Souk's Notes - Unofficial Summaries
Dr. Jon Krohn954: Recap of 2025 and Wishing You a Wonderful 2026 — with Jon Krohn (@JonKrohnLearns)
- Agentic AI is overhyped but powerful enough to prepare for now. The host says leading proprietary LLMs can replace humans on 2-hour software development or ML engineering tasks with 50% accuracy, a duration doubling every seven months. He urges listeners to get their data, security and governance ready to use increasingly autonomous machines in 2026 and beyond.
- The podcast's shift to video sharply grew its YouTube audience. The host says changes implemented in August raised daily YouTube watch time tenfold within two months, and subscribers rose from 25,000 a year earlier to over 140,000. His personal highlight was recording with Aurélien Géron, who he says had not done a podcast in a decade, before a live audience at the University of Auckland. Episode 1,000 is due in June.
All Things FinancialSilver Lease Rates indicate a sharp rally to close the price between physical and paper prices 1-1-2026
- He starts the year by testing the loudest claim in his feed and finds no evidence for it. The story is that the banks are short silver, going bankrupt, and being bailed out through Fed repos. His check is the simplest available: JP Morgan and Bank of America charts over the past year. If the two largest banks in the universe were short, in trouble, and on Fed life support, those prices would not be where they are. His reasoning is social as much as financial — New York is a small world where people move between firms and into government and back, everyone knows when someone is in trouble, and the standard response is to trade against them for the bonus. That is why positions are guarded, and why a secret of that size does not keep.
- His alternative explanation for the repos is arithmetic about issuance. Treasury sells new debt to the banks, which must pay for it before distributing it into the system. He asked how much was issued in December: roughly $500 to $600 billion a week across the first two weeks and again across the last two — close to $2.5 trillion in a single month. Repos of $28, $29, $71 billion against that funding requirement are unremarkable. He notes the same crowd made the same argument a month earlier about reverse repos disappearing, and it resolved the same way.
- Silver lease rates are the indicator he thinks people should be watching, and they are far outside normal. A lease rate is what you pay to borrow physical metal, quoted annually as a percentage of the metal rather than of cash. Normal in a balanced market is 0.1 to 2%; 2 to 5% means tightening; above 5% is severe stress; 8 to 10% is unusually high even for a spike, and indicates that vaults, refiners and eligible delivery stocks are genuinely strained. Silver is at 9 to 10%. He fairly presents the opposing case — the argument that when liquidity dries up paper markets break first, because leverage sits in the paper and not in the metal, so the paper gets sold — and notes silver is among the most paper-levered markets on the planet, with hundreds of claims per ounce. But both his indicators point the same way: backwardation historically resolves through a sharp upside move as paper catches up to physical, and extreme lease rates have preceded explosive rallies. What he leaves open is the magnitude, given demand from EVs and solar arriving as China withdraws exports.
Scott HanselmanVjekoslav Krajačić on File Pilot and a return to fast UIs
synopsis
Vjekoslav Krajačić, the developer behind File Pilot, ends up holding that native apps could be far faster and leaner if more developers built them the way games are built, though Krajačić doubts the next generation will learn to, with AI and deep stacks pulling programmers further from the hardware.
From quiet project to public beta. Krajačić started File Pilot in 2022 after leaving the game company they had run for over a decade, saw a gap because most file managers dated from the Windows 95 era, worked incognito for two or three years, had a few hundred alpha users, and drew wide attention with the public beta in early 2025.
Built like a game. Hanselman asks why games run at 100 frames a second while apps stutter on resize. Everything is custom: OpenGL rendering (a DirectX move is planned over driver bugs), an immediate-mode UI that redraws each frame, and Krajačić's own standard library. Hanselman's analysis found a 2.3 MB executable with 14 DLL dependencies and 312 imported functions.
Where it hurts. The worst problems are in old Windows APIs: the context menu stays single-threaded because shell extensions break off the main thread, and drag and drop still has unsolved bugs. Hanselman offers Microsoft contacts, and live finds network-share handling wanting.
AI and the next generation. Krajačić uses ChatGPT only as a better search engine and then verifies against docs, calling AI-written code dangerous for people who can't spot hallucinations. Hanselman agrees, saying seniors get boosted and learners slowed.
What's next. As recorded: free beta at 0.6.3, a $40 individual perpetual license on pre-order, exit from beta early to mid next year, then Unicode rendering, drag and drop from zip archives, possible macOS and Linux ports, and one more hire.
- Apps can be as fast as games if built like games. Krajačić, a former game-studio head who started File Pilot in 2022, built the Windows Explorer replacement from scratch in C with a custom OpenGL renderer and an immediate-mode UI that redraws everything every frame. Krajačić credits immediate mode for the speed and plans a move to DirectX early the following year because of OpenGL driver bugs.
- A minimal, self-contained binary is the design goal. Hanselman's own analysis found a 2.3 MB single executable with 14 DLL dependencies and 312 imported functions; Krajačić writes their own standard library rather than linking the C runtime, uses about 10 C files plus a roughly 200-line C++ file for DirectWrite, and supports Windows 7 and up.
- The hardest bugs live in old Windows APIs they cannot see into. The right-click context menu is the one part not multi-threaded, because shell extensions seem written for the main thread and break when moved; drag and drop still has bugs Krajačić can't fix, and few people still know these APIs. Hanselman offered an introduction to people at Microsoft; live, he also found that a UNC path to his server opened a local folder, and credentialed network shares are not yet supported.
- AI boosts those who can already solve problems, not newcomers. Krajačić is "not too optimistic" about the next generation of programmers, and uses ChatGPT as "improved Google" and then checks every parameter in the official docs, having been bitten by hallucinated code. Hanselman says he and Mark believe seniors get boosted while learners get slower and write worse code.
- Beta is free; paid release and new features are planned. As recorded, the beta (version 0.6.3) was free, pre-orders carried a 20% early-bird discount, and an individual perpetual license was $40; Krajačić aimed to leave beta in the beginning or middle of next year. Next: Unicode rendering for Japanese and Korean users, drag and drop from zip archives, possibly macOS and Linux ports, and hiring at least one more person.
Dr. Jon KrohnHow Dell Achieved 30x ROI on AI Projects (with John Roese)
- A disciplined AI program decoupled revenue growth from cost. Roese says that in the fiscal year Dell reported in March, the first with its four big AI areas running, revenue grew $10 billion while absolute cost fell 4%. That had not happened in Dell's 41 years, and he credits people, process and technology. The host notes Roese's 10-to-1 and 30-to-1 ROI ratios, which feed the "AI ROI flywheel" Roese described in episode 887.
- Insist on hard ROI metrics, and AI exposes work not worth doing. Roese's first principle is to target measurable things like cost, margin or revenue rather than goodwill. He cites a consulting report, he thinks McKinsey's, calling generative AI a "BS work detector": if applying AI tools to a task changes nothing in revenue or cost, the work may not be worth doing. He urges picking targets where improvement makes something material happen.
GeoHussarBiden and the Democrats in Congress are why inflation was so bad and still still persists
- He rejects the title's blame on Biden and congressional Democrats. The title is a setup he immediately calls untrue; this is part two of his argument that inflation is the Federal Reserve's fault, dismissing both the camp that blames Trump's first term and the camp that blames Biden and Democrats in Congress.
- First-term money-supply growth guaranteed the later inflation. Repeating his previous post, he says the US money supply saw its largest increase in recorded history under President 45, and that any drastic expansion guarantees inflation later. He places reckless money printing in the Fed's purview rather than a president's.
- Congress's role is triggering monetization through stimulus acts. He describes the mechanism: Congress passes stimulus, the Treasury issues hundreds of billions of dollars in Treasuries, and the Fed creates dollars to buy them and holds them on its balance sheet. The captions cut off as he turns to 2020.
Crosstalk SolutionsIndy Car Pit Practice
- No text was available for this short. The transcript file for this 14-second video is empty, so there is nothing from it to summarize.
Loss DogCritical things to think about before starting your own business! #oneluckydog #careergrowth
- Double whatever money and time you think a new business will need. Because costs run higher and timelines run longer than planned, he says the worst position to be in is partway through a venture, running short on both money and time, with no way to finish.
- Full commitment, not the idea itself, is the one make-or-break question. He argues there's no such thing as starting a business part-time; without total commitment it's "incredibly difficult," though he adds that whatever happens, the attempt itself teaches enough to make it worthwhile.
Scott HanselmanLoris Cro on the Rise of Zig
synopsis
Loris Cro, VP of community at the Zig Software Foundation, ends up presenting Zig as a modern C rather than a modern C++: a pre-1.0 language whose real advantage is working intimately with existing C and C++ code, run by a small nonprofit that keeps big tech off its board.
Why Zig exists. Cro came to Zig from bioinformatics, needing to teach Redis module writing without C's intimidation, and argues C and C++ carry 40-plus years of quirks, such as OpenCV bindings that compiled only because of include order. Zig, at 0.15.2, promises no backwards compatibility, which Cro treats as freedom to fix mistakes.
Against Rust. Hanselman framed Rust as optimising for memory safety and Zig for control and explicitness; Cro agreed, adding that Zig cross-compiles C and C++ dependencies, which Rust leaves to the user. Where Rust's guarantees stop at safe Rust code, Zig adds type-system detail for C interop and defer cleanup, behind Cro's claim that C libraries are easier to use from Zig than from C.
Comptime. Asked how comptime differs from #ifdef, Cro points to lazy evaluation: unresolved branches are not semantically analysed, so per-OS code needs no guarding, and types can be queried at compile time. The cost is that errors in uncalled code go unnoticed.
Governance and ecosystem. The 501(c)(3) refuses big-tech board seats and avoids reliance on one donor. It left GitHub for Codeberg mainly over GitHub Actions problems, preferring nonprofit services as more stable. Bun began with one person asking questions on Discord and, Cro notes, has since been bought by Anthropic; TigerBeetle allocates memory only at startup.
- Zig is unfinished by design and will keep breaking code. Cro, VP of community at the Zig Software Foundation, says the 0.15.2 release carries no backwards-compatibility promise, and calls that "our superpower" because the team can change things without fighting legacy commitments.
- C and C++ are full of rough corners that cut newcomers.
Cro's example: third-party C bindings for OpenCV (used by the GoCV package)
included a C++ header inside an
extern "C"block, and only compiled for Go because an include guard had already fired earlier in the build. - Zig's edge over Rust is integrating with existing C code.
Cro says Zig cross-compiles C and C++ code, which Rust does not, and claims it
is "easier to use C libraries from Zig than from C itself", citing pointer
types that distinguish one item from many, null terminators in the type
system, and
errdefer-style cleanup. Hanselman framed Rust as optimising for memory safety and Zig for control and explicitness, and Cro agreed. - Comptime relies on lazy evaluation, which makes multiplatform code natural. Branches resolved at compile time are not semantically analysed, so a Windows branch can call APIs absent on macOS without guards; the trade-off is that errors in uncalled functions go uncaught. Cross-compilation lets CI build all targets on one machine.
- The foundation avoids big-tech board seats and a single dominant donor. The Zig Software Foundation is a 501(c)(3) that balances corporate and public donations; it moved from GitHub to Codeberg mainly over GitHub Actions trouble, including a "safe sleep" process that ran for two months, and prefers nonprofit services such as every.org as more stable. Cro cites Bun (since bought by Anthropic, Cro says) and TigerBeetle, which allocates all memory at startup, as major Zig projects.
Dr. Jon KrohnWhy MCP is Dangerous for Enterprises (with John Roese)
- MCP is a keeper but not yet enterprise-grade. Roese, Dell's global CTO, says MCP is a tool-and-data-access abstraction, not an interworking protocol like A2A, and that it was built to let public LLMs consume data, so it still lacks security work (OAuth is being added). He calls it "incredibly powerful and incredibly dangerous" because it can do more than users expect, and expects most agents to use it long term.
- Dell wraps MCP in a central controlled environment. Under Dell's internal agentic guidelines, which he says are not yet public, all MCP servers are centralized so the company can impose the controls the protocol itself cannot.
- Disciplined, top-down AI deployment is how Dell avoids slop. Roese says Dell has deployed probably fewer than 30 AI use cases, with ROIs of 10:1 to 30:1, across sales, services, supply chain and engineering, rather than handing a tool to half a million people and hoping. The host brought up The Economist naming "slop" its word of the year at the time of recording; Roese answers that "it's better to do nothing than to do chaos."
Loss DogHow Should I Manage Zero DTE Risk__ _ O.L.D. LIVE! _ 12.30
synopsis
The hosts open a holiday-thin session with the S&P and Nasdaq roughly flat, bonds pinned in a weeks-long range, and gold and silver whipping violently against a backdrop of a silver short squeeze both hosts are personally caught in. Their running message for the day is defensive: with volume down 20-25% into year-end, size down, don't chase the metals move, and manage existing risk by rolling and reducing rather than by adding stops or new capital.
Zero-DTE risk and position sizing. A caller trading zero-DTE SPX iron condors on paper wants to add stop-losses because he's loss-averse. The hosts argue stops actively hurt results on defined-risk trades since a majority of trades that dip into losing territory come back to be winners; the better fix is trading smaller products -- ES (half the SPX size), or SPY and XSP (a tenth the size) -- rather than capping losses. If he still wants a stop, it must be a stop-limit, since a plain stop on a wide iron condor can fill far worse than intended.
Angel investing in friends and family. A caller considering funding a relative's startup is told to commit only an amount he'd be fully comfortable losing, since early-stage odds are long regardless of how good the idea or person seems; the position should be equity only, never a loan, and a strong outcome would create the option to invest again later rather than needing to be maximized up front.
Defending the silver squeeze. Two viewer emails describe being caught short calls or futures in the parabolic silver move, which one host attributes possibly to Chinese spot demand or margin-call chatter, without endorsing either explanation. Both hosts are in the same trade: hold or roll out to February rather than close, don't add to the position, and use a stop well above the market -- around 87 -- only if it would otherwise threaten the account. With IVR near 120, they'd rather add short premium via SLV strangles than trade the futures directly.
Starting a business, and a career pivot into investing. Answering a generic "how do I start a business" email, the hosts skip checklists in favor of one filter: total commitment, since half-measures and part-time efforts rarely survive. That theme carries into a call from a mid-30s tech veteran making $600k who has posted roughly 24% annual returns in value investing over six years and wants to go pro. The hosts, both veterans of running a hedge fund and a public mutual fund, tell him a fund path suits his hands-on temperament better than becoming an adviser, but warn that his whole adult investing life has only seen a bull market, and that raising outside capital -- roughly a $100,000 setup cost -- means becoming "a salesperson and a babysitter" who answers to other people's money.
AI versus the executive suite. Prompted by a viewer email, one host argues AI should replace expensive executives -- CFOs, CMOs, CTOs -- rather than entry-level staff, citing CEO pay as overpaid 40-60% against a $21 million S&P 500 benchmark; his co-host insists leadership and creativity can't be reduced to a language model. A live poll on replacing a CMO with AI closes 59-41 in favor.
Pairs trades and hedging philosophy. One segment walks through the most liquid pairs -- ES/NASDAQ, bonds/notes, gold/silver -- and the mean-reversion logic behind trading the spread rather than the direction; another call pushes back on a NASCAR braking analogy for hedging, with the hosts preferring to reduce position size or improve cost basis over adding hedges that consume capital. A caller's question about their post-tastytrade brokerage draws mention of the coming Lost Dog platform, which will display their portfolio (not individual trades) once launched. The show closes on quiet market action -- Meta higher, Micron still running post-earnings -- and a preview of New Year's prediction calls and 2026 contests tied to the platform launch.
- Loss-averse zero-DTE traders should size down, not add stops. A caller wanted to add stop-losses to zero-DTE SPX iron condors; the hosts say stops increase the odds of a losing trade because more than half of at-risk trades recover, so the fix is trading smaller contracts (ES, SPY, or XSP instead of full SPX) rather than capping losses. If a stop is used anyway, it must be a stop-limit, since a plain stop on a wide spread can fill at a far worse price than intended.
- Defend, don't exit, silver positions caught in the squeeze. With one host short January 66-strike silver calls and implied volatility rank near 120, both say not to add to the position, to roll out to February rather than close, and to set any stop well above the market (one mentions 87) rather than covering into the spike. The high IVR is framed as an argument for selling premium via SLV strangles instead of trading the futures outright.
- A strong value-investing track record doesn't guarantee an easy capital raise. A caller reported roughly 24% annual returns over six years and asked about starting a fund or becoming an adviser. The hosts, who have run a hedge fund and a public mutual fund, estimate it costs roughly $100,000 to properly set one up and warn that managing outside money makes you "a salesperson and a babysitter" answering to partners on every decision.
- AI should replace overpaid executives, not entry-level jobs. One host argues the average Fortune 500 CEO is overpaid 40-60% against a cited $21 million benchmark, and that AI could run finance, marketing and technology functions while staff move up to fill the gap. His co-host pushes back that leadership can't simply be programmed into a language model, and a live viewer poll split 59-41 in favor of replacing a chief marketing officer with AI.
GeoHussarIs Inflation Biden’s fault?? #inflation #economy #finance #learnfromme #analysis
- Money-supply growth, not Biden, drove the inflation. He calls Biden "a common scapegoat" and names the surge in money supply from roughly 2020 to 2022 as the main cause. On a 2015-to-present chart, he shows a gradual rise from 2015 to 2020, then a near-vertical climb starting in 2020.
- The steeper money-supply growth came under Biden's predecessor. Comparing the two terms on the chart, January 2017 to January 2021 against Biden's January 2021 to January 2025, he argues the earlier term's line is the steeper of the two. Blame by presidency, on his reading, points the other way.
Polis PanditMy Best Work From 2025 Across Writing and Filmmaking
- He decided to stop chasing the news cycle. Despite writing about tariffs, Charlie Kirk, and the AI bubble in 2025 and having some of those pieces do well financially, he says none were his best work because they weren't stories only he could tell — a big reason he shuttered The Political Prism.
- He identifies a matching problem in filmmaking: over 400 YouTube videos, mostly vlogs with no narrative arc or story, which he says don't work unless the creator is already famous.
- His stated filter for 2026 is whether his perspective is unique and adds value nobody else could — a standard he admits he doesn't always meet, comparing the temptation to weigh in on trending events to a child unable to resist a marshmallow.
- By that standard, he names his best 2025 work — the personal essay "4 Years Ago, I Entered the Wilderness" and the film "A Snow Day" — as proof that personal, non-newsy stories mattered more to him than his most popular pieces, even when they weren't the same thing.
Dr. Jon Krohn953: Beyond “Agent Washing”: AI Systems That Actually Deliver ROI— with Dell’s Global CTO John Roese
synopsis
Roese, Dell Technologies' global CTO and chief AI officer, returning to give his annual predictions, argues that AI pays off only when it is disciplined and narrowly targeted. Dell's evidence, as he gives it, is a fiscal year in which revenue grew $10 billion while absolute cost fell 4%. His 2026 predictions follow from that: governance, a new knowledge layer, agents as four-part software systems, resilient AI factories, and sovereign AI put to wider use.
Agentic AI and agent washing. Roese says his 2025 prediction was that "agentic" would be the word of the year. He defines it as the digitization of skills: an autonomous system that reasons and executes while the human supplies intent and validation. He calls the term wildly overused and says many "agents" are tools that bring autonomy to people. The host cites the METR finding that the length of human tasks AI can do at 50% accuracy doubles every seven months, now about 2–4 hours. Roese adds a second dimension: agent ensembles take on composite work that spans people and organizations.
MCP and slop. MCP is a tool and data-access standard, not an interworking protocol. Roese says it is early, lacks security, and is dangerous if used carelessly, so Dell centralizes its MCP servers under its unpublished internal agentic guidelines. The host raises The Economist's word of the year, "slop". Roese answers that Dell runs fewer than about 30 use cases with ROIs of 10:1 to 30:1, and that it is "better to do nothing than to do chaos."
ROI. Roese calls the revenue-cost decoupling a first in Dell's 41 years, a mix of people, process and technology. He repeats his rule that every project needs a hard ROI metric. He cites a report he attributes to McKinsey calling generative AI a "BS work detector."
Governance and data. More than a thousand uncoordinated jurisdictions make external governance "a disaster". His advice is to avoid heavily regulated use cases first. He then describes a knowledge layer of vector databases, knowledge graphs and memory, sitting between systems of record and compute. It is always warm, and agents write to it as well as read from it, so it should live near compute at the edge.
Agents, factories, sovereignty. Dell's agent has four parts: an LLM, a knowledge graph, MCP and A2A. In Dell's studies, agents sharing a knowledge graph and told to value consensus ignored a hallucinating peer. Roese says this mitigates hallucination but does not cure it. About 3,000 customers are building AI factories with Dell, and none are yet resilient. Because agents rarely touch primary data, he says resilience priorities and backup models must change, ransomware protection included. He predicts every country will have a sovereign AI strategy, with uses beyond government such as critical-infrastructure backup and government-certified agents. He says quantum is "just accelerating." The host discloses that he and Roese recorded a television ad together for Dell's AI factory with Nvidia.
- Targeted AI let Dell decouple revenue growth from cost. Roese says that in the fiscal year Dell reported in March, revenue grew by $10 billion while absolute cost fell 4%, which he says had not happened in the company's 41 years; he credits fewer than about 30 use cases with ROIs of 10:1 to 30:1, alongside people and process changes. He adds a consulting-firm report he attributes to McKinsey calling generative AI a "BS work detector": if automating a task changes no revenue or cost, the task may not have been worth doing.
- Most so-called agents are agent washing, not autonomous systems. Roese defines agentic AI as the second wave of generative AI, the digitization of skills, where the human gives intent and validation and stays "on the loop" rather than in it. He doubts customers who claim thousands of autonomous agents, and predicts 2026 is when people actually experience them, first on composite multi-team tasks such as generating a quote or resolving a complex support issue.
- A new knowledge layer belongs near compute, not near primary data. Roese predicts the layer of vector databases, knowledge graphs and agent memory between systems of record and compute will dominate 2026 data discussions. He argues it is persistent and always warm, fed from databases as a one-time, non-real-time event but consumed in real time, and that agents write their learned pathways back into it, so it should sit at the edge or device where the transactions happen.
- An agent is a four-part software system, not an LLM feature. Dell's model is an LLM for world knowledge and reasoning, knowledge graphs for specialized knowledge and memory, MCP for perceiving and using tools, and A2A for agents talking to each other. Roese reports Dell studies in which agents sharing a knowledge graph as "truth" and told that consensus matters ignored a hallucinating member; he says this mitigates hallucination without solving it.
- Governance will be central to every enterprise AI discussion in 2026. Roese says over a thousand jurisdictions regulate Dell without coordinating, and calls external governance "a disaster" that NIST and EU frameworks might partly normalize. His tactic is to avoid highly regulated use cases first, and he treats internal governance, choosing what to pursue, as the counter to slop, citing consultancies sued after staff used AI to generate government work.
Loss DogSilver Pairs Trades, Skew and the 1/3 Rule O.L.D. LIVE! 12.29
synopsis
The hosts open with gold and silver reversing hard after last week's parabolic squeeze -- silver down from an $82.70 high to the mid-70s, gold down roughly $200 -- and spend the day arguing that reduced risk, not prediction, is the right response to markets this violent: keep pairs trades on rather than chase direction, don't read skew as a forecast, and treat the choice between saving and spending, or renting and owning, as a personal-circumstance question rather than a formula.
The silver squeeze and pairs-trade mechanics. Both hosts are short silver from an average in the 50s to 60s and have ridden it through an $82.70 peak back down into the 70s, framing rumors about Chinese demand or margin calls as noise rather than an explanation worth trusting. A viewer's question about the gold-silver ratio trade prompts a detailed defense of pairs trading generally: buying one leg and selling the other against a highly correlated instrument reduces risk by roughly 85% versus an outright directional bet, by CME's own statistics, even though this particular spread has swung from about 100 to 55 in a week -- a $10,000 move per contract on the micro-sized legs. A separate question about the "1/3 rule" for credit spreads becomes an explainer on skew: when one side of a spread pays more credit than the other, it reflects the market's pricing of risk velocity, not a directional signal, and traders who read skew as a forecast are drawing a conclusion the pricing doesn't support.
Social Security timing. Two exchanges work through delaying Social Security: one viewer's research that waiting from 67 yields 7% more per year with a break-even age of 78-82 is confirmed as accurate by the hosts' own research, with the caveat that the strategy only pays off if you "wait and live." A caller raises the wrinkle of continuing to work while collecting early, describing a rule that benefits are docked a dollar for every two dollars earned above roughly $24,000; the hosts note the complexity of combining that with 401(k) contributions and defer a fuller answer to later research.
Meme stocks, sports betting, and generational risk-taking. A viewer's question about exchange-based sports betting versus crypto turns into the day's sharpest disagreement: one host calls current sports-betting apps predatory but argues moving betting onto exchanges would improve the product, and defends 2021's meme-stock mania as a net-positive transformational moment that introduced 20-30 million people to markets despite the era's chaos. The other pushes back that he lived through the meme-stock fallout -- blown-up accounts, frivolous lawsuits -- and won't credit it as unambiguously good. The disagreement extends into home ownership, where they debate whether younger buyers priced out of housing are rationally rotating into Bitcoin for higher expected upside.
Hiring value: athletes, degrees, and credentials. A caller's question about whether college athletics factors into employability produces a second flare-up: one host, a former D1 athlete, argues the work ethic required is a genuine hiring signal, while the other insists two decades of hiring shows no measurable difference between athletes, non-athletes, or advanced-degree holders, comparing it to overvaluing a PhD on a resume.
Career calls. A caller with a tastytrade account asks which mechanics from "The Unlucky Investor's Guide to Options" hold up in a small account; the answer is nearly all of them, with the emphasis on staying small and spreading risk across non-correlated underlyings while building experience. A Milwaukee electrical engineer weighing a pivot into financial engineering and coding is told the degree is rare enough to be a strong resume line, particularly for risk-systems programming roles that Chicago's high-frequency trading firms struggle to fill. The show closes with markets still down across the board -- Nasdaq off 166, gold down roughly $200 -- and a reminder that closing trades for 2025 tax purposes should happen soon given T+1 settlement.
- Options skew prices risk, not market direction. Asked about the "1/3 rule" for credit spreads, one host explains why one side of an iron condor can pay $1.80 versus $1.40 on the other: skew reflects the market's perceived velocity of risk, not a forecast, so an uneven credit between the put and call sides should never be read as a directional signal.
- Pairs trades cut risk sharply versus outright bets. One host has run the gold-silver ratio trade since around 66, adding at 56 and 70; even as the spread swung from roughly 100 to 55 in a week (a $10,000 move per contract on the micro size), he cites CME data showing pairs trades reduce risk by about 85% across the board and calls the structure the best approach in an unpredictable market.
- Hosts split on whether meme-stock mania helped young investors. One argues the 2021 boom, despite echoes of predatory sports-betting apps, was a net-positive "transformational moment" that pulled 20-30 million people into markets; the other counters he watched its damage firsthand -- blown-up accounts and frivolous lawsuits -- and isn't ready to call it a win. A viewer poll splits 48% yes to 52% no.
- Hosts clash over whether college athletics predicts hiring value. One, a former D1 athlete, argues the discipline it takes is a real plus on a resume; the other, who says he's hired far more people over the years, insists there's no measurable difference and compares it to over-crediting a candidate's PhD -- what someone actually produces matters, not the credential.
Loss DogMichael Saylor's Crypto Empire is a SCAM. #crypto #oneluckydog #wearelossdog
- The crypto-treasury trade behind MSTR has collapsed. The speaker calls it one of the great scams but stops short of a Madoff comparison. They say every company that copied the model is down about 60%, MSTR included, and name GameStop as a partial exception at about -25% because it has other businesses.
- MSTR's premium to its Bitcoin has turned into a discount. As the speaker recalls the figures, MSTR once traded around $100–110 billion against roughly $67 billion of crypto, more than 50% over its holdings. Now it is about $3 billion under, so it trades like a closed-end fund. They say it had to pay its last dividend out of its own pocket, and call this the outcome they had predicted.
- Convertible debt with no set terms was flawed from the start. The speaker objects that MSTR offered convertible debt without a fixed interest rate or conversion ratio, betting that it could keep pushing the underlying up. They call the structure a death spiral (captioned as "death rate").
- MSTR is a single point of failure for Bitcoin. The speaker's remaining fear, stated as a possibility and not a forecast, is that MSTR's troubles could make the whole Bitcoin market crash. The clip ends before they explain how.
GeoHussarWhy everybody should be paying attention to Silver.. #economics #inflation #silver #learn #finance
- Silver's record price will hit consumers through industry, not jewelry. He puts silver near $80, its highest ever, and argues it is essential to the power grid. As the most conductive metal, it goes into cars, smartphones, appliances, medical equipment and HVAC systems.
- Silver is consumed, while gold is stockpiled. He puts industrial use at 50–60% of annual silver consumption, against only 7–10% for gold. Gold mostly sits as heirloom jewelry, bullion and coins, while most silver is used up, and he presents that difference as the core of the problem.
Dr. Jon KrohnThe Best AI Leaders Never Stop Engineering Hands-On (with Dropbox's Josh Clemm)
- Engineering leaders must stay technical, but not too hands-on. Clemm argues that a leader who codes too much loses time for strategy, people development and org design, while one who is not technical comes across as giving orders from an ivory tower with no reasoning behind them. Side projects fill the gap when the workday doesn't: his LLM search engine Yattle.ai, with a model switcher and query classification that routes queries to models, and an earthquake app the host calls possibly the top one on Android.
- AI makes hands-on use a leadership requirement. Clemm says that telling teams to "go use more AI" without using the tools yourself is why there is a work-slop problem. Leaders who know the tools' limitations can give concrete advice on which tool to deploy, why, and how to handle its gotchas.
Dr. Jon KrohnThe Most Valuable AI Skill Today: Context Engineering (with Dropbox's Josh Clemm)
- Good context engineers think like constrained-systems programmers. Clemm, who has called context engineering the high-status job in AI right now, draws on his early work in embedded software, UAVs and lossy networks: people used to deciding exactly what data ships under hard limits will do well. His image is a Mars rover payload that travels five or six months and had better contain everything it needs.
- Search and retrieval experience transfers directly to context engineering. He argues that pulling the most relevant snippets out of a huge corpus, as Google has done for decades, is the core skill, along with retrieval and recommender-system work. He calls search the "unsung hero" of AI.
- Context engineering also requires verifying outputs with evals. Clemm says people who are good at defining tests up front and tying them to the inputs will thrive, because whatever goes into the context has to be checked against what comes out.
Dr. Jon KrohnThe Real Bottleneck in Agentic AI: Context (with Dropbox's Josh Clemm)
- Bigger context windows degrade quality well before they fill. Clemm says models advertised at 1 million or 2 million tokens show significant degradation when actually used, filling up effectively at more like 100K to 200K. He cites the NoLiMa benchmark, a harder needle-in-a-haystack test where accuracy falls as the document grows, and research showing LLMs "get lost" across multi-turn conversations.
- Dropbox Dash replaced many tools with one super tool. Clemm reports that Dash's accuracy dropped considerably as tools and their retrieved results were added to the context, so the team moved to a single tool over its content index. He says Cloudflare and Anthropic have blogged similar fixes, such as having the LLM write code to invoke tools or pick from a selection.
- The goal is a context packed with exactly the needed data. Too much context gives bad results, and too little makes the model fall back on its training data, which knows nothing about your workplace.
Dr. Jon Krohn952: How to Avoid Burnout and Get Promoted — with “The Fit Data Scientist” Penelope Lafeuille
synopsis
Lafeuille is a senior data scientist at Medidata Solutions who posts as "The Fit Data Scientist". Her position is that burnout in data science comes from overwork and a career pointed the wrong way, and that recovery has to be managed with the same care as work.
From burnout to a reset. In finance in New York she worked too much at the wrong thing. She switched to life-sciences data science and hired a fitness and lifestyle coach. She had been working out almost daily, overtrained and underfed, and now trains four times a week around her work and pickleball, keeping the schedule flexible.
Fuel and rest. She eats one gram of protein per pound of body weight and puts most of her carbs around workouts so she doesn't crash mid-work. Rest matters most: eight hours of quality sleep, reading before bed, short phone-free walks to downregulate, and a five-minute brain dump into a notebook before bed.
Her work and training. At Medidata, which makes clinical-trial software for pharmaceutical companies, she began on immuno-oncology analysis and now builds the backend data science. She uses R and Python and favours R for biostatistics. Her Columbia master's in management science and engineering included company projects that built her portfolio. She also credits ChatGPT and similar tools with getting her past perfectionism about writing in English.
- Her burnout came from overwork and a misaligned career. Lafeuille says that working in finance in New York after her master's, she worked too much at something that didn't match where she wanted to be in five years. She moved to data science in life sciences, the field she cares about, and hired a lifestyle and fitness coach, and credits the two changes with turning both her career and her personal life around.
- Daily workouts without recovery are overtraining, not stress relief. She was working out nearly every day and not fuelling properly. Her program is built around her work schedule and the activities she won't give up, such as pickleball tournaments: four gym sessions a week, usually Monday, Tuesday, Thursday and Friday, kept flexible.
- Eat enough protein and time carbs around workouts. Her two nutrition rules are one gram of protein per pound of body weight, for muscle growth and recovery, and eating most carbs (rice, potatoes, fruit, sometimes ice cream) before, during and after workouts, to avoid a glucose spike and sluggishness mid-work.
- Rest, especially sleep quality, matters most for avoiding burnout. She aims for at least eight hours of sleep, which means more time in bed, and reads instead of scrolling before bed. She takes five-minute phone-free walks between meetings to downregulate during the day, and before bed spends five minutes writing down everything on her mind, which also leaves her a task list for the morning.
- A master's with built-in company projects builds the portfolio. Her Columbia master's in management science and engineering, run jointly with the business school, placed students on roughly three-month projects as employees of real companies, Louis Vuitton among them. She says that answers the question she is asked most on LinkedIn, how to build a strong portfolio.
GeoHussarMerry Christmas everybody. Watch this beautiful prayer with me. #christmas #serbian #orthodox
- A family mocked for its humble Christmas is richer than its mockers. He defends a Serbian Orthodox family whose video, showing traditional bread, fish, a prayer and a candle-lighting, drew ridicule over its modest setting. He argues their closeness, gratitude and sense of culture are things "99% of the population in this country" lack, and closes by greeting them for the feast of Saint Nicholas.
Dr. Jon KrohnWant "AI Slop" Everywhere? Just Rush AI Adoption! (with Josh Clemm)
- Work slop comes from rushed AI adoption without clear goals. Clemm, responding to a Harvard Business Review report the host raised, defines work slop as plausible, impressive-looking content that turns out generic or hallucinated. He blames competitive pressure, citing reports that almost three-quarters of CEOs feel pressed to adopt AI, and says his first question to customers who say "we need AI" is what exactly they hope to accomplish.
- Treat AI output as a first draft needing human revision. His analogy is a college essay written overnight: you don't hand in the first draft, you reread it, strengthen it and add research, and AI output needs the same checks for quality, signal-to-noise and verifiable facts.
- Better work context, not quick MCP connectors, reduces slop. Clemm calls MCP tools a phenomenal way to get agents running but says they are slow, can't reach all content and use many tokens. Dropbox Dash instead ingests, understands and indexes content into both a lexical index using BM25, which he calls "still the workhorse" for queries like part numbers, and a vector index for semantic queries, for hybrid retrieval.
- Internal evals tell you whether a deployment works. He recommends turning customers' goals and target metrics into test sets that define "what good looks like" and serve as an internal benchmark against which AI deployments are judged.
- Use top performers to train the rest of the workforce. Clemm says about 55% of employees don't know how to use AI, so companies should run training, share-outs and demos led by the high performers who are likely already using the tools well.
Dr. Jon KrohnWhy Enterprise Search Is Broken and How Dropbox Is Rethinking It (with Josh Clemm)
- Enterprise search is broken because every app has its own. Clemm says finding something at work means using 12 to 20 separate search bars across apps, and positions Dropbox Dash's universal search as the context layer Dropbox is building for businesses and the first step in many intelligent workflows.
- Multiplayer AI is a teammate that knows you and your team. Clemm contrasts today's "single player" chatbots with AI that understands where you sit in the org and what projects you work on. On the data layer, Dash normalizes third-party content to markdown, extracts text, figures and image content, and builds a graph linking documents, meeting transcripts, Slack and Teams messages and people.
- Collaborative AI applications are the second, more speculative layer. His example is Dropbox's Stacks, where a team gathers URLs, slides and PDFs for something like a board deck in one place, comments on it and queries chat over it together. He says Dropbox and others are starting to explore how multiple people interact with chat agents.
Crosstalk SolutionsDJI has been BANNED
- The FCC has barred sales of new foreign-made drones, DJI included. The host reports that as of December 23rd, 2025 the FCC added all foreign-made drones to its Covered List, alongside Huawei, Dahua and Kaspersky; retailers may still sell models approved this year until that inventory runs out.
- The host sees the ban as resting on no public evidence. He compares it to the looming TP-Link ban: generic national-security claims of surveillance and data exfiltration. DJI, he reports, called the move disappointing and says it asked for security audits three times this year without any agency responding.
- Existing DJI drones still fly, but their software support is in doubt. The host finds no restriction on drones already in the US, but asks why DJI would keep updating software that checks in with the FAA for airspace warnings and flight limits once it cannot sell here.
- He expects consumer drones to get far more expensive. His advice is to buy now if planning to; few US-made drones are on sale today, and he predicts US companies will enter the market and the used-DJI market will "explode".
Dr. Jon Krohn951: Context Engineering, Multiplayer AI and Effective Search — with Dropbox’s Josh Clemm
synopsis
Clemm, Dropbox's VP of engineering, who leads Dropbox Dash and earlier led about 400 engineers at Uber, argues that the bottleneck in workplace AI is context, not model capability. The fix, in his view, is to ingest and index all of a team's work content ahead of time and give the model exactly what it needs and nothing more. He makes the same case for organizations: leaders must stay technical and work on the inputs of team productivity.
Universal search and multiplayer AI. Clemm describes Dash as the "2025 version of Dropbox": search, chat and then agentic features across files scattered over Slack, Google Docs, Jira and dozens of browser tabs. He says traditional enterprise search is broken because workers face 12 to 20 separate search bars. His "multiplayer AI" is not multiple agents but AI that knows a whole team. It rests on a graph of normalized content and people, and on collaborative products such as Stacks. He flags prompt injection and data exfiltration as the security problem with AI browsers from Perplexity and OpenAI, and argues that pre-ingesting content avoids crawling at query time.
Work slop. He attributes slop to pressured, goal-less rollouts. AI output should be treated as a first draft. His other remedies are hybrid lexical and vector indexes (BM25 being "still the workhorse"), internal evals, and training led by high performers. He adds that coding agents speed only one stage of software development, so bottlenecks move upstream to requirements and design.
Context engineering. Clemm says effective context windows fill at 100K to 200K tokens despite 1–2 million advertised. Dash's accuracy fell as tools were added, so it moved to a single super tool. He says Cloudflare and Anthropic have published similar approaches. Good context engineers, he argues, come from constrained-systems work (embedded software, UAVs, lossy networks), from search and recommender systems, and from evals.
RAG and side projects. RAG has moved from single-shot vector retrieval to hybrid and agentic retrieval, as coding agents exploring a codebase show. His roughly 100-line open-source answer engine grew out of a fantasy-football preview generator built on GPT-3 around 2021–2022 and was inspired by Perplexity. It continues as Yattle.ai, which routes queries between models. He argues that citations let an answers engine "prove its work."
Leadership. Past Dunbar's number (about 150), culture has to be replicated through leaders, small-group AMAs and a weekly check-in he began at Uber in the pandemic and still runs. He prefers fixing inputs (goals, strategy, org structure, seniority mix, operational rigor, tooling) to counting PRs. His "think like a farmer" point is that teams need resilience against whiplash from constantly leapfrogging frontier models. Over-reliance on A/B tests, he says, slows big companies down and erodes conviction.
- More context degrades agents, so Dash collapsed its tools into one. Clemm says models advertised at 1 million to 2 million tokens effectively fill at about 100K to 200K, citing the NoLiMa benchmark and multi-turn research. At Dropbox, Dash's accuracy "dropped off considerably" as tools and retrieved results were added, so it moved to a single "super tool" over its content index; too little context, he adds, leaves the model knowing nothing about your work.
- RAG isn't dead; it evolved into hybrid and agentic retrieval. Clemm argues the original single-shot vector lookup has given way to combining vector search with BM25 keyword search, which suits exact matches such as part numbers or function names. It has also given way to agents that retrieve repeatedly, choosing their next fetch from what came back. Coding agents such as Claude Code and Cursor exploring an indexed codebase are his example.
- Multiplayer AI understands a team, not just one user. Clemm calls today's chatbots "single player" and says the effective AI is a teammate that knows your org position and projects. That rests on a data layer in which Dash ingests third-party content, normalizes it to markdown, extracts figures and images, and builds a graph linking documents, messages and people, positioned against the 12 to 20 search bars he says workers juggle.
- Work slop comes from adopting AI without defined goals. Clemm cites reports that almost three-quarters of CEOs feel pressure to adopt AI and that about 55% of employees don't know how to use it. His remedies are to treat AI output as a first draft, build better work context than quick MCP connectors (which he calls slow and token-hungry), write internal evals defining "what good looks like", and have top performers train everyone else.
- Outsourcing every decision to A/B tests breeds organizational lethargy. Clemm says risk-averse market leaders lean on data until conviction disappears, and sees the same risk in people handing decisions to AI. He recalls a fast-moving Uber Eats competitor that turned out to have no A/B testing framework at all, and recommends qualitative customer contact to offset the quantitative data.
Polis PanditFrom Fisherman's Food to Fine Dining
- Bouillabaisse's luxury status was layered on later, not built in. It began as dockside fishermen's food in Marseille — a way to use fish scraps nobody would buy, cooked quickly with no fixed recipe, no plating, and no audience beyond other fishermen.
- The transformation tracks French industrialization: once railroads connected Marseille to Paris, professional chefs adopted the dish for a Parisian audience seeking performance rather than sustenance, fixing recipes, narrowing ingredients, and swapping scraps for saffron and exotic fish.
- He generalizes this into an "optimization trap": professionals cooking or creating for an audience (restaurant chefs, record labels, David Chang shifting from Momofuku's ramen stove to scalable chili crunch oil) inevitably optimize for consistency and reputation in a way originators, who only had to satisfy themselves, never needed to.
- His conclusion is that forgetting origin stories costs more than nostalgia. Using the Michael Jordan-to-LeBron James comparison as an analogy, he argues origins supply motive and context, and losing them — especially as AI reshapes culture — means losing "a piece of ourselves."
Loss DogBalancing Work While Trading Small and Often O.L.D. LIVE! 12.22
synopsis
After a long stretch of Bears-game banter, the hosts settle into a day whose through-line is risk management as a personal, non-formulaic choice: how much capital to raise, how much debt to accept, how actively to trade around a day job, and how much to trust historical averages versus recent ranges. Markets themselves barely move -- "Groundhog Day," one host calls it -- so the substance is mostly in the calls and letters.
Degrees, wealth, and never stopping. A caller stacking multiple degrees and minors is told they still matter to hiring even once internships enter the picture, with one host recalling choosing a candidate with two master's degrees over an otherwise-equal rival. A viewer's open-ended question about "what wealth means" to one host draws a firm answer: he doesn't stop taking risk once he's successful, because stopping was never the point -- risk-taking is what created the wealth in the first place, and he contrasts this with people who fantasize about walking away entirely after one big win.
Global labor competition and the H-1B fee. A caller in southern India describes the scale of new technical universities and companies now paying the $100,000 H-1B fee just to place workers in the US at wages still favorable relative to home-country pay. The hosts say their platform isn't yet built to address non-US labor markets but describe active work building an India-specific dataset, citing a large existing viewer base there; one calls the visa fee "the most ridiculous thing in the world," since it discourages the very talent inflow that benefits the US economy.
Silver, skew, and betting against a friend. A caller notes that a friend of the hosts, a bullish gold-and-silver trader, disagrees with their short-silver position, and asks why they don't just defer to his view. The answer is that no single person's fundamental call outweighs their own "pot odds" read: with silver at record highs, the risk/reward favors the downside regardless of who else is bullish, and the caller's invocation of the 100-year average gold-silver ratio (which would imply silver near $100) is waved off as irrelevant next to the ratio's actual recent multi-year range, currently near its lows.
Fitting active trading around a day job. A Montreal caller inspired by one host's "trade small, trade often" philosophy asks how to combine active options trading with a 9-to-5 job. The hosts differ slightly: one says defined-risk structures (verticals, iron condors, calendars) matter most if you can't watch the screen; the other says strategy choice matters less than position size, since keeping size small is itself the risk control, regardless of what's traded.
Capital raises, startup debt, and job offers. One letter asks about minimum viable Series A size; the hosts say $250,000 is too small for most firms' diligence effort, and that raises of $500,000 to $2 million (or $10-20 million for private equity) are, counterintuitively, easier to close than smaller ones. Another caller offered a half-debt, half-equity startup deal is told to insist on pure equity, since debt in a pre-revenue company carries the same practical risk as equity but caps the upside and has no collateral behind it. A third letter about a job offer conditioned on a six-month trial period draws criticism from the hosts as a symptom of a loosened labor market that they expect to reverse once hiring demand picks back up.
Law school and AI. A caller weighing law school against AI disruption risk is told law is likely the profession most exposed to AI, since much of it is research and precedent-based, but that lawyers won't disappear and following genuine interest matters more than trying to time a fast-changing landscape. The show closes with quiet markets, a note on the holiday trading schedule, and an announcement of the new Lost Dog merchandise store.
- Raising too little capital is harder than raising more. A caller was told $250,000 "looks bad" for a raise; the hosts confirm that smaller checks demand the same due-diligence work from investors as larger ones, so most venture firms want a $500,000-$2 million minimum (private equity firms want $10-20 million) -- asking for more money is often the easier ask than asking for less.
- Recent, not century-long, history should drive the gold-silver ratio bet. A caller notes the 100-year average ratio implies silver should trade near $100 against $4,000 gold; the host dismisses that comparison as irrelevant and says he's trading off the ratio's actual multi-year range, currently near its lows, adding that selling both the call and put side statistically beats selling just one, regardless of which way a friend's fundamental call leans.
- Early-stage startup money should always be structured as equity, never debt. Responding to a caller offered a half-equity, half-debt deal, the host argues an investor takes essentially full risk either way, so debt paying even double the risk-free rate isn't worth it if the company is statistically likely to fail; there's no collateral behind a startup loan, so all the upside protection should come from equity.
- Interim, trial-basis job offers reflect employers gaining leverage. A recent grad's job offer was qualified as a six-month trial before becoming permanent; the hosts call this a real and growing trend driven by a looser labor market, but one flatly calls it disingenuous -- comparing it to "dating" without either side fully committing -- and expects it to fade once hiring tightens again.
GeoHussarThe US Department of Justice has failed to comply with “The Files” transparency law #releasethefiles
- The DOJ has failed to comply with the files transparency law. He says the law, which the president "was forced to sign," explicitly requires the attorney general to release most unclassified documents in the files. What has been published, he says, is heavily redacted.
- Responsibility sits with Bondi, not Patel. He describes Attorney General Pam Bondi as holding one of the five most powerful posts in the country, over the FBI, ATF and DEA. FBI Director Kash Patel, on this reading, is her subordinate, and Congress used the act to get around the department's power.
Crosstalk SolutionsThe Rare UniFi Switch You Can't Buy Anymore
- The discontinued UniFi US-8-150W switch is worth hunting down. The host calls his near-new unit his last one: 8 ports and 150 W, where each port can run either 802.3at PoE or 24 V passive PoE, a choice current UniFi switches no longer offer.
- Older UniFi gear still runs well under a current controller. For unannounced software testing he runs a Cloud Key Gen2 Plus on UniFi OS with a U6 Lite, the switch and a UXG-Lite gateway, plus a Minisforum MS-01 as the server; he prefers the U6 Plus to the U6 Lite today.
AI Analyst LabWhy Unit Tests Break for AI Products
- AI output quality can't be checked with rule-based tests. The speaker argues that traditional software testing works because a deterministic function can check whether a goal was met. Accuracy, hallucination or the quality of a legal draft is subjective and has many dimensions, and none of them reduces to a simple automated check.
AI Analyst LabFuture Proof Your Data Career in 2026
- AI-powered analytics and AI product evaluation are the next focus. Shane names these as the lab's two themes for 2026 and says the themes recur with every guest. One is AI letting people make data decisions with less reliance on a data person. The other is that evaluating AI products needs different tools, a continuous loop, and input from researchers and subject-matter experts, which he thinks data scientists, analysts and ML engineers are well placed for. He announces a 2026 interview series with the founders and leaders building each kind of product (as of December 2025).
AI Analyst LabHow 1 Year of Podcasting Change Us
- The clip asks a question and stops before it is answered. A co-host asks whether the hosts' reason for running the podcast changed over its first year. The first respondent says "an unambiguous yes", and the clip ends before giving a reason.
Y CombinatorThe Truth About The AI Bubble
synopsis
Reviewing 2025 on the Lightcone podcast, the YC partners conclude the AI economy has stabilized into distinct, durable layers — models, applications, and infrastructure — each with a workable playbook, even as the details underneath shifted more than anyone expected.
A changing of the guard among models. Diana reports that in YC's Winter 2026 application data, Anthropic overtook OpenAI as founders' top API choice, after OpenAI held over 90% share in early batches and Anthropic hovered near 20-25% through most of 2024-25 before a recent surge past half. She credits Anthropic's internal focus on coding performance, plus a possible "bleed-through" from founders' own coding habits. Gemini has also risen sharply, from single digits to about 23%.
Model choice as arbitrage, not loyalty. Partners describe both consumers and founders treating models as interchangeable, hand-picking whichever performs best per task — one series-B company reportedly routes Gemini 3 for context engineering into OpenAI for execution — likening it to swapping chip architectures between Intel and AMD.
Reframing the bubble question. Drawing on economist Carlota Perez's installation-versus-deployment framework, the partners argue the massive current capex — data centers, power constraints prompting ideas like space-based data centers and fusion startups — sits on big incumbents' balance sheets, not founders', so a pullback wouldn't strand startups; the resulting compute glut instead sets up a deployment phase for application-layer companies, comparable to excess telecom fiber later enabling YouTube.
Fine-tuning's fading moat, and hiring. They note early fine-tuned "AI-headed" companies got overtaken once frontier models leapt ahead, though newer domain-specific fine-tunes on open-source bases can still beat general models narrowly. They also observe last year's zero-hire, high-ARR startups have mostly returned to normal hiring post-Series A, as rising customer expectations, not headcount efficiency, now seem to be AI tooling's dominant effect.
- Anthropic overtook OpenAI as the top model choice among YC's Winter 2026 batch, a reversal from OpenAI's earlier dominance. Partner Diana says OpenAI held 90%+ share in early batches and Anthropic hovered near 20-25% through 2024 and early 2025 before surging past 50% in the last few months; she attributes it partly to Anthropic making coding performance an internal north-star eval, feeding the batch's wave of coding-agent and vibe-coding startups. Gemini has also climbed from roughly 2-3% to about 23% over the same period.
- Founders increasingly treat models as commodity, interchangeable inputs rather than picking one loyally. Partners describe series-B-level founders building an orchestration layer that swaps models per task based on proprietary evals — one startup reportedly uses Gemini 3 for context engineering and feeds the result to OpenAI for execution — comparing it to how chipmakers Intel and AMD get swapped out as each releases a better architecture.
- The "AI bubble" framing is reinterpreted through economist Carlota Perez's installation-versus-deployment model of tech revolutions. Partners argue the current heavy infrastructure capex — gigawatt data centers, and even fringe bets like space-based data centers and fusion-power startups such as Zephyr Fusion — sits on the balance sheets of big companies, not startups, so a slowdown there wouldn't strand founders; the resulting compute glut is framed as setting up the "deployment phase" for application-layer startups, comparable to how excess telecom fiber from the dot-com buildout later enabled YouTube.
- Early fine-tuning-based moats are proving fragile against leapfrogging frontier models. Partners note a YC company that beat GPT-3.5 through fine-tuning was later "blown out of the water" once GPT-4.5 and 5.1 shipped, while newer domain-specific models fine-tuned with reinforcement learning on open-source bases can still beat general models on narrow tasks — one YC healthcare startup is cited as outperforming OpenAI's models on healthcare benchmarks with only an 8-billion-parameter model.
Dr. Jon KrohnHow Digital Twins Let Us Break AI, Before It Breaks Society (with Alex "Sandy" Pentland)
- Living labs expose the unanticipated effects that matter most. Pentland argues that social media looked wonderful at launch and turned darker over time, and that the only reliable way to see such effects is to have a group of people agree to "live in the future". His example is a town trying out an AI transportation system; the Consumer Reports loyal agents project is applying the same approach to AI agents.
- Digital twins let you test designs without harming people. Since you can't run every experiment on people who have lives to get on with, Pentland describes building a model of how people behave, validating that it predicts what is observed, and then testing designs on it. He says falling compute costs and growing data make this feasible, that aggregation can preserve privacy, and that if the twin "does something really stupid, at least you haven't killed somebody."