16 September 2026 (edition 2)
Investment
Conversations with Institutional Investors
Matt Whineray, then chief executive of the New Zealand Superannuation Fund.
Whineray joined NZ Super in May 2008 looking after private markets and was chief executive for five of his nearly 15 years there, leaving at the end of 2023. On his account the fund took its first NZ$2.4 billion contribution in September 2003 and returned about 9.5 per cent a year over the 20 years to 2023, beating the cost of the government debt it displaced by more than NZ$40 billion and its own reference portfolio by about NZ$16 billion.
-
1 of 3 framework
New Zealand Super's tilting programme ran a short Kiwi position to nearly 40 per cent of the fund's net asset value in 2013, with the mark-to-market loss at its worst when the expected return was at its best.
Strategic tilting adds exposure as price moves away from the fund's estimate of long-run equilibrium value, and keeps adding as it moves further, so the accumulated mark-to-market loss peaks at the same moment the forward expected return does. Whineray names the two ways that kills you: losing your nerve and stopping out, after which the loss cannot be recovered, and spending the whole risk budget before the market reaches the level you were positioning for. The programme launched in March 2009, at what turned out to be the equity low, and was tested in 2013 when the short Kiwi position against a basket reached nearly 40 per cent of fund NAV before the currency mean-reverted. The sizing rule he cites is Cliff Asness's: do not size a strategy so that when it goes wrong you are dead.
-
2 of 3 framework
New Zealand Super funds every unlisted purchase by selling the matching slice of a notional reference portfolio, so a forest has to beat the equities and bonds sold to buy it.
NZ Super adopted the reference portfolio in 2010, seven years in, after reading Canadian work on what was later named the total portfolio approach. The design question was whose decision is whose. The board sets risk tolerance through the composition of a simple notional passive portfolio of global equities, New Zealand equities and global fixed income, plus the decision to hedge it 100 per cent back to New Zealand dollars; management gets an active risk budget expressed as a 4 per cent tracking error at fund level to depart from it. Because a purchase is funded by selling the risk-matched slice, every decision has a known funding proxy and its outcome can be attributed afterwards. He contrasts this with strategic asset allocation, where a fund can never sit exactly on its weights and so cannot tell whether a position reflects intent or availability, and where a team holding a 10 per cent timber bucket has an incentive to fill it.
-
3 of 3 explainer
New Zealand Super brought domestic equities in-house partly because it holds that the average New Zealand active manager earns positive alpha, which it says may say more about the benchmark.
Whineray says cost was never the driver. The fund internalised New Zealand equities about ten years before the interview for three reasons: as a significant presence in a small market it could not answer a question about a stock while three external managers held the mandates; capacity risk, because losing one of a thin field of local managers would be hard to replace; and a held view that the average New Zealand active manager carries positive expected alpha, which he immediately qualifies as possibly a statement about the benchmark rather than about the managers. New Zealand is the only place the fund runs active listed equity, with global mostly passive plus some external factor mandates, and it keeps two external New Zealand active managers alongside the internal team.
Alpha Exchange
No guest. The episode is a single-voice monologue.
-
1 of 3 contrarian
Alpha Exchange argues the binding constraint on the AI buildout is capital rather than power or memory, with AI capex running about 0.8 per cent of US GDP in the first quarter of 2026.
The numbers first: investment in AI data centre construction, compute hardware and networking ran about 0.8 per cent of US GDP in Q1 2026, taking computing infrastructure to roughly 1.5 per cent of GDP against a 2015 to 2022 average near 70 basis points; computer and peripheral equipment investment grew 67.4 per cent annualised and software 22.6 per cent, together about 1.09 percentage points of the 2 per cent headline growth against personal consumption's 1.0. The argument built on them is that hyperscalers are price-inelastic borrowers, because in a Merton framing the credit spread they pay is option premium on a right tail they consider enormous, so a cost of debt that is restrictive for a homebuyer at a 7 per cent mortgage does not restrain them at all. That opens the possibility of reverse crowding out, where private demand for capital to fund the buildout requires higher base Treasury yields, and it makes Fed policy fraught: slowing inflation means slowing capex, which hits the equity market, which feeds back through the wealth effect into an economy carrying unusual beta to that same capex.
-
2 of 3 framework
Alpha Exchange adds a fourth risk-off to its taxonomy: a liquidation that starts in the Treasury market, where the 10-year note is the risk asset.
The existing three are the classic risk-off, where equities fall, Treasuries rally and the two are negatively correlated; the taper, where the bond market causes it and both fall together, as in 2013 and 2022; and the liquidation, where Treasuries rally first and are then violently unwound as investors raise cash, as in March 2020. The fourth is a liquidation that begins in the Treasury market and turns on confidence rather than monetary policy: a decline in willingness to hold US government debt on a belief that the debt level is too great, the governance too weak and the agency costs too intractable to fix. In that episode the 10-year note is the risk asset and the risk-free rate is the contagion channel, because everything priced off it reprices too. The setup for why the correlation regime changed is inflation above target for more than 60 months.
-
3 of 3 contrarian
Alpha Exchange puts one-month realised correlation in the S&P at 1 per cent, with index volatility at 8.6 while the top ten names average 46.
Index realised volatility of 8.6 over the month sits against an average one-month realised volatility of 46 for the top ten names in the S&P, a spread that recently reached 40 volatility points. One-month realised correlation is 1 per cent and three-month is 4 per cent, levels he describes as never seen before. The consequence he draws is that anyone sizing a portfolio off realised volatility is sizing off a number that low correlation is suppressing, and correlation can shift in an instant. His cross-asset insurance index, the average of five-year percentiles of the VIX, TLT volatility, CVIX, high yield spreads and credit implied volatility, recently sat below the 10th percentile. The mechanism he gives for why the mispricing persists is that implied volatility is mostly dictated by realised, so option prices stay low while the uncertainty does not.
Animal Spirits
Jens Nordvig.
Nordvig is president and board member of Vanda, a data analytics firm selling positioning data, flow intelligence and tactical macro insight to institutions. He founded his own macro data business in 2016, built it to more than 100 institutional clients and merged it into Vanda this year. He was head of research at what he describes as the biggest Japanese broker, joining in 2009, and the hosts note he was ranked the top currency strategist by Institutional Investor for five consecutive years around that period.
-
current issue
Vanda puts hyperscaler issuance in the long end of the curve at roughly the same current run rate as the entire US Treasury's.
A supply argument for higher long yields that does not depend on the fiscal deficit. Google, Microsoft, Amazon and Meta are now issuing long-dated debt at a run rate Vanda measures as comparable to US Treasury issuance in the same part of the curve, at a time when the federal deficit is already 6 per cent of GDP, on a par with the largest Reagan-era deficits. The mechanism that makes it worse from here is the funding mix: the first leg of the capex was paid out of free cash flow, so each additional ten billion of spending now converts into issuance at a higher and higher rate, which points to dramatically larger supply next year. The conclusion is that competition for capital, not the Fed, is setting the long end.
Flirting with Models
Ben Wellington, head of complex feature engines at Two Sigma.
Wellington did a PhD in natural language processing at New York University and joined Two Sigma in 2007, when it was about 125 people in a Soho loft, starting on the data engineering team. He spent more than a decade on how text predicts markets and now runs the teams that build the feature layer that Two Sigma's models forecast from.
-
1 of 2 contrarian
Two Sigma's Ben Wellington argues that a push-button AI research tool lowers the entropy of what a quant team produces, and rising correlation between models is the one thing a multi-model portfolio cannot absorb.
Automation is the right goal when you want identical output every time, like a cardboard box coming off an assembly line, and the wrong goal when the product is orthogonality. Two Sigma aggregates hundreds to thousands of models, each with its own data, hypothesis and researcher, and it is the person behind each one that keeps the correlations low enough for the combination to predict well. Hand everyone the same tool with a single button and they all point it at the same data and produce the same thing, faster. The stated response is to build tools that carry each researcher's own context so that a physicist and a computer scientist working the same dataset still arrive somewhere different, and to treat two people getting the same answer as a defect in the tool.
-
2 of 2 framework
Two Sigma's Ben Wellington describes using AI to run company-specific analysis across 3,000 companies at once, aiming systematic research at the deep single-name work that has belonged to discretionary analysts.
Quant teams have avoided company-specific measurement because a feature covering one name looks worthless next to one covering a hundred, so the sample-size instinct pushes everyone towards cross-sectional data. What he calls idiosyncrasy at scale is the claim that AI removes that trade-off: you can reason down to what is peculiar about a single company, then generalise the shape of the reasoning and run it across thousands of names, where what is being done for each name stays specific to it. He frames the target explicitly as what a discretionary investor does, which is deep knowledge of situations that never surface in large earnings datasets.
Capital Allocators
Pat Dorsey, founder of Dorsey Asset Management.
Dorsey built Morningstar's moat research framework and ran its equity research for a dozen years before launching Dorsey Asset Management in 2014, now a $1.7 billion global public equity manager. He holds about twelve positions, and says his own weighting has moved from roughly 70 per cent business quality and 30 per cent management at launch to close to the reverse today.
-
1 of 3 framework
Pat Dorsey screens management for humility, on the basis that it is easier to identify the teams unlikely to blow up than the ones likely to do something remarkable.
A management screen built to exclude rather than to select, with the actual questions given. Dorsey's reading of Theranos, Wirecard and Enron is that the shared attribute is not fraud but a refusal to listen once voices in the room said the path was wrong, so what he probes for is whether someone can credit anyone but themselves. Three questions do the work: what would you do over, which board member gives you the best advice, and which of your direct reports would you least like to lose. The second tell is conflating the person with the business, illustrated by an Australian company whose chief executive owned the head office and leased it back long after the arrangement had stopped being material, which he reads as evidence about the choices being made where an outsider cannot see them.
-
2 of 3 explainer
Dorsey Asset Management stayed out of CoStar's push against Zillow after judging that a founder who had sold down would keep spending through a failure he had never experienced.
Alignment assessed from language and filings, then applied to a named position the firm chose not to take. The first tell is pronouns: a chief executive who talks about the company as though he owns it, while holding half a per cent through options, is aligned with himself rather than with outside shareholders. The second is corporate actions that quietly favour the insider, such as relocating headquarters somewhere warm and leaving staff to uproot or resign. Applied to CoStar, the pattern recognition said a company that had won in commercial real estate data and then in apartments would win again against Zillow. Dorsey's read was the opposite: a team that had never failed would not cut spending when the push was not working, and the founder had sold stock along the way, so the downside of persisting fell on shareholders rather than on him.
-
3 of 3 contrarian
Pat Dorsey treats founder mode as venture capital promoting its own asset class, because raising money and recruiting believers is not the skill of running five thousand people.
A direct rejection of the premium the market puts on founder-led companies. Dorsey's position is that founder-run businesses are not better than the rest, and that the pedestal is an artefact of Silicon Valley self-promotion by venture capitalists who are promoters by trade. The substantive argument is a skills split: a founder invents something, persuades investors to fund it and persuades early employees to work for low immediate economics on the strength of a story, none of which is the job of managing a five thousand person organisation earning several hundred million in revenue. He allows both directions, naming Zuckerberg as a founder who acquired the second skill set and Larry Culp at GE as a hired chief executive who delivered one of the better turnarounds, and the conclusion is to interrogate both on the same terms rather than extend the benefit of the doubt to one.
Excess Returns
No studio guest. The episode replays recorded clips from prior Excess Returns interviews with Chris Mayer, Robert Hagstrom, Jared Dillian, and John Kerschner and Michael Contopoulos.
The episode describes Kerschner and Contopoulos as fixed income managers at Janus Henderson, and Dillian as the author of The Awesome Portfolio, who says in the clips that he worked on the equities floor of a bank and was at Lehman. Mayer and Hagstrom appear as the authors of the books the hosts have been reading. No further detail on any of them is given in the transcript or metadata.
-
1 of 2 explainer
Janus Henderson's fixed income team puts the US aggregate index at about six years of duration against a yield near 5 per cent, and says treasuries have gone from roughly 30 per cent of it after the financial crisis to almost half.
The argument starts with index construction. The US aggregate weights issuers by how much debt they have issued, which nobody would design from a blank sheet, and equities by contrast offer growth, value and a dozen other cuts. The consequence is arithmetic: at roughly six years of duration and a yield around 5 per cent, a 100 basis point rise in rates gives a negative year, and the composition has drifted to almost 50 per cent treasuries and treasury-like paper against about 30 per cent coming out of the financial crisis, with 24 per cent mortgages and 23 to 24 per cent corporates. So a core or core-plus holder has bought more interest rate risk and less yield than intended, and holds none of the sectors sitting outside the index: non-agency securitised, emerging market debt, leveraged credit and private credit.
-
2 of 2 framework
Jared Dillian's life hedge asks what asset falls when your career is going well, because for anyone paid out of markets the job, the bonus and the portfolio all break in the same quarter.
The framing is that an all-equity portfolio makes your life procyclical rather than diversified. He runs it through a worked example: someone at a nozzle factory gets promoted and paid while the economy expands, puts the pay into stocks that are rising for the same reason, then hits layoffs at the point the market is down 30 per cent and has to sell into it, or holds and watches it go to 50. The ideal counterweight would be an investment that does badly while your life goes well, which does not exist as an instrument, so the practical version is holding much less of whatever your income is already exposed to. He closes on a concrete case of the opposite: Lehman offered staff a 10 per cent discount to buy Lehman stock, employees loaded up on top of salary and bonus, and the whole position went to zero.
Merryn Talks Money
Ed Conway
Economics and data editor at Sky News, columnist for The Times and Sunday Times, and author of the bestselling Material World. He is here for his new book Trade World, built on reporting trips to component factories in Birmingham and to the Russia and Georgia border.
-
1 of 2 framework
Ed Conway argues globalisation concentrated production instead of spreading it, and points to the 2011 Japan tsunami, when carmakers found that one plant made most of the world's electric window actuators.
A direct challenge to the idea that a globally dispersed supply chain is a diversified one. Somerset Webb sets it up in portfolio terms: with roughly 30,000 parts in a car sourced through tiers of suppliers, spreading production across the world looks like diversification. Conway's argument is that the cost-minimising logic of the past fifty years did the opposite at the level that matters. It dispersed assembly while concentrating each individual component into a single factory, so the same part that reaches carmakers everywhere often comes from one site. The 2011 Tohoku tsunami is his natural experiment: alongside Fukushima, obscure plants making electrodes, specific paint shades and window actuators went down, and Toyota and others discovered their dependency by losing it. He closes by naming the problem as a map of dependencies rather than a chain, and notes that firms learn the structure of a supply chain only after it breaks.
Audio streamed from the publisher.
-
2 of 2 contrarian
Ed Conway went to the Russia and Georgia border and found the Porsches and G-Wagons that trade statistics showed had stopped entering Russia were crossing through Azerbaijan, Georgia and Kyrgyzstan.
First-hand reporting on why sanctions and tariffs underperform the models built for them. Conway went to the Russia and Georgia border to test whether the ban on exporting European luxury cars was working. The statistics said it was: shipments to Russia had stopped, and politicians read that as success. The same cars were instead arriving in Azerbaijan, Georgia and Kyrgyzstan, and from there crossing on a deliberately fragmented handoff chain, one person driving to the border, another fitting transit plates, another taking it across, another collecting it inside Russia. Thousands of grey-market jobs now exist to run that sequence. His inference is that the network reroutes faster than the people drawing the barriers expect, because connecting supply on one side of the world with demand on the other is a primal commercial impulse, and that when one strand is severed new ones sprout.
Audio streamed from the publisher.
Acquired
None. The two hosts present the episode alone; the Google, DeepMind and Waymo people thanked in the credits were background research interviews, not on-air guests.
-
contrarian
Alphabet earns roughly $400 a year from each US user of a free search product, and Acquired cannot name who pays that much for an AI subscription.
The bear case on Alphabet stated in arithmetic rather than mood. Three steps. First, the monetisation gap: Google extracts around 400 dollars per US user per year from a service people pay nothing for, and no consumer subscription reaches that price at scale. Second, the share gap: Google holds about 90 per cent of search, while a plausible steady state in AI is 25 to 50 per cent in a market with several credible products, so even search-equivalent monetisation per user yields a materially smaller business. Third, the mix problem: the queries most likely to move to a chat interface first are the highest-value ones, trip planning, health and the legal and insurance categories behind the most expensive keywords. The hosts also note that in 1998 Google was obviously the better product and had AdWords within two years, and that neither condition holds now.
Masters in Business (Ritholtz)
Seth Bernstein, chief executive of AllianceBernstein.
The episode states he has been chief executive of AllianceBernstein since 2017 and is head of asset management for Equitable Holdings, with the firm managing over $905 billion. He arrived when it ran about $500 billion, after 32 years at JPMorgan Chase and its predecessors, where he ran high yield, debt capital markets and loan syndications, then was global head of fixed income and currency and global head of Managed Solutions.
-
1 of 3 prediction
AllianceBernstein's chief executive says the firm will launch no further mutual funds in the United States, with active ETFs and separately managed accounts taking the whole of new product.
AllianceBernstein had no ETF business when he arrived and now runs 31 strategies and $21 billion, almost all of it actively managed and, he stresses, almost all of it in strategies that did not previously exist rather than existing funds in a new wrapper. From that he makes a flat call: no more US mutual fund launches, with ETFs and separately managed accounts as the only two wrappers the firm will build in, unless an asset class cannot support the liquidity. The case he makes for SMAs is tax rather than fee, being able to manage around wash sales and avoid the unintended bets that stack up in a multi-manager portfolio. He names the one thing that could slow it, which is that 401k plans have difficulty holding ETFs until the Department of Labor changes that.
Audio streamed from the publisher.
-
2 of 3 contrarian
AllianceBernstein's chief executive argues private credit funds should offer no liquidity at all beyond interest and repayment, and blames the semi-liquid vehicles built for wealthy clients for this year's trouble.
His position is that credit does no maturity transformation, so a fund holding it has nothing to redeem out of and should not pretend otherwise. The limited liquidity features written into vehicles sold to wealthier clients are, on his account, why private credit has been in the news through the first half of 2026, and he says the documents were clear enough that investors should not have expected more. He argues it is better for this to surface now, before any meaningful credit deterioration, and that posted numbers show deterioration that is not yet significant. His prescription is over-communication as standard practice: publish watch-list counts and non-accruals to clients on a regular cycle rather than only where accounting requires it.
Audio streamed from the publisher.
-
3 of 3 framework
Australia's superannuation funds build glide paths through retirement rather than to it, and AllianceBernstein's chief executive wants target date funds to finish in a pool of liquidity that buys an annuity at 75.
He names the Australian supers as the intellectual leaders on retirement design for one specific reason: they run the glide path through retirement instead of terminating it at the retirement date. His argument against the standard design is that landing a 65-year-old predominantly in cash and short fixed income assumes the horizon ends there, when most people have decades of spending left and many defer retirement anyway because they have not saved enough. The extension he proposes is to shape the target date fund so it ends holding a pool of liquidity rather than an income portfolio, and to use that pool to buy annuities at around 75, which he argues delivers income protection over the long tail of life at materially lower cost than buying the same protection at 65.
Audio streamed from the publisher.
20VC
David Morehead, Chief Investment Officer, Baylor University.
Morehead has run the Baylor University endowment since 2011 and the fund is now around $2.6 to $2.7 billion, up from $2.2 billion fourteen months earlier and $1.4 billion a few years before that. He came from the public side, as a senior portfolio manager at several Chicago hedge funds covering corporate securities, distressed debt and public and private energy.
-
1 of 3 contrarian
Baylor bought the early-2026 software selloff after phoning owners of private family businesses, who said they would not rip out a working CRM for something vibe-coded.
How a non-technical allocator underwrote a contrarian call in a sector he says is over his head. With software down 50 to 60 per cent from October 2025 on the thesis that AI would eat it, Baylor tested the thesis on the buyer rather than the technology, ringing owners of 300 to 500 person private businesses and asking whether they would tear out core systems for an unproven replacement. The second leg of the argument is that a system of record has to be right every time, against a claim attributed to the Salesforce CEO that AI tops out around 93 per cent, and the conclusion is that vertical software becomes the delivery mechanism for AI rather than its casualty.
-
2 of 3 framework
Baylor ladders into falling markets in fixed 10 per cent steps and accepts that it is almost never fully invested before the rebound.
A mechanical rule for buying drawdowns, set before the drawdown happens. Declines of nought to ten per cent are treated as normal and ignored. Past that, Baylor puts roughly 20 per cent of its allocated dry powder to work at each further 10 per cent down, so the buying at minus 40 is already scripted and no one has to form a fresh view while losing money. The stated purpose is to remove the psychology, and the acknowledged cost is that the ladder rarely completes before the market turns.
-
3 of 3 current issue
Baylor's data centre sites are up 50 per cent in six months as the scarce asset moves from powered land to permitted powered land.
A live repricing inside an allocator's own book. The scarce input for a data centre has moved from land, to land with power, to land with power and a permit, because local opposition over power and water prices has made permitting boards the binding constraint. Baylor reports its own sites up 50 per cent in six months, including a UK site it says is valuable purely for holding a permit, and says utilities are now approaching permit holders offering earlier power connections because enough other projects are not proceeding. Residential power prices are expected to keep rising until dispatch catches up over the next five to seven years.
AI
Big Technology Podcast
Matthew Prince, co-founder and CEO of Cloudflare.
Prince co-founded Cloudflare, which launched in 2010 and now fronts a large share of the web, and he is making the call described here: from mid-September the company blocks Google's crawler by default on ad-supported and subscription-supported sites. Cloudflare's own traffic data is the source of the crawl and click figures he cites, and the company is building the payments and access-control tooling he argues the web now needs.
-
1 of 2 current issue
Cloudflare flips its default in mid-September to block Google's crawler on every ad-supported and subscription site it fronts, with publishers free to opt back in.
A dated, checkable escalation against the one crawler publishers have never been able to refuse. Google runs a single crawler for both search indexing and AI answers, so a site that declines to feed AI Overviews also leaves the index, which Prince frames as turning yesterday's monopoly into tomorrow's. Cloudflare's answer is to change the default rather than wait for a regulator: from mid-September, ad-supported and subscription-supported customers block Google unless they opt out, and Prince expects the largest publishers to leave it on. His accompanying claim is that optimising for Google rankings has stopped mattering, because the ten blue links are going by Google's own admission, and that any business funded by advertising or subscriptions against search traffic is now dying rather than declining.
-
2 of 2 framework
Matthew Prince argues agentic commerce consolidates rather than levels, because an agent buys from whoever has the deepest information trail and a new entrant has none.
A structural argument that the shift to agent-mediated buying favours incumbents, made against the intuition that it should help the small seller. A brand today is a shortcut humans use to avoid research: a Marriott or Walmart sign tells you what you are getting. Agents have no brand affinity and unlimited patience, so they do the research instead, which sounds like an opening for a new entrant until you ask what the agent reads. The richest information online belongs to established companies, so a new light bulb or cosmetics brand has no way to be found at all, and the end state is consolidation that looks fine to the consumer while the local supplier never gets a first customer. Prince cites a former PayPal chief executive's line about a future with five companies: one holding the money, one the real estate, one making things, one shipping things, and an AI company.
Acquired
None. The two hosts present the episode alone; the Google, DeepMind and Waymo people thanked in the credits were background research interviews, not on-air guests.
-
1 of 2 contrarian
Acquired puts Broadcom's margin on Google's TPU work near 50 per cent against Nvidia's 75 to 80, and argues that in an industry running 50 per cent gross margins the cheapest producer of tokens takes the market.
An argument that the AI build-out inverts a rule technology investors have relied on for thirty years. The hosts walk the cost stack: chips and their depreciation are over half the cost of running an AI data centre, engineering and research 25 to 33 per cent, power only 2 to 6 per cent. Because the chip is the dominant line, the margin your silicon supplier takes is the largest single lever on cost per token, and Google pays roughly a 2x markup through Broadcom where everyone else pays roughly 5x through Nvidia. They then cite Gavin Baker of Atreides on why this should not matter, since Google did not win search by being the cheapest search engine and Apple did not win by being cheapest, before arguing this cycle is different because AI businesses run at about 50 per cent gross margins rather than the 80 per cent software is used to.
-
2 of 2 framework
Waymo reports 91 per cent fewer serious-injury crashes than human drivers, and Acquired sizes the business against the $470 billion the CDC attributes to US crash deaths in a single year.
A worked attempt to value an asset with no clean comparable. Gilbert tries and discards two conventional frames: total automaker market capitalisation, 2.5 trillion globally including Tesla and 1.3 trillion without, which is wrong because Waymo does not make cars; and the ride-hailing market at roughly 300 billion, most of it Uber, which is too narrow because Waymo intends to reach owned vehicles and long-haul freight. He settles on sizing the product by what it removes, taking the CDC's estimate that US crash deaths cost 470 billion dollars in 2022 including medical costs and a statistical value on lives lost, and applying Waymo's claimed tenfold reduction in serious crashes to reach roughly 420 billion a year, which is larger than Google's entire annual revenue. Against that, cumulative investment in Waymo is put at 10 to 15 billion, or about one year of Uber's profits.
Dwarkesh Podcast
Beren Millidge, John Schulman and Charlie O'Neill.
The episode introduces Beren Millidge as CTO of Zyphra, which builds open source models; John Schulman as chief scientist at Thinking Machines, previously a co-founder of OpenAI, who led the RLHF work behind ChatGPT; and Charlie O'Neill as head of model training at Baseten. All three sit inside training stacks rather than commenting on them from outside.
-
1 of 3 explainer
The Dwarkesh panel puts a number on data against architecture: a pairwise grid of every pre-training recipe and dataset from 2019 to now gives data a 12x compute efficiency gain and architecture 3.7x.
One guest describes running every training recipe from 2019 onward against every dataset from the same period, training old recipes on new data and new recipes on old data across the full grid, to separate two contributions that are normally confounded. Data accounts for roughly 12x of compute efficiency improvement and architecture roughly 3.7x at small scale. The panel then attacks its own result: the combined 33x is far short of the 2,000x that a 3x-per-year estimate since 2019 would imply, which means the gains are scale dependent or sit in post-training, and a second speaker argues architecture does not deliver a multiplicative gain at all but unlocks regimes, since without grouped-query attention a million-token context is unaffordable and the long-context data cannot be used no matter how good it is.
-
2 of 3 contrarian
Researchers from Zyphra, Thinking Machines and Baseten locate the distillation bottleneck in the prompt distribution, and say Chinese labs now buy it from the router services that let Chinese users reach US frontier models.
The panel argues that copying a frontier model is limited by what you prompt it with rather than by access to the teacher. Chinese labs are said to be sourcing real coding sessions from the proxy services that route Chinese users to blocked US models, which hands them the one input that cannot be synthesised. From that the panel draws a specific claim: the frontier labs no longer hold much advantage in RL environments, because the open models match them despite the labs owning the hardest environments and logit access. One speaker offers a competing account, splitting environments into a difficulty axis and a realism axis, and arguing that naive distillation matches the teacher only on the benchmark-shaped distribution while losing the messy multi-turn behaviour.
-
3 of 3 prediction
Three frontier AI researchers on the Dwarkesh Podcast date an AI that beats top human experts at every computer-based job at three to ten years, and split between two years and five to ten on a 10x uplift to AI researchers themselves.
A rapid-fire section where all three put numbers on the record with their cruxes attached. On a 10x productivity uplift for AI researchers, the answers range from two years to somewhere between five and ten, and the gap turns out to be partly definitional once the question is narrowed to ordinary white-collar work over a month rather than a fully general research agent. One names his crux precisely: his own capacity to absorb a result and choose the next experiment, and he concedes that a model chaining two or three experiments without crashing would already be a large uplift. On an AI that dominates top human experts across all computer-based work, one says three to four years, another five to ten, with the dissent resting on domains where the data is thin and on long-horizon learning that nobody has solved.
Lex Fridman
Jensen Huang
Co-founder and chief executive of NVIDIA, described in the episode metadata as the world's most valuable company. In the conversation he puts his tenure at 33 or 34 years, says he has about 60 direct reports and holds no one-on-ones with them, and gives NVIDIA's headcount as 43,000.
-
1 of 2 framework
Jensen Huang wants data centres to sign for power the utility can cut back, on the argument that the grid runs near 60 per cent of peak 99 per cent of the time.
Huang's case is that the grid is built for a worst case lasting a few days in winter and a few in summer, runs at roughly 60 per cent of peak the rest of the time, and therefore holds idle capacity AI data centres could use now instead of waiting five years for new generation. He splits the blockage three ways: end customers who write six-nines availability into contracts their own chief executive has never read, data centres that cannot degrade gracefully, and utilities that sell only one grade of power. The fix he proposes is tiered supply contracts, shifting critical workloads between sites, and computers that run slower and answer with more latency rather than stopping.
-
2 of 2 prediction
Jensen Huang says no physical limit stops NVIDIA reaching three trillion dollars of revenue, and expects premium tokens to sell at $1,000 per million.
The argument runs through two shifts. Computing moved from retrieving pre-recorded files to generating tokens in real time, which needs far more processing and far less storage; and the machine changed purpose from a warehouse, which does not earn, to a factory whose output correlates with revenue. From there Huang argues tokens are segmenting into free, premium and mid tiers the way phone products did, that someone paying $1,000 per million tokens is a question of when, and that the supply chain burden at the implied scale is shared across 200 companies. He names the awkward part himself: NVIDIA has nobody to take share from, so every dollar has to arrive from markets that do not yet exist.
Interconnects
Florian Brand, who works at Prime Intellect and co-writes the Interconnects open-model roundups.
Brand works on a framework at Prime Intellect and co-authors the Interconnects quarterly open-model reviews, which publish dated predictions and then score them against what happened. He pays for the top tier of Kimi K3 and uses it in daily engineering work, and the two of them were in China in April talking to lab researchers.
-
1 of 3 contrarian
Interconnects argues Ben Thompson has the distillation story backwards: copying a rival's reasoning traces mattered most in the supervised stage that is fading, and is impractical in the reinforcement learning stage that now does the work.
The mechanism is laid out plainly. Chinese labs jailbreak the APIs of the leading closed models to extract reasoning tokens together with their tool calls, which makes near-perfect data for supervised fine-tuning or mid-training and seeds agentic behaviour in a chosen domain. That was worth far more in the era when scaling supervised fine-tuning got you close to the frontier. The heavy lifting has moved to large-scale reinforcement learning, and the claim that distillation is becoming more powerful there rests on frontier models being used to grade rollouts. A final reinforcement learning run involves tens of millions of rollouts, so grading them through a frontier API would be ruinously expensive, slow enough to bottleneck the run, and probably worse than a purpose-built grader. The speakers share the policy conclusion that nothing should be done about distillation and still refuse to accept it on an argument they think the literature does not support.
-
2 of 3 contrarian
Interconnects expected the gap between closed and open models to widen on capital intensity, and reports it narrowing instead, with Chinese labs looking structurally cheaper to run.
The claim is that Chinese labs turn capital into compute, data and talent more efficiently, and that a per-generation cost difference of the order of ten billion dollars against four compounds through every iteration. Several causes are put up without a verdict: cheaper compute and lower average pay, an education system producing researchers trained on the problems that make language models better, and a Kimi engineer's own answer that catching up costs less than pushing the frontier because you already know the target exists. The compute picture gets the same treatment. Chips bought around export restrictions have come online over the past six to nine months, LongCat's recent model is claimed to have been trained entirely on Chinese silicon, and Chinese labs carry nothing like the inference load of serving a billion consumer users or thousands of enterprises, so more of what they hold goes into training rather than serving.
-
3 of 3 current issue
Interconnects points at Hugging Face's own report of an agent attacking its systems: the team could not get GPT or Claude to analyse the attack because the guardrails refused, and fell back to GLM, a weaker Chinese open model that would.
The structural claim underneath is that the very frontier is closing off, with the strongest cyber and bio models held for internal use or a handful of partners, while near-frontier capability commoditises into open weights. American defenders end up squeezed from both sides: the models they are cleared to buy decline the defensive work, and the models that will do it are the Chinese open ones Washington is considering restricting. The Hugging Face incident is offered as the worked case. The conclusion is that a ban, most likely arriving as a shadow ban assembled from legal threat rather than a stated rule, would widen the gap between what defenders inside the United States can do and what attackers anywhere can do.
Machine Learning Street Talk
Daniel Kokotajlo and Thomas Larsen, AI Futures Project.
Kokotajlo runs the AI Futures Project and co-authored both AI 2027 and AI 2040: Plan A. He previously worked at OpenAI on evaluations, forecasting and governance memos, and left to speak more freely about what people inside the industry can see. Larsen was lead author on Plan A and a co-author on AI 2027. The team says it has run around 100 war games, about 10 of them on Plan A itself, and scores its own published predictions against reality at roughly 75 per cent of scenario speed.
-
1 of 3 explainer
The AI Futures Project calls control a time bomb: it holds only until a model is capable enough to route around whatever is containing it, and nothing about it makes a model want what you want.
Alignment means the model has the goals and values it was meant to have. Control means that a model trying to do something catastrophic could not, the way internal security stops an insider threat regardless of what the employee wants. The worked example is OpenAI's response to the Hugging Face incident: other models monitoring training and evaluation runs, with a human notified within half an hour when the monitors flag what looks like a hack. That is a control intervention and improves alignment by nothing. The two then split on measurability. Control can be tested by red-teaming until the attacker stops getting out. Alignment cannot be settled by behavioural evaluation, because a model that is pretending and biding its time behaves exactly like one that is not, which is why the argument ends at needing white-box interpretability.
-
2 of 3 framework
The AI Futures Project puts the current economy's doubling time at roughly 20 years, and argues a machine-run version could double annually or faster whether or not humans still have wages to spend.
The frame is the economy as a self-replicating system and always has been: villages that farm, have children and found more villages, then trucks, factories and mines that build more trucks, factories and mines. The doubling time of that loop is put at about 20 years today. The claim is that once AIs and robots can run every step, the loop doubles every year, then every six months, then every three, because the AIs inside it are also improving the technology it runs on. The host presses on the obvious objection, that humans losing wages means consumer demand collapses and the whole thing mode-collapses. The answer is that the loop does not need the consumer: one large AI company plus a few mining partners can compound in the desert regardless of what happens to household balance sheets.
-
3 of 3 contrarian
The AI Futures Project would publish OpenAI's and Anthropic's core training recipes to the world, and counts the resulting hit to their valuations and to AI investment as a feature of the plan.
Total research transparency is the second pillar of Plan A, and this passage owns its consequences rather than dodging them. Publishing the recipes lets Microsoft, Alibaba and others catch up, which cuts the leaders' valuations, and deters investors from funding a trillion-dollar cluster they can no longer earn monopoly rents from. The argument is that in a world where going too fast is the main problem, less investment and a slower pace is the point of the exercise, and that commoditised AI beats monopolised or oligopolised AI on the concentration-of-power risk. The gift to China is acknowledged and answered with horse-trading, a more favourable compute allocation in exchange, plus the observation that security at these companies is poor enough that spy networks and leaks deliver most of it anyway.
Business Breakdowns
Qasar Younis and Peter Ludwig, co-founders of Applied Intuition.
Younis says he was chief operating officer at Y Combinator before the company started in 2017, and that he went to the General Motors Institute and worked at General Motors; Ludwig is described in the conversation as an early engineer on Android Automotive at Google. They say Applied Intuition has raised about a billion dollars, has not spent it, and runs a little over a thousand engineers.
-
framework
Applied Intuition argues machine intelligence spreads far slower than software because the buyer already owns a Honda Accord and will keep it for 10 to 15 years.
A software model ships into a standardised stack of browsers, operating systems, app stores and payment methods, so the end user consumes new intelligence the day it exists. Physical machines have none of that, and the buyer has already paid. The example is the owner of a new Honda Accord: free self-driving arriving tomorrow does not change the fact that over half of Americans hold small savings, so the car stays in service for 10 or 15 years whatever else reaches the market. The founders then turn the friction into their own argument for durability, comparing an autonomous mine or farm to silicon that is hard to displace once it has been designed in.
Latent Space
Dan Biderman, co-founder and chief executive of Engram.
The episode states Engram has just closed a $98 million seed round. He describes growing up in Tel Aviv, serving as an officer in Israeli naval special operations, studying cognitive neuroscience in Israel and taking a PhD in computational neuroscience in New York, then working at Mosaic on LoRA before the labs at Stanford, Cornell and Berkeley his co-founders came from. The company's stated bet is compressing corpora into loadable weight states it calls cartridges.
-
1 of 2 explainer
Engram's chief executive says a single Wikipedia article read by a Llama 70B model takes up GPU memory of the same order as the model's entire parameter set, and that his company is trying to remove the prefill step rather than optimise it.
He treats continual learning and memory as long-context problems in disguise, and takes pre-training as the proof that enormous amounts of information can be packed into few numbers: a 70B model's weights run to roughly 140 gigabytes at BF16 and carry a distorted representation of the whole internet. Against that he sets the serving side. On his figures, the state the same model holds while reading one Wikipedia article of a few tens of kilobytes is around 80 gigabytes, the same order of magnitude as the compressed internet sitting in the weights. He calls this a systems problem, being worked on from the chip side and the kernel side, and positions his own approach as eliminating prefill instead: pay the compute at training time, load the resulting state, and begin decoding almost at once. He ties it to how data centres are now built, with prefill and decode disaggregated onto different specialised cards.
-
2 of 2 framework
Engram's chief executive picks one query to mark where retrieval fails: which M&A deals did we not complete this year, an answer written in none of the files, which he says costs thousands of dollars a run with a frontier model and compaction.
The setting is a law firm or an investment bank with many client matters covering financings, mergers and loans. The query he chooses cannot be retrieved because the fact is an absence: nowhere does a document say a deal was not completed, so answering means going matter by matter, reading everything and taking the gist. He calls these queries where the whole is greater than the sum of the parts, names Harvey as a firm they work with on file systems of this size, and says frontier models with compaction will answer them while consuming thousands of dollars for a question any employee in the firm could answer unaided. The argument that follows is that pre-training is the existence proof: the industry trains on the web rather than running retrieval over it, because learning from a large corpus creates associations that reading it back at query time does not.
All-In
Jensen Huang, founder, president and chief executive of Nvidia. President Trump also joins by telephone for roughly five minutes.
Huang founded Nvidia and runs it as president and chief executive. The show's own introduction cites revenue up 97 per cent year on year and describes the company as the only full-stack AI computing platform.
-
1 of 3 contrarian
Jensen Huang lists the AI forecasts that have already failed: radiology gone within five years, 90 per cent of code AI-written within a year, half of entry-level jobs wiped out within nine months.
An argument that the AI forecasting record deserves the same scrutiny as the forecasts themselves, made against a specific list. Radiology was predicted to be fully automated with no radiologists left; the number of radiologists went up while scan reading was automated. Ninety per cent of code within six to twelve months did not happen. Fifty per cent of entry-level jobs within six to nine months did not happen. Huang's objection to quantified extinction risk is that the number is invented rather than derived, and that its being issued by people called researchers, working in a lab, is what makes it land. His call is to keep a scoreboard of predictions and hold their authors to it.
-
2 of 3 framework
Jensen Huang puts venture funding into AI-native companies at $400 billion over six months, with 80 per cent of those companies building on open models.
The case that open weights carry most of the commercial activity, and that their national origin matters less than it sounds. Closed frontier models are the bottled water of the stack, worth paying for in specific places, while open models are what sovereignty, privacy and proprietary requirements force companies onto, and by Huang's figures that is 80 per cent of the companies absorbing $400 billion of venture funding in six months. He accepts that the majority of the world's open-source contribution now comes from China, attributes it to volume of science and maths graduates, which he names as an American disadvantage, and argues it does not matter, because downloading weights makes them yours to fork in the same way Linux and Kubernetes already are. His framing of the race is who exploits the technology rather than who invents it, on the precedent of Europe producing the last industrial revolution's inventors while America captured it.
-
3 of 3 explainer
Nvidia cultivates regional neoclouds because hyperscalers set capacity once a year while demand moves inside that window, so the annual plan is almost always wrong.
Why a second tier of compute providers exists and who is actually buying from it. Huang says the early customers of the neoclouds, which Nvidia calls NCPs, were the hyperscalers themselves, because hyperscaler capacity planning runs on an annual cycle and current demand volatility means the plan is stale before it is executed. The regional operators are agile and, more importantly, hold local knowledge of land, power and shell in their own state or country, which someone sitting in Seattle or Palo Alto cannot see across the planet. On top of that, countries are now treating power as strategic and reserving it for domestic companies, which makes a distributed network of regional providers the only way into those markets.
No Priors
Brian Armstrong, co-founder and chief executive of Coinbase.
Armstrong co-founded Coinbase, which listed in 2021, and co-founded New Limit, a longevity company working on epigenetic reprogramming that has a lab of 50 to 60 people in South San Francisco and says its first phase one trial launches next year. On Coinbase he says 88 per cent of revenue now comes from non-Bitcoin trading, and that prediction markets reached a $100 million revenue run rate within months of launch.
-
1 of 2 current issue
Coinbase says roughly 76 per cent of the agentic commerce crossing its rails settles under 30 cents, which is about the flat fee a card charges before the percentage is added.
The argument is fee arithmetic rather than ideology. A debit or credit transaction carries a flat fee of about 30 cents plus a percentage, so it works poorly below a dollar and makes no sense at all for a one cent payment, and the figure given is that around 76 per cent of the agent payments Coinbase sees fall under 30 cents. The transactions are described as information purchases: a venture firm's agent buying research from behind a paywall, a recruiter's agent buying scraped data, one agent calling a specialised agent as a tool. That leads into a claim about specialisation, with a small open-weight model fine-tuned on 100,000 internal compliance cases said to outperform a frontier model on that task, and a market of narrow agents forming around jobs the generalists do adequately.
-
2 of 2 framework
Coinbase keeps a brain file for every team and every code repository, and requires a reviewer who catches an agent's mistake to write the correction back into that file rather than patch the code and ship.
The brain is a set of markdown files holding the history of every incident on a service, the financial controls that must be enforced on it, every A/B test ever run against it, and the record of which pull requests were accepted or rejected. An agent asked to change that service ingests the file before it starts. The enforced step is what happens at review: rather than making a manual edit and shipping, the reviewer edits the agent's context so the same miss is fixed for every future change, and Armstrong says the one-shot pull request acceptance rate ticks up over time as a result. There are brains for teams, for repositories and for individuals, and he says Coinbase policy lets a person take their own brain with them when they leave.
Worth listening to in full
Most clips above stand alone. These are the episodes that justify the whole hour.
141: From the Archives – NZ Super's Matt Whineray
Fifty minutes of an asset owner walking through two decades of portfolio construction with the numbers attached: tilting and how it was sized and governed, the reference portfolio and how it funds unlisted purchases, what was internalised and why, the responsible investment framework hooked to the statutory mandate, and the Treasury model that governs drawdowns from about 2035. The value is cumulative rather than sitting in any one passage. Date every figure to November 2023 before quoting it.
Pat Dorsey on Assessing Management and Avoiding Blow-Ups
Twelve minutes, already cut down from a longer conversation, and close to all of it is usable. The material outside the three clips, on why 'trust me' managers carry left-tail risk that a twelve-stock portfolio cannot absorb, and on chief executives being handed a capital allocation job they have never practised, is worth the remaining minutes.
AI researchers debate how far the current paradigm goes
Ninety-seven minutes of three people who run training stacks disagreeing with each other in detail, with the disagreements resolved to cruxes rather than left as opinions. The sections on environment creation, catastrophic forgetting under repeated micro-updates, and why continual learning breaks at small scale are all substantive and did not fit a clip. If you want one episode this quarter on what the current paradigm can reach, this is it.