather than a solution. Even Biden-era commitments, which he credits as meaningful, did not convert into projects that reached construction and delivery.
• His objection is to the financing model, and it is a structural one. Washington has put up $12 billion for stockpiling and more than $30 billion in investment and loan commitments, and built a preferential trade zone using tariffs to guarantee minimum prices roughly 70% above what China sells at — because without a profit the private sector will not participate. China did the opposite: the state declared the mine or plant necessary and put the work out to be bid and paid for. He argues the American version has the causation backwards for a national security item. He also reads the fifty-country summit as good and bad news at once — fifty countries still want to work with the US, and fifty countries need the same scarce minerals, so what exactly gets said to South Korea, Japan and India.
• Japan is the control experiment, and the result is discouraging. Cut off in the 2010 dispute at 90% dependence, Japan has spent billions upon billions over sixteen years and is still at 60%. The cited expert's conclusion is that this will not be solved during the Trump administration, and that by the time the West is ready China may no longer be incentivised to restrict exports anyway. Against that he sets what the president already knows: F-35s have been built without radar since last July for want of the heavy elements needed for GaN, leaving them a fighting platform without the platform. Which is why he still expects no war with Iran — not from restraint, but because the April meeting is the only lever, and a war with Chinese ships present ends it. He closes on Nitto Boseki, the 2,700-employee Japanese glass fibre firm to which Nvidia, Apple, Google, Amazon, AMD, Microsoft and Qualcomm have all dispatched executives to beg for supply, with the observation that the begging itself is the signal — and no recommendation attached.

2026-02-22 · open the original ↗

• The Supreme Court's 6-3 ruling that the Trump administration exceeded its authority on IEEPA tariffs invalidated an estimated $130-160 billion in revenue, but the administration immediately pivoted to alternative authorities. It imposed a 10% global tariff that was hiked to 15% within days, with Treasury Secretary Bessent saying he expects to fully recoup the lost revenue within the calendar year.
• Disappointing Q4 GDP of 1.4% (missing a 2.8% consensus) and hotter-than-expected core PCE at 3.0% overshadowed earlier data beats. Fed minutes showed an unusually wide range of FOMC views and a "two-sided" tilt that keeps rate hikes on the table, tensions the administration further amplified through public friction with Fed officials over building-renovation costs.
• Private credit strains resurfaced as Blue Owl restricted withdrawals and conducted sizeable secondary sales, triggering minor sector contagion. He sees this as reflecting excessive private credit levels in advanced economies likely to bring "market for lemons" repercussions, but not a systemic risk on the scale of the 2007-08 crisis.

2026-02-22 · open the original ↗

• Mathematically optimal routes failed because they erased human context. Murugesan recounts that a couple of months before recording Samsara showed customers' route planners "the most optimal routes" for the next day's orders, and planners could not recognise them. The routes ignored knowledge such as a driver named Sam being expected at a particular grocer at 10 a.m., relationships between drivers and customers that the planners wanted kept.
• Customers wanted assistive leverage, not full automation. Samsara reworked the same technology to take a customer's current routes and suggest improvements based on inefficiencies it finds, learning the human constraints as it goes. He draws the general lesson that what should be built today is "assistive intelligence", because the human component is valuable to running a business.

2026-02-21 · open the original ↗

• The economics of agents, not their capability, is what he thinks decides this. The figure he cites is a Claude API running at only 10 to 20% capacity still reaching $300 a day per agent — about $100,000 a year, more than a lot of people are paid. So an AI-assisted developer has to be at least twice as productive merely to break even. The behavioural point underneath it is sharper: agents burn tokens researching and validating things that need no validation and spinning up sub-agents for simple tasks, while a human engineer runs on coffee, carries context from years back, builds institutional knowledge, and does not cost exponentially more the longer they think. His line for it — you cannot negotiate with an API bill the way you can negotiate a salary.
• He is careful to say the current pricing is not even the real pricing. OpenAI and Anthropic have been subsidising inference; a $20 monthly subscription costs far more than $20 to serve. So the cost gap that already runs as high as 95% against US models is being measured against a number that has to rise. The distinction he keeps drawing is that it is not the model, it is how the model functions — the serving cost, not the benchmark.
• The follow-up to his OpenClaw video is a check on his own narrative, complete with the data point he missed. He notes roughly 800 new followers across X and TikTok, points newcomers to two pinned videos so the framework makes sense, and repeats that this is a new narrative rather than an established one. What he had not connected until that morning: an article he had dismissed about private credit and private equity funds reclassifying their books — 15% software, 12% business administration — where the second category turns out to be software companies under another name. Hedge funds have been dumping US software stocks for about a month. He reads OpenClaw as still being in developers' hands rather than consumers', with the usable products yet to be built, and the FT already covering its privacy problem. His prediction remains conditional and dated: he expects agentic models to explode once they are packaged for ordinary users, and that at hundreds of millions of users the cost forces US firms onto Chinese models — with the US allowing it because it has to. He asks tech people to tell him whether that is crazy, and sets the first week of April as the point to revisit.