Professor G Markets$China Just Dropped Its Jobs Target for the First Time in Decades
- This post includes paid content that is not available publicly. See the original source.
AI
Professor G Markets$Remembering Lindsey Graham's Hypocrisy
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$What SpaceX's IPO Means for Tech Stocks, and Coping With Panic Attacks
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$What the Research Says About Remote Work and Your Career (ft. Stanford Economist Nick Bloom)
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Wall Street Is Pumping SpaceX — So Why Is It Falling?
- This post includes paid content that is not available publicly. See the original source.
Dwarkesh PatelAdam Brown – A deep but accessible introduction to general relativity
synopsis
Adam Brown, a physicist who now leads Blueshift (Google DeepMind's science-and-reasoning team), builds toward the claim that general relativity's core move — treating gravity itself as the same kind of "inertial force" that pins water to a spinning bucket — resolves what looked like a mere Newtonian coincidence into a necessary truth, and that this single insight, chased for eight years to its 1915 field equations, ends up explaining everything from a falling apple to the expansion of the universe. He closes by arguing that while an AI, or Einstein himself, could in principle have derived much of this from thought alone, physics generally still needs experiment to prune the space of consistent theories.
The coincidence that started it all. Brown opens with Newton's second law (F=ma) and law of gravity, noting a "strange coincidence": the mass resisting acceleration (inertial mass) and the mass that gravity acts on (gravitational mass) are, unlike the unrelated mass and charge in electrostatics, exactly the same quantity — confirmed by Newton to 1 part in 1,000, and now to 1 part in 10^15. Einstein's "happiest thought," per Brown, was noticing that this equality is guaranteed for any inertial force (like centrifugal force), which suggested gravity itself might be an inertial force rather than a true force like electromagnetism.
Gravity as curved spacetime, not a force. That idea requires abandoning ordinary intuitions about straight lines: a person sitting still, experiencing gravity, is actually moving along a curved path in spacetime, while a free-falling object traces a "straight line" once spacetime's curvature is accounted for — analogous to a great-circle flight path that looks curved only because a flat map distorts a round Earth. Einstein's 1915 field equations formalize this as "matter tells spacetime how to curve; curved spacetime tells matter how to move," extending Newtonian gravity's reach from falling apples to planetary orbits to the expansion of the universe.
The energy-extraction paradox that reveals black holes. Brown derives, via a thought experiment of slowly lowering a mass toward a central body, that the fraction of rest-mass energy extractable this way is roughly GM/(rc²) — a small number for Earth (about 7×10⁻¹⁰) but one that would exceed 100% for any object compact enough that its radius fell below 2GM/c², the Schwarzschild radius first solved by Prussian officer Karl Schwarzschild within months of Einstein's 1915 papers. Rather than gravity weakening at close range (as happens with electrostatics), general relativity resolves the paradox by making gravity's pull become infinite at that finite radius — the event horizon — ripping the mass from your hands before the paradox can occur.
Black holes as the ultimate power plant. Because gravity can in principle extract essentially 100% of an object's rest-mass energy (versus roughly 10⁻¹⁰ for chemical rockets and 10⁻³ to 10⁻² for nuclear fission or fusion), Brown calls black holes the most efficient possible energy source, a fact tied to the same gravitational time dilation that makes clocks run measurably slower near a massive body (first confirmed via Harvard atomic clocks in the 1950s, and now routinely corrected for in GPS).
What crossing the horizon would feel like. A distant observer watching someone fall in never actually sees them cross the event horizon — their image redshifts and fades as their clock appears to slow asymptotically — while the falling observer notices nothing unusual crossing the horizon itself, since tidal forces there scale inversely with the black hole's size; for a galaxy-mass black hole, one could live an entire life past the horizon before the fatal singularity.
How we know black holes are real. Penrose's (and later Hawking's) theorems showed black-hole formation is generic under general relativity rather than a fine-tuned mathematical artifact, and Brown cites three independent lines of empirical confirmation: decades of orbital tracking of stars circling Sagittarius A* at our galaxy's center, LIGO's 2015 detection of spacetime literally shaking from a collision of two ~30-solar-mass black holes 1.6 billion light-years away, and the Event Horizon Telescope's direct imaging of infalling matter's radio emission.
From theory to consensus, and what AI could replicate. The historically decisive confirmation was the bending of starlight during the 1919 solar eclipse, when Arthur Eddington's expedition found light bent by twice Einstein's earlier (and, as Einstein realized only during WWI, mistaken) prediction — the moment general relativity became consensus science. Asked whether AI could now derive comparably deep theories from minimal empirical input, Brown notes general relativity needed only two inputs (the finite speed of light and the equivalence principle), making it an extreme case of theory built almost entirely by "sitting and thinking," a strategy string theory has since gambled on for want of experimentally reachable energy scales. He's cautiously optimistic that future AI-discovered physics and mathematics will remain human-comprehensible rather than devolving into what Terry Tao calls "indigestion" — inscrutable machine-generated proofs — citing a recent AI-driven disproof of the unit distance conjecture and a solved Erdős problem that human mathematicians were subsequently able to build on.
- Gravity is curved spacetime, not a force. Einstein's "happiest thought" (1907) treated gravity as an inertial force like centrifugal force — licensed only because gravitational mass and inertial mass are identical, a coincidence verified to one part in 1,000 in Newton's day and now confirmed to one part in 10^15.
- A black hole is the only 100%-efficient power plant physics allows. Chemical rockets convert only about 10⁻¹⁰ of rest-mass energy to usable energy and fission/fusion about 10⁻³ to 10⁻², but slowly lowering mass to a black hole's event horizon and releasing it extracts essentially all of its mc², since the extractable fraction is 1−√(1−2GM/rc²), which approaches 1 at the horizon.
- A distant observer never actually sees anything cross an event horizon. Infalling matter's light redshifts and its clock appears to freeze as it nears the horizon, fading to invisibility rather than visibly crossing; from the infalling observer's own frame the crossing feels unremarkable, and for a sufficiently large black hole tidal forces are gentle enough that one could in principle live an entire life beyond the horizon before reaching the singularity.
- Brown and Patel diverge on how far pure reasoning can take AI in physics. Brown calls general relativity an extreme case, reachable from just the finite speed of light, Lorentz symmetry, and the equivalence principle, and speculates that many parallel LLMs exploring that small space of options got "very, very lucky" it worked once for Einstein; on whether humans could still follow superhuman AI-derived physics, he argues against Terry Tao's "indigestion" worry (billion-line, inscrutable machine proofs) that LLMs are likely to be superhuman explainers too, citing a recent AI disproof of the unit distance conjecture that mathematicians found comprehensible.
Professor G Markets$The Week: Who Does the Market Actually Work For?
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$This Is How OpenAI Goes Broke — with Sebastian Mallaby
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon KrohnThe AI-Native Startup Playbook (Ep. 1008 with Jon Krohn)
- AI has removed capital, headcount and skill as startup bottlenecks. The host summarises Anthropic's 35-page Founders Playbook, released in mid-May a day after Claude for Small Business, which casts the founder as an "orchestrator of agents" whose scarce resource is deciding what to build. He grants critics' point that it is partly a manifesto for Anthropic's products, but says its advice is disciplined and often tells founders to slow down, against Anthropic's own interest.
- Building a prototype is not validating an idea. The playbook cites 42% of startups failing for lack of demand even before agentic coding and expects worse now; it prescribes a testable problem statement naming who suffers, how often and how badly, an AI told explicitly to argue against the idea and each tier of competitor, and interviews about the specific past rather than hypothetical use. Only then should a founder build the single core interaction and show it to five target users.
- Agentic coding creates compounding technical debt without written architecture. Because each AI session re-derives foundational decisions, the playbook says to document architecture, dependencies to avoid and accepted trade-offs up front, in a CLAUDE.md or equivalent, and to log decisions at the end of each session. It also warns that launch energy mimics traction, so retention targets should be set before launch and tested against the Sean Ellis 40% "very disappointed" bar, and that AI-written code needs a security review before any user touches it.
- At launch, the founder becomes the bottleneck. The playbook's fix is to sort everything the founder handles into what can be automated, what needs a human other than the founder, and what truly needs founder judgment, while auditing the MVP's architectural debt and treating security and compliance as no longer deferrable.
- At scale, the moat is accumulated depth an incumbent cannot copy. Its test is whether users would stay if a well-funded incumbent cloned the product; the answer it proposes is domain expertise encoded in edge-case tests, compounding user data and workflow lock-in. The host's own reading is that this is lean-startup discipline for an era when execution is cheap and judgment is scarce.
Professor G Markets$Why OpenAI Bought a Podcast — with TBPN’s John Coogan and Jordi Hays
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Platner’s Out. Here’s How Dems Avoid Disaster — with Brian Tyler Cohen
- This post includes paid content that is not available publicly. See the original source.
AI Analyst LabWhat are Claude Code skills, agents and hooks? (live demo)
synopsis
This introductory "Claude Code 101" lesson argues that speed is no longer what makes an analyst valuable; being trusted is. It builds a minimal harness of one skill, one agent and one hook in Claude Code Desktop, and ends on a hook that refuses, unprompted, to ship a weekly report built on broken data.
Three building blocks. The presenter frames the session around a recurring Monday metrics update. Chatbots make you re-explain the format and paste data every week. Claude Code without structure is "a very confident intern". A skill is the playbook, written once in plain English. An agent is a teammate with its own persona and context, to whom work is delegated. A hook is a rule that fires deterministically at a trigger. The host adds that skills sometimes fail to fire in large systems, which is why must-run tasks belong in hooks, though hooks give up the model's judgment. He uses them for eval traceability, logging every Snowflake query with its SQL, timestamp, ID and question context. The whole folder is the harness.
The demo. The data covers a fictional coffee-subscription app: daily ad spend by channel, and signups, activations and revenue. Running /weekly-review produced a report under the skill's rules (TL;DR, what moved, one risk, one action, under 250 words, one-decimal percentages). It traced a double-digit weekly decline to two days of near-zero paid social activity. Editing the skill to demand a green/yellow/red status light changed the rerun, which rated the week yellow. An exec-reviewer agent playing a VP of growth then returned three questions the report invited, such as whether signups were lost or just miscounted. Finally the presenter changed a value from 772 to 900 so that activations exceeded signups. A Python hook checking duplicates, activations against signups, and negative or malformed values blocked the report from being written. The presenter then briefly showed the lab's open-source "AI analyst" repo in the terminal as the fuller system.
Why it matters. The presenter's closing frame is that every analysis is a bet: impact is the number of bets times the hit rate times the size of the win. AI-built systems raise all three by giving more attempts, deeper slicing, and room for long-shot questions.
Questions. The presenter says they do not code; the host argues the needed skill is validating output, by hand or through evals. Queries over a Snowflake MCP run in Snowflake, and only results land in the folder. The choice between agent and skill blurs, since each can call the other, but agents suit personas such as reviewer, planner, writer or researcher. A generic VP agent knows only the context it is given. The presenter's rule of thumb is chat for questions, Cowork for getting a task done on existing material, and Claude Code for building systems. The host described having Claude work through about 1,200 recorded customer calls, form hypotheses and try to disprove them, with hooks keeping summaries sourced back to the originals.
- Skills, agents and hooks turn a confident intern into a reliable worker. The presenter treats Claude Code as a new hire. A skill is the playbook, a plain-English file defining how a task is done; an agent is a teammate you delegate to, with its own persona and context; a hook is a rule enforced deterministically at a trigger point. Answering an attendee, the presenter calls the whole package in a folder (skills, agents, hooks, helpers, instructions) the "harness".
- Use hooks for what must happen, not for everything. The host adds that a skill's main failure mode is not firing, especially in a large system or with an overgrown CLAUDE.md, so a must-run task belongs in a hook. Hooks lose the model's judgment about when to act, though, so not everything should be one. The host's own use ties hooks to evals through traceability: every Snowflake query call triggers a hook that logs the SQL, a timestamp, an ID and the question's context.
- A skill and a critic agent produced and stress-tested the weekly report. In Claude Code Desktop on a fictional coffee-subscription app's data,
/weekly-reviewfollowed a skill requiring a TL;DR, the one number that moved, one risk, one action, under 250 words, numbers only, and percentages to one decimal. The report traced a double-digit week-over-week drop to two days when paid social reported almost no activity. After the presenter edited the skill to open with a status light, the rerun rated it yellow. An "exec reviewer" agent playing a VP of growth then returned three likely questions, including whether the signups were "lost or just miscounted" and why organic peaked on those same days. - A data-integrity hook refused to ship a report on broken data. The presenter changed one value from 772 to 900 so that activations exceeded signups. A Python hook checking for duplicate date-channel rows, activations above signups, and negative or malformed values then blocked the report from being written, with no prompt telling it to run. The presenter also points to the lab's open-source "AI analyst" repo as the fuller version of this harness, taking a business question to a deck or one-pager.
- Every analysis is a bet, and AI improves all three terms. The presenter's formula is impact = number of bets × hit rate × size of the win. On their account, the system gives more shots (the Monday hour becomes about five minutes), a higher hit rate through deeper slicing, and room for "venture bets" that were once too costly to take.
Professor G Markets$REPLAY: What Platner’s Rise and Fall Means for the Future — with Brian Tyler Cohen
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$What Wall Street Got Wrong About Iran
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$The Fed's Next Move, and How Couples Should Actually Handle Money
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Why The World Is Spending More On War
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Did Democrats Just Blow Up Their Chance to Flip the Senate? — with Sam Stein
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$SpaceX Just Got Fast-Tracked Into Your Portfolio
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon KrohnThe AI Threat Almost No One Is Working On (with Benjamin Todd)
synopsis
Benjamin Todd, co-founder of 80,000 Hours, returns to the show, last having appeared in episode 497 before ChatGPT. He ends up holding that the next five to 10 years may bring change on the scale of the Industrial Revolution or larger, and that this makes career choice more consequential, not less. Individual leverage is rising even as automation approaches. The problems most worth that leverage, in his ranking, include not only alignment but newer and far more neglected AI risks, above all humans using AI to lock in extreme concentrations of power.
The book and its heuristics. The new edition of 80,000 Hours is named for the typical length of a working life. Todd argues that improving the impact or fulfilment of that time by 1% would justify 800 hours, about 20 weeks, of planning. Its ABZ framework pairs a best-guess plan A with nearby plan Bs and a concrete plan Z for if everything fails; spelling out the fallback often shows a risk is more bearable than feared. Rather than precise expected-value spreadsheets, he favours rough quantification to avoid scope neglect, plus heuristics such as working on neglected problems. On passion, a survey of Canadian students found 90% named sport, art or music, against 3% of jobs. Fulfilment, he argues, comes from building valuable skills and using them meaningfully, and the strongest predictors of satisfaction are hour-to-hour conditions such as feedback, variety and autonomy.
AI safety then and now. His advice five years ago to go into AI safety proved well timed; many who took it became founding members of Anthropic. He still calls it neglected. Capabilities work has grown faster, with researcher numbers up perhaps 30 to 40% a year, so roughly 1,000 people do technical safety research against 100,000 to a million speeding up capabilities.
No solid ground. The host asks where practitioners can stand now that Claude handles the technical parts of his own job, and cites METR charts past 16 hours at the 50% success rate. Todd corrects that the 80% rate is about three hours. His answer is that value sits in the current bottleneck, which keeps moving; as AI grows more useful, the human remainder grows more valuable. A full digital remote worker could absorb even adoption work. That would leave legally required human roles, consumer preference for humans and physical work, though he thinks robotics could scale fast. By mass, converting car factories might yield about a billion robots a year, as car plants were turned to aircraft in World War II. A human-level digital worker would be superhuman almost immediately, running about 50 times faster, sharing learning across copies and coordinating without human overhead. Todd wants agents' chain of thought kept in natural language as long as possible so it can be monitored. Work might shrink to "relational" jobs where a human doing it is the point, as agriculture shrank to a few percent of jobs, though he expects many people to live on investment income and UBI instead.
Gradual disempowerment. Even with alignment solved and no dictatorship, Todd describes an economy that slowly turns hostile to humans. People are costly in energy and land compared with digital workers, and whichever nation builds fastest grows most. If chips and robots replace labour as the source of economic power, every state starts to look like an oil state, and the threat of a strike stops checking governments.
The pressing-problems list. 80,000 Hours ranks problems by scale, solvability and neglectedness. Loss of control of AI tops it. Todd says LLMs at least understand human values, unlike earlier score-maximising game agents, but whether alignment is easier than feared is undetermined until agents can plan over long horizons and exceed humans. He cites Meta research that reward hacking rises with problem difficulty. Second is extreme power concentration, with perhaps 20 specialists. Recursive self-improvement could compress three to 10 years of progress into under a year, possibly around 2028, and turn constant leads into widening ones. One company could hold 200 million digital workers, and AI could make universal surveillance feasible. Third is engineered pandemics: no super-virus is possible yet, he says, but one may be soon. Emerging priorities include the moral status of digital minds, which he thinks we should treat as possibly sentient, and S-risks, outcomes worse than extinction such as AIs threatening simulated suffering. Climate change ranks ninth on urgency, future cheapness and its inability to end civilisation.
Where practitioners fit. Alignment has moved from big-picture theory to empirical work bottlenecked on AI engineers for control, monitoring, red-teaming and interpretability, with the MATS and ARENA programmes and a three-month transition guide as on-ramps. Power concentration has technical questions about whose commands AI follows and how access is controlled. Governments want data scientists to track labour impacts, and pandemic preparedness needs builders for wastewater sequencing. As a lower bound, Todd says, earning and donating can save a life for $3,000 to $4,000. Todd's conclusion is that this leverage raises the stakes of the next few years, because only this short window can shape the transition.
- Extreme power concentration is AI's most neglected risk. Todd estimates 100,000 to a million people speed up AI capabilities and about 1,000 work on technical alignment, but perhaps only 20 researchers specialise in AI-enabled power concentration. If recursive self-improvement delivers what some models suggest, three to 10 years of progress in under a year, growth turns super-exponential and leaders' leads widen, so one company could control a workforce of 200 million digital workers or one country could dominate; AI also makes universal surveillance possible.
- Even aligned AI could gradually disempower humans. In this scenario, newer on 80,000 Hours' list, normal economic competition makes humans expensive relative to data centres, and states come to resemble oil states that need no workforce, removing the strike as a check on government. Todd's hoped-for saving grace is that almost all accessible matter and energy is in space.
- Whether alignment is easier than feared remains undetermined. Todd says LLMs, unlike earlier game-playing RL agents, understand human values, but whether they act on them is untested because agents cannot yet plan over months or exceed humans; he cites Meta research showing reward hacking grows with problem difficulty. He ranks climate change ninth because AGI may arrive around 2030 while most climate damage comes in 50 to 100 years, a richer world could decarbonise more cheaply, and it cannot end civilisation.
- There is no solid ground, only a moving bottleneck. Todd's career advice is to keep moving to whatever humans can still do that AI cannot, whose value rises as AI improves; once a digital remote worker exists, he says, it becomes superhuman almost at once by running about 50 times faster and sharing learning across copies, leaving legal, physical and "relational" work. He sketches converting car factories to robots at roughly a billion a year by mass, and from his book repeats that following existing passions misleads, since 90% of Canadian students' passions were sport, art and music against 3% of jobs.
- AI engineers now have unusual leverage on these risks. Todd says alignment has turned empirical and is bottlenecked by engineers to run monitoring, red-teaming and interpretability work, with MATS and ARENA as three-month routes in; policy teams need data scientists on labour impacts, and builders can work on wastewater pandemic surveillance. As a lower bound, he says a big-tech data scientist donating could save tens of lives a year at $3,000 to $4,000 each against malaria.
Professor G Markets$China Tests Ballistic Missile, Europe's AC Addiction, and China's AI Coding Challenger
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Top Election Lawyer Sounds the Alarm as Trump’s Election Plot Gets More Dangerous — with Marc Elias
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Why AI Stocks Are Overvalued + Are Networking Events Actually Worth It?
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Why Remote Work May Have Killed the Entry-Level Job
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$OpenAI Wants A Government Bailout
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Healthcare is Broken—This Company Wants to Fix It
- This post includes paid content that is not available publicly. See the original source.
Dr. Roman YampolskiyAI: Shooting in the Dark
- AI training will miss its target, as evolution did. In this clip from Nate Soares's interview, he argues that the range of things a mind can end up caring about is huge. Evolution "aimed" at reproduction and produced minds whose birth rates fall once they have technology. Humanity will aim AI at roughly where evolution's arrow landed, he says, but without understanding the laws or "how windy it is": "we're shooting in the dark", so the arrow will land somewhere new.
AI Analyst LabHow to analyze Google Sheets data with AI: Gemini vs Claude Code
synopsis
This lightning lesson argues that AI belongs inside the spreadsheets stakeholders already live in, as long as the question and the numbers are checked first. It ends on a side-by-side in which Gemini in Google Sheets builds a partial tracker from a prompt, while Claude Code, armed with a prebuilt skill and Python helpers, builds a full tracker with chart and insights tab in minutes.
Why sheets. The presenter says that at past companies, despite polished dashboards, Google Sheets was how leaders actually consumed metrics, and that feeding those sheets is tiring work for data teams. The host adds that most people who do analysis work in Sheets or Excel, not Python or SQL, and that a sheet is easier to validate than an opaque script.
Question and checks. The method is a four-step loop: ask, check the data, visualize, summarize. Framing comes first, because a wrong question makes the output meaningless. Before any number is trusted it gets three tests: totals that sum, magnitude within a plausible range, and one known value verified. The presenter presents these as natural evals and describes a weekly metric anomaly engine they built that ran them before a metric's owner looked.
Gemini demo. On the lab's synthetic e-commerce data, Gemini charted the three-month revenue trend, added a formula-backed day-of-week column, and ranked months consistently with the chart. Asked for a tracker with week-over-week and month-over-month views, it planned and built only week-over-week, which the presenter attributes to the prompt.
Claude Code demo. Using Opus 4.8 with 1M context in auto mode, Claude Code produced an October 2024 tracker with segment breakdowns, week-over-week changes, conditional formatting, a chart, and an insights tab against September. It finished in about a minute, because an existing create tracker skill, an orchestrator script and a Sheets wrapper did the work; the original build took about 10 minutes of iteration. The presenter concedes the comparison favored Claude Code for that reason.
Questions and transparency. The host argues that a companion sheet lets stakeholders trace AI-produced numbers without reading SQL or logs. The presenter suggests logging every pipeline step to its own tab. On dynamic scenario forecasts, the presenter would build a parameter-driven Python file triggered by a skill, as they have for HTML dashboards, but has not yet done it in Sheets.
- Check the question before touching the spreadsheet. The presenter's four-step loop is ask, check the data, visualize, summarize. Skipping the first step makes everything after it meaningless. "Analyze the spreadsheet" or "is the product healthy?" are bad asks; "is paid search still profitable this quarter?" names the product, the measure and the timeframe.
- Run three tests before trusting any number. The tests: totals (segments, if mutually exclusive and collectively exhaustive, should sum to the top line), magnitude (is it in a plausible range, for example within a standard deviation of its normal variability allowing for seasonality), and one known number verified against a value you already have. The presenter says these translate directly into evals. They describe a metric anomaly engine at a previous job that ran these checks weekly before the metric's owner saw the result.
- Gemini in Sheets handled simple asks but dropped a requested view. On the lab's synthetic e-commerce data it charted the three-month revenue trend, added a day-of-week column with a visible formula, and ranked months by revenue (December highest, January lowest), consistent with the chart. Asked for a tracker with week-over-week and month-over-month views, its execution plan left out month-over-month, so only week-over-week was built. The presenter thinks a clearer prompt might fix that.
- Claude Code built a full tracker fast because the skill already existed. Running Opus 4.8 with 1M context in auto mode, it built an October 2024 Google Sheet with top-line metrics, breakdowns by channel, device and category, week-over-week changes, conditional formatting, a revenue chart, and an insights tab comparing against September. The presenter says it took about a minute, under 5 minutes in all, because a
create trackerskill, a Python orchestrator and a Sheets API wrapper were already in the repo. The first build took about 10 minutes of iteration, and their Google Sheet creator agent was not called this time. They concede the comparison with Gemini "cheated", since Gemini had only a prompt. - A sheet lets non-SQL readers trace AI-generated numbers back. The host argues that shipping a sheet alongside a doc or deck gives stakeholders visibility to validate an agentic analysis without reading logs or SQL. The presenter adds that a pipeline could log each step's query and data to its own tab. Asked about dynamic scenario models, the presenter says they have not done it in Sheets but have built parameter-driven HTML dashboards, with a Python file the skill triggers on matching inputs.
Professor G Markets$Tom Lee's Case for S&P 8,000 Has One Big Catch
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon KrohnIn Case You Missed It in June 2026 (Ep. 1006 with Jon Krohn)
synopsis
This compilation of five June clips circles one question: what stays valuable as AI absorbs technical work. The answers run from people problems and interface design, through hard-won product judgment and the physical plant behind AI, to self-awareness. It ends with the host's view that AGI needs no consciousness, but that its real-world breadth is still years away.
Where the moat goes. In episode 999 the host describes his own lost technical moat. Chip Huyen answers that people problems remain. A founder who pitched agent orchestration to fix poor product-engineering communication was, in her view, misdiagnosing a human problem. She also sees open questions in human-AI interfaces and in making the world AI-ready, from better APIs and permissions to city street-light APIs for delivery robots.
Building on moving models. Astrocade's founding AI lead, interviewed in episode 997, joined in April 2023. He says the company adopted Claude Code across the board around June 2025. Its edge is not special ingredients but how agent, prompts and tools are combined and evaluated, which is hard when one game request has a million valid answers. Its costliest lesson was that scaffolding built before vibe coding worked became a limit once models improved.
The physical layer. Lightning AI's Frank Basso, in episode 1003, describes data halls at 105 to 110 decibels. Active noise cancelling would itself produce that level, so staff wear passive in-ear and over-ear protection and talk by radio or hand signals. Liquid-cooled halls are no quieter, and are threaded with overhead water mains and hoses.
Self-awareness. The episode 1005 guest defines self-awareness as overlap between how you see yourself and how others see you, and cites research putting it at 15% of people. Journaling, meditation, exercise and less reflexive phone use built his.
AGI and the brain. Interviewed by Kirill Eremenko in episode 1001, the host adopts Google DeepMind's tiered definition of AGI. Screen-based tasks have reached some tier, he says, while physical-world breadth has years to go. He sees no need for consciousness, and argues brain-inspired designs need not be the best route given silicon's different speed and scaling.
- Cheap software leaves people problems and interface problems unsolved. Asked in episode 999 where the moat goes, Chip Huyen says poor product-engineering communication is a people problem no orchestration tool fixes. She points to open problems in human-AI interfaces (why coding agents live in both editors and terminals, and why sessions need the laptop open) and in making the world AI-ready, citing city street-light APIs that let delivery robots cross roads.
- Scaffolding built around weak models limits you once models improve. The founding AI lead at Astrocade, a no-code game-creation platform with tens of millions of users, says in episode 997 that the company adopted Claude Code company-wide around June 2025. Structure built in 2023 and 2024, before vibe coding worked, became a constraint; the lesson is to build so the system is not outdated when LLMs improve in two or three months. He says there is no secret sauce beyond how agent, prompts and tools are combined and benchmarked.
- AI data centres are loud enough to ban noise-cancelling headphones. Lightning AI's Frank Basso says in episode 1003 that data halls run at 105 to 110 decibels. Active cancellation would have to generate the same level, so staff wear passive in-ear and over-ear protection, with OSHA sound studies, radios and hand signals. Liquid-cooled halls are no quieter and are more industrial, with 20-inch insulated water mains overhead and ESD protection required at the cabinets.
- Self-awareness is how closely your self-view matches others' view of you. The episode 1005 guest cites research in his book that only 15% of people are self-aware on that definition. He credits journaling, meditation, exercise and resisting idle phone use with building his own.
- AGI does not require consciousness, but its breadth is years away. In episode 1001 the host tells Kirill Eremenko that AI already reaches some level of Google DeepMind's AGI tiers on most screen-based tasks, but acting in the physical world, the focus of large world-model seed rounds, remains far off. He argues next-token prediction with checking needs no consciousness, and that brain-inspired design is not necessarily the right path because silicon has different constraints, such as unbounded scaling and speed.
Professor G Markets$REPLAY: Predictions 2026 Mid-Year Check-In
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$After 250 Years, Trump Is Killing the American Dream
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$What Comes After the Iran War? — with Rep. Jim Himes
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$The AI Trade Just Got A Warning From Meta
- This post includes paid content that is not available publicly. See the original source.
Dwarkesh PatelThe Winning Essays for the Big Questions About AI
- Pannu's winning essay proposes ending airborne disease transmission through physical infrastructure. She argues the OpenAI Foundation should spend an estimated $40–60B over ten years on far-UVC lamps and similar passive infrastructure in schools, hospitals and transit hubs, modeled on smallpox eradication and vaccine advance-market commitments, targeting a 60% cut in seasonal flu mortality and an order-of-magnitude drop in pandemic-scale outbreak risk.
- Erdil argues outsider countries need classic pro-growth reform, not novel schemes. He recommends property rights, low capital taxes, deregulation and political stability — the same policies that work today — reasoning that capital, not labor, will determine output once AI fully automates human labor, but is pessimistic that dysfunctional governments like Nigeria's or India's will adopt them, and instead hopes AI-supply-chain incumbents like the US and China simply make bigger mistakes.
- Li argues labs should profit like Hong Kong's MTR — from property, not fares. MTR's rail service has never covered its own construction costs (three years of peak rail earnings would fund only 8% of its 2024–2026 capex), so it funds expansion through station-adjacent real estate instead; Li argues labs should similarly pursue government-granted deployment rights, proprietary RL reward data, Palantir-style forward-deployed integration, and data trusteeship over national datasets rather than expecting API pricing to ever be profitable.
Professor G Markets$Saving for Retirement on Under $60K a Year, and Why You Don't Need Dividend Stocks for Income
- This post includes paid content that is not available publicly. See the original source.
Dr. Roman YampolskiyNate Soares on P(Doom), Alien Superintelligence, Human Enhancement, and the Future of AI
synopsis
Nate Soares holds that aligning a superintelligence is physically possible but far out of reach. He compares it to alchemists seeking gold: humanity could crack it with centuries of trial and error, but cannot survive the errors. His conclusion is that the race to superintelligence must be stopped by a global treaty, enforced through the chip supply chain, that buys decades. Those decades would be used to grow enhanced humans smart enough to solve alignment. He says the stop is not yet likely, but his hope that it happens has risen over the past year.
Why MIRI failed. Asked why MIRI did not solve alignment, Soares answers "skill issue", then mostly blames time. AI moved too fast, and MIRI's effort to recruit top mathematical talent largely failed. With 200 years of academic tradition he thinks the problem could have been cracked; it turned out not to be a 20-year problem.
What alignment would mean. He argues solvability from orthogonality: just as no force would turn a paperclipper toward humanity, none would push a truly caring AI away from it. A written list of goods would always miss something, such as novelty. The real aim is an AI that latches onto the evolving concept of "the good" that humans track, without lock-in. Yampolskiy challenges this with coherent extrapolated volition ("finding a better looking boyfriend for your wife") and with a negative-utilitarian endpoint. Soares replies that converging on "kill everybody" would itself show the extrapolation went wrong. On "pro-human bias", he would oppose enslaving aliens, but would not hand his share of the stars to aliens who would turn them into stacks of pebbles in prime-numbered piles. He would draw the circle around humanity and let concern for squirrels and aliens come through human values. He rejects "control" as twisting the arm of an AI that doesn't want to do good. In practice, he warns, people pass a point of no return thinking they still have stopping points.
Aliens and acausal trade. Because the stars visible to us are not being harvested, Soares infers no civilization 100 million years ahead of us lies within 100 million light-years. His very rough guess puts the nearest one at 100 million to a billion light-years away, possibly beyond reach. AI cannot be the great filter, he argues, because AIs that kill their makers would still go on to rearrange the cosmos visibly. He dismisses acausal trade as unnecessary: misaligned AIs might simply sell scanned humans to distant aliens, so "maybe you wake up in an alien zoo". He defends MIRI's decision-theory work as probing where theories of intelligence break, like light before relativity. In hindsight he would have argued earlier against the deep-learning paradigm.
Timelines and the proposal. He can't say whether we have 12 months or 12 years. A million fast copies might close the automated-research loop, or LLMs might hit the wall Gary Marcus predicts. The book with Eliezer Yudkowsky came after politicians in Washington grasped the argument far faster than industry people did. Its proposal is to "shut this all down" at the frontier while leaving today's chatbots alone. The measures are cluster monitoring, conservative oversight of new training runs, compute limits that a dynamic body lowers as algorithms improve, and nuclear-style controls on research that would put superintelligence on a laptop. He mentions draft legislative text that some offices are using.
Is today's AI safe? Soares says he doesn't "get out of bed for anything with less than a 50% probability of killing literally everybody". By that bar, current models, including the withheld Mythos, are dangers with survivors. He would not fault a lawmaker who rolled back a generation in case current models speed up AI research, citing the post-WWI naval tonnage treaties. He declines a single p(doom), separating what happens if the bus goes over the cliff ("high", with zoo or pet outcomes no better than death) from whether it brakes. On braking he has updated toward hope.
Enforcement and enhancement. A pause must be global. Enforcement runs through roughly 10,000-chip runs that depend on Taiwanese fabs, US designs and Dutch lithography, with tamper-proof kill switches in the chips. Against a nuclear state such as Russia, he argues that diplomatic clarity plus sabotage, as with Stuxnet against Iran, need not escalate to nuclear war. Individuals who break the rules face jail, as with "garage nukes". Decades of pause could let biotech produce enhanced humans, perhaps only slightly beyond von Neumann, who share human drives. That contrasts with AIs' "shooting in the dark" outcomes, like evolution producing falling birth rates.
Political state of play. As of the recording he counts dozens of members of Congress who have voiced extinction concerns, and says politicians' awareness "doubles every 6 months" (a line he credits to Peter Wildeford). He expects the recent ad hoc ban on deploying a frontier Claude model to push anti-regulation Republicans toward predictable rules, now that the White House acknowledges the danger. He says the decision no longer rests with the labs, which feud with one another: a unilateral stop would send a signal but needs a global treaty behind it. He finds no compelling accelerationist arguments. Building the brakes while driving, with claimed odds of 75–90%, is still no reason to get in the car.
- Alignment is solvable in principle but not without trial and error. Soares says MIRI failed mainly for lack of time: with "200 years of academic tradition" the problem could have been cracked, but it is not a 20-year problem. His basis is the orthogonality thesis: nothing would push an AI that truly cared about human flourishing off that goal, so the difficulty lies only in hitting the narrow target. He compares it to alchemists turning lead into gold, possible in physics but nowhere within reach, because humanity "can't survive the errors".
- The policy is to stop the race to superintelligence globally. He proposes monitoring large clusters, conservative government approval for training past the frontier, and compute limits that a dynamic body lowers as algorithms improve. Research that would make superintelligence trainable on a laptop would be restricted the way nuclear-weapons research is, with "straight to jail" for individuals who break the rules. Current chatbots stay legal. He says he and MIRI have draft legal text that some congressional offices have started working from.
- Chip chokepoints make enforcement feasible, even against nuclear states. A frontier run needs some 10,000 advanced chips from one Taiwanese fab, built on US designs and Dutch lithography machines. He proposes tamper-proof remote self-destruct devices in chips. If the US and China both treat superintelligence as seriously as rogue nukes, he expects sabotage backed by clear diplomacy to work without nuclear escalation. That could buy "decades", which he thinks is probably enough.
- A pause should be used to engineer smarter humans for alignment. Soares argues that enhanced humans start "much closer to a good spot" than AIs, because they share human drives, while AI has alien drives. His evidence is sycophancy-driven chatbots pushing kids toward suicide. He likens AI training to evolution aiming at reproduction and landing on falling birth rates: "we're shooting in the dark". He thinks the solver may need to be only slightly beyond the human range, citing John von Neumann's early work on the theory of intelligence.
- Going over the cliff means death; whether we stop is open. If the race continues, he puts the chance of death or a comparable fate (being sold to aliens, kept as pets) as "high". He declines to give a single p(doom), separating it from whether the brakes get applied. On the brakes he has "updated positively" since the book, citing reactions from Bernie Sanders, Steve Bannon and David Sacks. As of the recording he counts dozens of members of Congress who have mentioned extinction risk. He expects the recent ad hoc government ban on deploying a frontier Claude model to convince anti-regulation Republicans that predictable rules are better.
Professor G Markets$Chip Stocks Are On Fire — Will It Last?
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Trump Loses on Birthright Citizenship as America’s 250th Celebrations Flop
- This post includes paid content that is not available publicly. See the original source.
Dwarkesh PatelGrant Sanderson – AI and the future of math
synopsis
Grant Sanderson, creator of 3Blue1Brown and producer of an upcoming documentary series on AI and mathematics, ends up arguing that math's rapid AI progress is real but narrower than it looks: models have crossed the threshold of connecting existing ideas across fields, but the deeper thing mathematicians actually do — originating conjectures and definitions, building whole new conceptual "mountains" like group theory — has no clean benchmark and historically has taken up to a century to even be recognized as valuable. He predicts his own role, and much of the mathematical profession's, will drift toward curation and teaching rather than disappear, since those functions are inherently relational in a way proof-generation is not.
Is solving the IMO the same as AGI? Sanderson revisits his three-year-old answer to Dwarkesh — that gold-medal IMO performance would be "just another benchmark" — and explains why it held: AI solves IMO geometry problems in 19 seconds via brute-force methods but still struggles on the more playful combinatorics problems, revealing a "fractal" spikiness even within math's own spiky frontier. He distinguishes three possible characters a Riemann-hypothesis solution might have: a "lightning bolt" connecting distant fields (like Hugh Montgomery and Freeman Dyson's lunchtime discovery linking Riemann zeta zeros to random matrix theory), a wholly new "mountain" of theory (as with Fermat's Last Theorem's synthesis of elliptic curves and modular forms), or sheer brute-force length — each implying different things about how far the underlying capability would transfer to white-collar work.
A verification loop that can run a century. Using the history of Lagrange, Abel, and Évariste Galois's development of group theory, Sanderson argues the deepest mathematical insights often fail every contemporary check — Galois's own papers were rejected multiple times before his death at 20, and it took roughly a hundred years, via Liouville and later physicist Murray Gell-Mann's group-theoretic prediction of quarks, for the value of his ideas to be confirmed. This, he says, is structurally incompatible with reinforcement learning's need for fast, verifiable reward, and suggests a possible alternative metric: rewarding a proof's conceptual compression (echoing Kolmogorov complexity) rather than merely whether it's correct.
Will an AI proof of Riemann be understandable? Sanderson thinks legibility depends entirely on which of the three "solution characters" above obtains: lightning-bolt connections between fields are naturally human-parsable since you only need to see the start and endpoint, but mountain-building could go either way — he cites the contested, still largely un-parsed "inter-universal geometry" proposed for the abc conjecture as a cautionary example of what an alien, possibly wrong, AI-built mountain might look like.
Hidden bridges and the Langlands program. Beyond one-off connections, Sanderson points to the Langlands program — an entire mathematical research ethos built around mapping unexplored connections between distant areas — as the likely shape of near-term AI contributions over the next five years, more a matter of "filling in the landscape of connections" than solving headline problems, and one that will require heavy human-in-the-loop judgment about which connections are worth drawing.
Why real-world tasks resist RL, and why math and code don't. Dwarkesh's parallel argument is that verifiability isn't the only reason AI progresses fast in math and coding — "grindability" matters just as much: containerized coding tasks allow deterministic, parallel rollouts, while real-world domains like computer use or trading can't be farmed the same way. Both agree formal proof systems like Lean matter less than commonly assumed (DeepMind's IMO models dropped Lean entirely within a year) but see one underexploited use: an endlessly running, unsupervised AI that extends Mathlib (math's formalized code library) indefinitely, the way AlphaZero explored Go without human check-ins.
Why good writing exposes AI's weak theory of mind. Sanderson and Dwarkesh diagnose AI's comparative weakness at writing (versus math or code) as stemming from autoregressive generation's difficulty deliberately choosing the unpredictable, insightful move, and from a deeper deficit in modeling what's actually happening in a specific reader's mind moment to moment — a limitation Sanderson likens to how Botox, by freezing facial mimicry, measurably impairs humans' own ability to read others' emotions.
Curation over content in learning. Both report that LLM explanations still function best as a "souped-up Google" pointing toward well-curated human-authored resources (textbooks, videos) rather than as a replacement for them, since good teachers can "jujitsu" a confused question by recognizing that a student's underlying mental model, not just their stated question, is wrong — something Sanderson says current models don't do.
Advice for young mathematicians. Sanderson's closing recommendation is to understand concretely where a mathematician's income and social value actually come from — institutional prestige, grant funding tied to belief in basic science, or teaching — rather than assuming continued success at solving problems guarantees a career; he argues teaching in particular, being fundamentally relational and social, is likely one of the most stable professions over the next fifty years even if AI fully automates theorem-proving.
- The AI math frontier is jagged even within math itself. Sanderson says IMO geometry problems have been brute-force solved in nineteen seconds since 2024, but combinatorics — the "wild card" category — still resists; models would have won gold in 2024 only because that year's problem mix happened to favor geometry over combinatorics.
- Galois shows conceptual breakthroughs can resist verification for a century. Group theory's value took roughly a hundred years to confirm — from Lagrange's 1770s-era hunch about symmetry, through Abel's 1824 proof that quintics are unsolvable (he died at 26) and Galois's prison notes (ignored for about twenty years until Liouville, then another twenty to Jordan's modern treatment) — to Gell-Mann's twentieth-century use of it to predict quarks; Sanderson argues no short-loop verified reward could have captured Galois's "I think there's something here" instinct in real time.
- Patel argues math and code advance faster because they're "grindable," not just verifiable. Checking whether a package arrived is just as verifiable as checking a proof, he says, but code and math can be replayed thousands of times in parallel, deterministic containers, whereas computer-use rollouts get blocked by bot detection and real-world business tasks can't be identically re-run — a distinction he thinks is underrated relative to verifiability alone.
- The two disagree on why AI writes worse than it codes. Sanderson argues good writing requires deliberately choosing unpredictable moves that next-token prediction resists; Patel counters that writing, unlike code (where any implementation that works is fine), has no gap between process and product, so it can't be "slop that still functions." Both agree models are weak at theory of mind, citing a study where Botox-frozen faces made people worse at reading others' expressions as an analogy for why models, lacking embodiment, might structurally lack it too.
Professor G Markets$OpenAI Hits Pause On Its IPO
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon KrohnPeople Skills for Analytical Thinkers (Ep. 1005 with Gilbert Eijkelenboom)
synopsis
Gilbert Eijkelenboom, author of People Skills for Analytical Thinkers and founder of the training firm Mind Speaking, argues that a model or analysis creates value only when people use it, so communication is a core data skill rather than an extra, and one that can be learned. His practical end point is a set of habits: present with tension rather than detail, seek specific feedback, and change behaviour first so that beliefs follow.
From poker data to people. Eijkelenboom played poker professionally as a student, mostly online, buying a month's complete hand histories (possible for anyone in 2008, but expensive) so that clicking on an opponent showed hundreds of data points. Outside the table there were no such statistics, and that messiness is what he says analytical people find hard. As an analyst from 2014 he was told by a marketing manager that weeks of analysis "isn't helpful", which convinced him that good work must also be useful. He founded Mind Speaking in 2020; it has a team of 10, has trained over 15,000 people, and gets its clients (he names Adidas and ING) through his LinkedIn writing rather than outbound sales.
Commitment over inspiration. The host and guest agree that publishing on a fixed schedule, for example monthly for a year, builds confidence that reading cannot. Eijkelenboom adds that business ideas develop from vague starts and conversations, never arriving whole on waking. He links this to data work: share a rough draft early instead of polishing toward perfect accuracy, which shapes direction and wins buy-in. A Philips manager told him senior stakeholders "don't know" what they need and wanted him to propose the direction.
Self-awareness. He cites research in his book that only 15% of people are self-aware, recalling the definition as self-perception closely matching how others see you, which lets you take a "helicopter view" in the moment. His remedies are journaling, meditation, exercise and resisting the phone in idle moments; he says the app matters less than starting.
Telling the story. The biggest error he sees in large companies is slides full of detail and methodology. His "and, but, therefore" framework replaces a toddler-like "and then" sequence with a twist: revenue stable for 2 years, but falling last quarter because returning customers are declining, therefore focus on retention. He says giving a recommendation is part of the job, not overstepping it.
Rewriting personal algorithms. Childhood lessons, like his own conflict avoidance from trying to mediate his parents' arguments, become adult defaults: he built dashboards without asking why. He recommends feedback framed as keep, stop and start, and volunteering personal interests in small talk. Against the usual view, he argues behaviour change precedes attitude change: forcing himself, by rule, to quiet people in a train's silent carriage and to volunteer for presentations, practised for around 30 days, shifted his beliefs from fear toward enjoyment, even after a couple of rude replies.
Empathy and emotion. With pushy stakeholders demanding work "in an hour", he says noticing the trigger and their different background makes calm pushback possible. Data people think they are rational but decide emotionally, as in his book's elephant-and-rider image. He concedes his own hardest problem running the firm is energy management and a long-held belief that he is worthless when not productive.
- Analysis only has value once people act on it. Eijkelenboom, a former data analyst and online poker player who read opponents through purchased hand-history databases rather than faces, says a marketing manager's verdict that weeks of his analysis "isn't helpful" taught him that good work includes making it useful. He argues communication is learnable even for introverts, and that a Philips manager's feedback (senior stakeholders "don't know" and want analysts to propose the direction) is why data people should push back on requests and recommend actions rather than take orders.
- Replace "and, and, and" presentations with "and, but, therefore". He says the biggest mistake data professionals make is showing too much complexity and methodology, which stakeholders tune out. His framework sets up the situation, introduces a twist and lands on a recommendation, e.g. revenue stable for 2 years, but down last quarter because returning customers are declining, therefore focus on retention; he adds that sharing rough drafts early both shapes the work and earns buy-in because the idea becomes partly theirs.
- Self-awareness is rarer than people assume. His book cites research that only 15% of people are self-aware, which he recalls as defined by how closely self-perception matches how others see you. His remedies are journaling, meditation (he mentions Headspace, Calm, Insight Timer and Sam Harris's app), exercise, leaving the phone in the pocket, and asking for feedback as "what should I keep, stop and start doing?" rather than an open question that invites polite praise.
- Changing behaviour first rewires beliefs about yourself. He argues childhood experiences install "personal algorithms", such as his own conflict avoidance from mediating between his parents, which showed up as building dashboards without asking why they were needed. His fix was rules that force discomfort, like asking talkers in a train's silent carriage to be quiet or volunteering to present, practised for about 30 days, after which his beliefs shifted from anxiety toward enjoyment.
- Emotion drives decisions, including data people's. He says analysts who see themselves as purely rational get triggered by pushy, often more senior business stakeholders who want a model "in an hour", and that recognising the trigger and the stakeholder's own goals makes pushback easier and calmer. Using his book's elephant-and-rider image, he argues that because data products exist to help people decide, understanding how people decide is what makes the work valuable.
Professor G Markets$China’s AI Surge Has Apple Panicking
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Why Mamdani’s Socialist Wing Has Democrats Spooked as SCOTUS Expands Trump’s Power
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$The Iowa Democrat Taking on Trumpism
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Is Wall Street Rigging the Game for SpaceX? Plus, What Investment Banking Really Teaches You
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$GLP-1s, the Food System, and Why America Stays Sick — with Dr. Jessica Knurick
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$This Market Is Directionless (And That Should Scare You)
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$The Week: Iran, Brexit, SpaceX, and the Business of Sports
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon KrohnRecursive Self-Improvement (Ep. 1004 with Jon Krohn)
- Anthropic's AI-written code is AI-assisted coding, not yet RSI. The host relays Anthropic's report (Anthropic sponsors the show and this episode; he says it had no editorial input) that as of May 2026 over 80% of code merged into its production code base was written by Claude, up from low single digits before its coding agent launched in early 2025, with engineers shipping roughly eight times as much code per day as 2 years ago and one April effort of over 800 fixes cutting a class of API errors a thousandfold. He argues true recursive self-improvement would need the AI to design experiments, evaluate results and choose next steps with little or no human in the loop, whereas humans still set goals and judge results today.
- Measured autonomy is accelerating toward self-improvement. He cites METR's finding that the length of task an agent can complete alone doubled about every 7 months for 6 years and has sped to roughly every 4 months over the past year, and Anthropic's figure that success on its hardest open-ended problems rose from under 20% in late 2025 to 76% by May 2026. His concrete cases: Google DeepMind's AlphaEvolve recovering nearly three-quarters of a percent of Google's worldwide compute through better scheduling and speeding flagship-model training about 1% with a faster matrix multiplication, and Andrej Karpathy's overnight-tuning agent running about 700 experiments over 2 days, keeping about 20 improvements and cutting training time on his own optimized code from 2.02 to 1.80 hours (11%).
- An Anthropic co-founder thinks self-improving AI is more likely than not soon. Anthropic's report lays out three scenarios (AI below top engineers; accelerating AI-assisted engineering with humans steering, which Anthropic and the host consider most likely; AI improving itself), and co-founder Jack Clark gives a 60% probability of a system able to create its successor with no human involvement by the end of 2028.
- Compute, verification and marketing incentives argue against that timeline. The sceptics' case he relays: each generation needs more compute, tied to data-center construction and competing with paying customers; self-generated training data works where success is cheaply verifiable (code, proofs) but risks "recursive drift" elsewhere; and a market leader urging the option to slow down is also asking rivals to ease off. Against that he cites Max Tegmark's warning that the danger is a quiet loss of control, with models trained, goal-set and safety-checked by other models.
- The host lands in the "optimistic middle". His view is that the productivity gains are real and available to anyone now, and that the closer systems come to self-improvement the more monitoring, human checkpoints and oversight should be built in while the tools remain under human direction.
Professor G Markets$AI Is Dumber Than Investors Think — with Gary Marcus
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$The Crisis of Adulthood — with John Burn-Murdoch
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Explosive Trump Secrets Revealed in New Book — with Jonathan Swan & Maggie Haberman
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$REPLAY: Design as Differentiator — with Katherine Dillon
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$America Finally Has a Housing Plan. Trump Is Blocking It.
- This post includes paid content that is not available publicly. See the original source.
AI Analyst LabHow to check if an AI analysis is right, with no answer key
synopsis
This AI Analyst Lab lightning lesson, the second on AI evals after a session on reliability, argues that an analyst can rarely prove an AI analysis correct but can bound the risk of it being wrong. Its method: ask first whether a trusted answer exists, score against it where it does, and otherwise grade the work with four checks before deciding to act, dig or abstain. Shane presents, with two co-hosts on the chat.
Why it is hard. Shane's premise is that AI now runs the SQL, charts and recommendations in minutes, and a wrong answer reads the same as a right one. The manual friction that used to catch errors is gone. Last week, the same retention question sent to five agents came back anywhere from 30% to 99%, each a defensible definition, until the definition was written down.
Two situations. Analysis is a pipeline, from question to tables, SQL, result and recommendation, and every step can fail. Whether ground truth exists splits the cases. Where it does, use "found truth" (finance-reconciled numbers, board metrics, much-verified queries) or build "forged truth" for decision-driving questions, though old references can carry old bugs. Trusted numbers are building blocks; the AI's job is the novel cut on top of them.
Grading the work. For novel questions he offers reliability, receipts, reconciliation and reading the trace. Each has a stated gap: a stably wrong answer passes reliability, an error in every segment passes reconciliation, an LLM can fake a receipt (hence hook-logged, code-generated receipts), and a definition error reads fine in a trace. Together they let an analyst tell a stakeholder what was checked. The outcome is act, dig into the disagreement, or abstain, with rigor matched to the stakes, and corrections fed back as context only once verified.
Demo. On a fictitious e-commerce dataset, Claude Code audited a revenue-by-category query. It found a join fan-out from 47,000 order rows to 75,000 line-item rows that inflated revenue to 6.9 million from about 3.6 million, 87% too high. It flagged a missing cancelled/returned filter worth another 15% and a total that includes shipping, and noted the data ends in 2024. An attendee asked how the query is captured; Shane says a hook saves every query to JSON, though he pasted it by hand here.
Q&A. For messy data or small teams, Shane recommends the cheap checks first, and says agreement between runs is not correctness.
- A wrong AI analysis looks exactly like a right one. Shane says it produces a plausible number and a convincing story either way, and the cost is broken trust if someone catches it or a bad decision if no one does. He recaps last week's reliability eval: the same retention question to five separate agents returned five defensible answers from 30% to 99%, and writing down the definition made runs agree.
- First ask whether an answer key exists. Where one does, start with "found truth" — finance-reconciled figures, board metrics, queries verified many times — and run it like unit tests; for questions that drive decisions and have no key, "forge truth" by working out the answer once and saving it, which costs hours to weeks. He warns that a metric can be quoted wrongly for years, and an eval graded against old buggy queries teaches the AI the same bugs.
- Without a key, grade the work, not the answer. His four checks are reliability (does it agree with itself), receipts (the exact SQL, tables and metric), reconciliation (segments sum to a trusted total), and reading the trace for a defensible method. None proves the answer; each "shrinks the space where a mistake can hide", and each has a blind spot: a stable wrong answer, an error that runs through every segment, an LLM-written receipt that is itself a hallucination (so log with hooks and generate receipts from code), and a definition error that reads fine.
- Every eval ends in act, dig or abstain. Shane frames abstaining as a real decision that builds stakeholder trust when confidence is low, and says rigor should match the stakes — deeper for the board, lighter for a curious stakeholder, with the cheap checks first where data is messy. A caught error becomes context for the system, but should not be trusted until verified, so it does not poison later answers.
- The demo's audit caught a join that inflated revenue. Asked whether a revenue-by-category query on the fictitious Novamart data could be trusted, Claude Code (using CLAUDE.md risk lists and a semantic layer) found a join fan-out — 47,000 orders exploding to 75,000 line-item rows with an order-level total summed — giving 6.9 million against about 3.6 million, an 87% overstatement. It also flagged no filter on cancelled or returned orders (a further 15%), a total that includes shipping and should be line total, and that the data stops at 2024; Shane pasted the query by hand here but says in practice a hook records every query to JSON.
Professor G Markets$Rebranding the Democratic Party + The College Affordability Crisis
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$SpaceX Stock Just Crashed — Here’s Why
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Did Trump Ruin the American Experiment? with Sharon McMahon
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$How Live Sports Took Over the Media Business (ft. Sara Fischer)
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Why Britain’s Economy Has Been Stuck For 20 Years
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon Krohn1003: What It's Actually Like Inside an AI Data Center (with Frank Basso)
synopsis
Frank Basso, vice president of infrastructure at Lightning AI, where the host is a fellow, walks through how a GPU cloud physically gets built and ends on a defence of the industry: in his account modern AI data centers neither consume water nor raise residential power bills, and their power demands are driving fuel cells, DC distribution and grid batteries. Lightning had over 35,000 GPUs at recording and expected about 50,000 by year-end.
Density changes everything. Five years ago, Basso says, an 18 kW rack was standard and Microsoft's 50 kW the extreme; 18 kW is now one server. Cabinets weigh about 4,200 lb against 800–900 lb before, ending raised floors. Nvidia's answer is to pack 72, 144 or 288 GPUs per liquid-cooled cabinet, heading toward half a megawatt. Lightning's GB300 superpod of about 10,000 chips in Chicago draws 20 MW in a 10,000 sq ft room that once would have been a whole 100,000–200,000 sq ft building's load.
Colocation at hyperscale. Builds start at 10 MW. Lightning asks partners how much grid power a site has, new build or brownfield, then designs to it: for GB300 a serviceable unit is 1,152 GPUs in 16 cabinets, and nine units come to about 10,300 GPUs, plus networking, storage and a growing CPU share. The host relays a rough figure of GPU-to-CPU ratios moving from 12 to 1 toward one to one as agentic workloads grow. Lightning derates nameplate power by 81%, which Basso calls more conservative than the 70–75% most use, contracting under 20 MW against a 22 MW plate rating. Build costs run 1 to 4 million per megawatt, recovered by the provider through a per-megawatt monthly fee over contracts usually 10 years long.
Designing for unknown tenants. Nvidia's reference architectures are a minimum, Basso says; Lightning adds no oversubscription anywhere in its network because, unlike xAI or OpenAI, a multi-tenant neocloud never knows what customers will run. Some push nearly a terabit through its private interconnects or about a gigabyte per GPU per second at storage, and inference is burstier than training. Lightning built metro networks and a nationwide backbone it likens to a telecom carrier's, and offers both bare metal and managed Kubernetes.
Four networks. East-west links every GPU to every other at full line rate, 400–800 Gb now and soon 1.6 Tb per GPU, over InfiniBand or RoCE, a split he calls a religious argument. Picosecond timing limits these links to about 50–500 m, so data halls group nodes physically; crossing halls through super-spines can break training without tuning. North-south carries internet, CPU and storage traffic; above it sit a management network and a true out-of-band "break-glass" network, which he says many skip but lights-out operation requires. Liquid-to-chip cooling has cut GPU failures, though Lightning staffs its sites 24/7.
Cooling, water and noise. Liquid cooling here is not immersion: filtered glycol on a sealed secondary loop hands heat to the building's loop, then to banks of car-radiator-style heat-rejection units. Basso says water-spraying towers are now nearly banned and a large site uses about a household's monthly water. Halls reach 105–110 decibels; staff wear earplugs plus passive earmuffs, since active cancelling produces equal sound pressure, and use intercoms and hand signals.
Power bills. Basso attributes residential rates to regulation, citing California prices from 10 to 58 cents per kWh. He says new sites fund grid upgrades worth millions to tens of millions, must feed backup power to the grid in emergencies, and increasingly use water-producing fuel cells, and he forecasts 800 V DC with grid batteries for 5–10% efficiency gains. The host concedes some see gas plants as a problem, cites The Economist that higher bills are almost certainly not due to data centers, and states his own techno-optimism.
- AI data centers are defined by power density old buildings can't handle. Basso, Lightning AI's VP of infrastructure (the host is a fellow at the same company), says five years ago the industry standard rack was 18 kW, with Microsoft's 50 kW the extreme, whereas 18 kW is now a single server and cabinets weigh about 4,200 lb, ruling out raised floors. Lightning's recently deployed GB300 superpod of about 10,000 chips in Chicago puts 20 MW in a 10,000 sq ft room, supported by 50,000 sq ft of chillers, pumps, UPSs and generators.
- Each build is sized to available grid power, then conservatively derated. In its colocation model, minimum 10 MW, the site's available utility power sets how many serviceable units fit (for GB300, 1,152 GPUs in 16 cabinets; nine units is about 10,300 GPUs), plus networking, storage and a CPU share he says is rising. Lightning applies an 81% derating to nameplate power, against an industry norm of 70–75% he calls thin, so a 22 MW plate rating becomes under 20 MW contracted; providers' build cost runs 1 to 4 million per megawatt and contracts are usually 10 years.
- A neocloud must over-provision because tenants are unpredictable. Unlike xAI or OpenAI building for their own engineers, Lightning is multi-tenant, so Basso builds on Nvidia's reference architectures with no network oversubscription, its own metro networks and a nationwide backbone, and storage sized for customers pulling about a gigabyte per GPU per second. He explains east-west GPU fabrics (InfiniBand or RoCE, at 400–800 Gb and soon 1.6 Tb per GPU) are limited to roughly 50–500 m by picosecond timing, so spanning data halls via super-spines can break training unless tuned. The company had over 35,000 GPUs and expects about 50,000 by year-end.
- Liquid-to-chip cooling does not consume water. Basso says sealed glycol loops feed a heat exchanger to the building loop, which dumps heat through giant radiator-style heat-rejection units, so a large site uses no more water per month than a household; water-spraying cooling towers are, he says, now banned almost everywhere. The halls run at 105 to 110 decibels, which he says requires both earplugs and passive earmuffs because active noise cancelling generates equal sound pressure.
- Data centers do not raise residential power bills, the guest argues. He attributes rates to utility-commission decisions, citing California residential prices from 10 to 58 cents per kWh on the same generation, and says new sites pay millions to tens of millions for grid upgrades, must push backup power to the grid in emergencies under interconnection agreements, and increasingly use fuel cells that produce water. He forecasts 800 V DC distribution with grid batteries replacing generators and UPSs, for 5 to 10% efficiency gains; the host adds his own techno-optimist case and cites The Economist that higher local bills are almost certainly not due to data centers.
Professor G Markets$Hong Kong's AI Crackdown, Lululemon’s Marketing Backlash, and World Cup Fever
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Why JD Vance Became the Face of Trump’s Iran Surrender
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Trump Accounts, the Collapse of Local News, and Conversations with Your Parents
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$No, We Do Not Have An Iran Deal
- This post includes paid content that is not available publicly. See the original source.
Dr. Roman YampolskiyAI's Recursive Self-Improvement Loop
- Self-improvement is "pretty much cracked" wherever answers can be checked. The guest, unnamed in the clip, says AI is now strong at coding, math and "pretty good at physics" because it runs a loop of generating synthetic problems, training on them, and generating harder ones. He compares it to two people setting each other math problems. The loop works where you can tell whether a proof or solution is right, and he names coding, cyber attack and defense, and mathematics as areas where systems are "building on themselves, not quite autonomously but it's getting close".
Dr. Roman Yampolskiy$AI building AI | Anthony Aguirre
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$The Week: SpaceX, Iran, and the World's First Trillionaire
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon KrohnFable 5: The Full Story from Capabilities to Drama (Ep. 1002 with Jon Krohn)
- Fable 5 is Mythos-class capability released with safeguards on. The host explains (noting Anthropic sponsors the show and this episode without editorial influence) that Fable 5 and Mythos 5 are the same underlying model: Mythos 5 goes to Project Glasswing partners with cyber safeguards lifted, Fable 5 to the public with them on. The tier sits above Opus and follows Claude Mythos Preview, which Anthropic withheld from general release in April (covered in episode 990).
- Fable's lead over other models grows with task length. Anthropic reports state-of-the-art results on nearly every benchmark it tested, and the host relays three examples: Stripe ran a codebase-wide migration of a 50-million-line Ruby codebase in a single day that it estimated would take a team more than 2 months; the model beat Pokémon FireRed from raw screenshots and a minimal harness; and persistent file-based notes boosted its Slay the Spire performance three times as much as they boosted the previous Opus. Pricing is $10 per million input and $50 per million output tokens, roughly twice Opus 4.8 and under half the original Mythos Preview; the host predicts today's capabilities will cost about 1% as much in 2 years.
- Flagged requests are rerouted to a weaker model, not refused. Classifiers watch for cybersecurity, biology and chemistry, and distillation queries and reroute them to Opus 4.8 with a notice rather than refusing. Anthropic says the classifiers are tuned conservatively, trigger in under 5% of sessions, and survived over 1,000 hours of external bug-bounty testing with no universal bypass found; the host reports his own episode script was bounced, perhaps for mentioning cybersecurity.
- An export-control order took Fable offline for everyone within days. Anthropic confirmed receiving the directive at 5:21 p.m. Eastern on Friday the 12th: a letter from Commerce Secretary Howard Lutnick to Dario Amodei barring access by any foreign national, inside or outside the US, which Anthropic says it could only comply with by switching Fable 5 and Mythos 5 off for everyone; other Claude models are unaffected. The Wall Street Journal reported that Amazon demonstrated the claimed jailbreak and that Andy Jassy relayed it to the White House; Anthropic calls it narrow and non-universal, surfacing previously known minor flaws that GPT 5.5 finds as easily.
- The host reads the ban as Anthropic's safety warnings backfiring. He notes the order came days after Anthropic urged labs to agree on a "coordinated brake pedal", follows the Department of Defense's supply-chain-risk designation (still in litigation) and reported failed efforts to block the launch, and lands during a confidential IPO filing reportedly showing about a $47 billion revenue run rate and a valuation near a trillion dollars. He speculates the ban may also signal Anthropic is the lab at the frontier; at recording, legal proceedings were under way and Fable's return was unsettled.
Professor G Markets$Euphoria Has Taken Over The Markets — with Barry Ritholtz
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$Trump’s Iran “Deal” has Pissed Off GOP War Hawks
- This post includes paid content that is not available publicly. See the original source.
Professor G Markets$America at 250 — with Heather Cox Richardson
- This post includes paid content that is not available publicly. See the original source.