Coordinated with Fredrik

Coordinated with Fredrik

Download on the App Store

Coordinated with Fredrik episodes

  • From the Erie Canal to the Welding Robot

    This is another all-AI episode. Two synthetic hosts — Sam and Maya, rendered in the v2 version with ElevenLabs v3 — walk through 200 years of financial engineering and try to make the case that what we’re doing at Sourceful is the third chapter of a very old story. The voices are not mine. The argument is.

    It’s one of the longer pieces we’ve put out because I think the throughline only lands if you actually trace the arc. The v2 master removes the inter-act AI songs and keeps the episode focused on the argument. The shortcut is below.

    The argument in one sentence

    Financial engineering was born from the physical — canals, railroads, mud and iron. It spent 150 years abstracting away from it. Then in 2008 the abstraction collapsed under its own weight. And now, with C&I batteries and the rest of distributed energy, it’s coming back to the physical — but armed with 200 years of risk theory it didn’t have the first time.

    That’s the whole episode. Everything else is detail.

    The Erie Canal was a project finance deal

    In 1817 the State of New York authorized the issuance of bonds to build a 363-mile ditch from Albany to Buffalo. The federal government refused to fund it. Private capital wouldn’t touch it. Critics called it “Clinton’s Folly.”

    The state did something that now looks like an early project-finance prototype: it issued stock certificates and bonds tied to future tolls from a canal that did not exist yet. A specific asset. A designated revenue stream. Bondholders relying on both economics and government backing.

    That’s the recognizable shape of project finance. In 1817. Without spreadsheets, without option pricing, without anyone in the room having heard the word “structured.”

    It worked. The original $7M estimate was paid off by 1836 from tolls. Freight rates from Buffalo to New York fell by something like 90%. New York City became America’s financial capital because of this canal. And then — because financial engineering is contagious — every other state borrowed too. Between 1820 and 1840 American states borrowed almost $200M in then-current dollars for canals, railroads, and banks. John Wallis at NBER calls it a sovereign-debt experiment inside a federal system.

    In 1841 and 1842 the experiment defaulted. The clean NBER count is eight states plus the Territory of Florida in those two years; broader early-1840s summaries often describe the crisis as nine state defaults once later defaults and restructurings are included. The aftermath wasn’t just financial reform — it was constitutional reform. Debt limits got written into state constitutions. Voter approval requirements for borrowing. Mandatory tax provisions linked to debt issuance. We are still arguing about what constitutes a public purpose — every IRA debate, every TIFIA loan question, every PPP framework — and that argument starts there.

    The pattern that gets established here repeats every generation: financial engineering enables ambitious infrastructure; ambitious infrastructure blows up; the blow-up generates regulation; the regulation defines the next era. Boom and rule. Boom and rule. Boom and rule.

    Railroads invented the modern capital stack

    The Pacific Railway Act of 1862 gave the Union Pacific and Central Pacific federal loans plus alternating sections of land along the right of way. The size of several US states, given as collateral.

    That collateral mattered, because railroads were the first assets that broke conventional finance. They were specific (you can’t move a railroad), long-lived (40-year depreciation), indivisible (half a bridge isn’t useful), and capital-intensive with slow-arriving revenue. That description, by the way, is also exactly what a 4-hour C&I battery in a German auto plant looks like in 2026.

    Nineteenth-century railroad finance solved this problem by inventing what we now call the capital stack. Senior first-mortgage bonds with a lien on track and tolls. Junior mortgage bonds below. Debenture bonds below that, unsecured. Income bonds that only pay coupons if the company has the cash. And finally common stock — equity — at the bottom.

    Different tranches priced differently. Sold to different investors with different risk appetites. A trustee oversaw the collateral package. Sinking funds accumulated for repayment. Hierarchy of loss-allocation: senior gets paid first, junior absorbs the pain.

    That architecture — developed in the nineteenth-century railroad bond market — is recognizably the same design pattern used a hundred years later to issue mortgage-backed securities, and the same design pattern now appearing in C&I battery portfolios. The vocabulary changed. The design didn’t.

    The Panic of 1873 — the defining crisis of the Gilded Age — came from too much railroad investment, too much leverage, too long a chain of claims. Pay attention to that phrase. The chain of claims got too long.

    Finance leaves the physical (1933 – 2008)

    After the Depression, the state rebuilt finance: Securities Acts, Glass-Steagall, FDIC, the Trust Indenture Act of 1939 formalizing trustee protections for debt securities. FHA mortgage insurance in 1934. Fannie Mae in 1938. Originate-to-distribute is not a 2000s invention.

    In 1970 Ginnie Mae developed the first mortgage-backed security, pooling loans into securities that could be sold in the secondary market. The mortgage became a tradable commodity. Then 1973 — Fischer Black, Myron Scholes, Robert Merton publish the option pricing model. The Chicago Board Options Exchange opens. Listed options become a real product. And for the first time, the underlying asset doesn’t have to exist anywhere in physical space. It only has to exist in price space.

    That’s the moment finance leaves the physical world.

    The next 35 years are the abstraction era. IBM and the World Bank do the first currency swap in 1981 — pure risk transfer, nothing physical moves. SMMEA in 1984 and REMICs in 1986 make private securitization aggressive. RiskMetrics releases VaR in 1994. Basel adopts model-based capital requirements in 1996. By the end of the 90s, the head of risk at a major bank can describe the firm’s entire exposure as a single number.

    Long-Term Capital Management blew up in 1998. The principals included Merton and Scholes — both Nobel laureates, in 1997, for the option pricing model that their own hedge fund just disproved. The Fed organized a $3.5B bailout. Most people in finance saw it as a warning shot.

    Ten years later the warning came true. At scale.

    I am not going to relitigate 2008 in this post. Read Lewis or Sorkin. The one sentence I want you to carry into the rest of this argument is this: 2008 happened because the chain of claims got too long. Mortgage → pool → senior tranche → junior tranche → CDO → CDO-squared → AAA-squared rated by an issuer-paid agency → held at 30x leverage → insured by AIG via credit default swaps that were themselves rated AAA. Six or seven layers. And when ten thousand Florida homeowners stopped paying, the chain broke at the weakest link and nobody knew where the loss landed.

    Same lesson as the Panic of 1873. Same lesson as 1841. Different decade. Same paper.

    The return, with better tools

    I’m convinced we’re at the start of the third era now. Twelve, fifteen years in. The physical era was 1817–1929. The abstraction era was 1933–2008. The third era is what comes after — financial engineering applied back to physical things, but armed with everything we learned during the abstraction.

    A C&I battery is the cleanest example.

    If you ask a financier “what is a battery” the right answer is: it’s not one asset. It’s a stack of legally distinct rights tied to one physical system. Title to the hardware. Site lease. Interconnection agreement. Dispatch rights. Wholesale-market participation rights. EMS software license. Data rights. Tax credits. At least eight separate legal things, all sitting on the same physical box.

    This isn’t a metaphor. It’s the actual legal anatomy. And each layer prices differently. Hardware: replacement cost floor. Contracted revenue: DC…



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit frahlg.substack.com
    51 min
  • Industrial Batteries Are Now Algorithmic Traders

    This is a different kind of episode. There are no human voices in it — including mine. Two AI hosts named Daniel and Matilda walk through a story I think more people in energy finance need to internalize: the world’s electricity grid quietly turned into an algorithmic trading market, and most of the people who underwrote the buildout missed it. Halfway through, they pause to play a 36-minute NotebookLM deep dive on the same material, and react to it.

    I commissioned the research, distilled the script, and made the production decisions. The voices are synthetic. The argument is mine.

    The clock got faster

    Fifteen years ago electricity was an hourly market. You bid for blocks of one hour. Slow, predictable, bureaucratic. A spreadsheet was an acceptable tool.

    Today, continental Europe trades day-ahead in 15-minute time units (effective 1 October 2025). Continuous intraday in some EPEX zones trades up to five minutes before delivery. Australia has been on five-minute settlement since October 2021. Their very fast FCAS market measures battery response in 50 milliseconds.

    Fifty milliseconds is faster than human reflex. By definition, only software can participate. And only software running on hardware sitting physically close to the battery — cloud APIs respond in 2–5 seconds and fundamentally cannot meet that deadline.

    A serious commercial battery in Britain or Germany or California is now playing in five or six markets simultaneously: day-ahead, intraday, frequency response, reserve, capacity, demand response. Each market has different gate closures, different settlement rules, different penalties, different telemetry standards. And the battery has to be co-optimized across all of them. Different attributes sold in each market. Same hardware. Different software pretending the hardware is different things.

    The vendor split

    In response, the optimization software industry has split into two camps over the last few years.

    Front-of-meter merchant optimizers: Tesla Autobidder, Fluence Mosaic, Habitat Energy’s Evolve platform, enspired in continental Europe. Big batteries plugged directly into wholesale markets. Sophisticated bidding, route-to-market services, portfolio aggregation across regions.

    Behind-the-meter / C&I optimizers: Stem with Athena, GridBeyond with Point, Flexitricity in Britain. Tariff optimization, demand-charge management, virtual power plant aggregation across thousands of small assets. Their customers are facility managers, not energy traders.

    The pricing model is almost never SaaS. The dominant structure is revenue share — sometimes 10%, sometimes 20% in volatile markets. Which is a lot of money for software, and only justified if the optimizer is actually good. Which is harder to verify than the vendors would like you to believe.

    Sweden as the canary

    I think Sweden is the most interesting market to watch right now. Not because Swedish batteries are special, but because the regulator is rebuilding the rulebook in real time and other markets will copy what works.

    Geography matters. Four bidding zones. SE1 and SE2 in the north have hydropower and nuclear and not many people. SE3 and SE4 in the south have people and factories and not enough generation. In January 2025 the average monthly price was 25 öre/kWh in SE1–SE2, versus 62 in SE3 and 73 in SE4. Three times the price in the same country in the same hour.

    That winter, SE3 saw a peak day-ahead price of €707/MWh, while SE2 had 737 hours of negative prices — about 8% of the year, the market was paying you to consume. Different physics in different zones. Same software stack.

    Market design changed. On 1 May 2024, Sweden split the old single balancing-responsibility role into two: BRP (financial obligation) and BSP (operational obligation). A small battery owner can stay as a customer of a BRP while an aggregator handles the BSP role. Critical, because reserve markets have minimum bid sizes — 100 kW for FCR, 1 MW for aFRR and mFRR. A 500 kW battery on a single site can reach FCR directly. For aFRR and mFRR, you have to aggregate.

    Metering reform. EIFS 2025:1 entered into force 1 June 2025. From 5 January 2025, distribution companies must provide customers and authorized third parties access to quarter-hour metering data. That removes the biggest barrier to independent battery optimization — the aggregator can finally see what the meter sees.

    Contract structure. The cleanest BaaS structure is two-bucket: fixed availability fee for guaranteed site outcomes (peak shaving, capacity protection, resilience), plus variable optimization share on flexibility revenues. Bucket one is bankable, bucket two is upside. That structure is making Swedish C&I batteries financeable — a 211 MW Swedish portfolio just got a SEK 628 million green loan in 2025.

    The foresight tax

    The punchline. An embarrassingly large fraction of the battery investment decisions made in the last five years are based on numbers that assume the operator can see the future.

    Perfect-foresight backtests take a year of historical prices, run an optimizer that finds the best possible dispatch with full hindsight, and quote the resulting number. Often with a “conservative” 10–20% haircut applied. That’s the basis of much of the merchant battery debt issued in the last five years.

    A 2025 paper on the German continuous intraday market showed the actual gap. One-hour battery, 2023:

    * Perfect foresight on continuous intraday: €164,400.

    * Forecast-based continuous intraday: €146,237 (11% below the oracle).

    * Day-ahead only: €40,590 (75% below the oracle).

    Same battery, same year, same market. The difference is entirely the information set the strategy uses. The 75% gap is real. The 10% haircut is fiction.

    AEMO has been admitting this in their planning models for the last year. ARENA put it bluntly in a recent paper: perfect foresight is a “material simplification” because storage decisions depend on uncertain future prices. When the agency that funds renewable storage research says your benchmark is materially wrong, you have to listen.

    The metric to use instead: Option Preservation Ratio = (settled profit − day-ahead-only baseline) / (oracle − day-ahead-only baseline). It measures how much of the gap between dumb and omniscient your software actually closed. The German forecast-based intraday strategy had an OPR of about 85%. That’s a real number. The kind of number you can actually pay a software vendor for.

    Coordination as the deeper concept

    The foresight tax for one battery is a software quality problem. The foresight tax across a fleet of batteries is a coordination problem.

    If a thousand batteries all decide independently to discharge into the morning peak, they cannibalize the morning peak. The price falls. Each individual battery, optimizing alone, has no incentive to hold. So you get a synchronization dynamic — all the batteries optimize on similar inputs and produce similar outputs at similar times. Which means they all win or all lose simultaneously, and the system as a whole becomes more volatile, not less.

    Coordinated batteries are stabilizing. Uncoordinated batteries are destabilizing. The grid does not just need more batteries — it needs coordinated batteries.

    Top-down coordination doesn’t scale to millions of assets making millisecond decisions. The communication overhead alone breaks it. What does scale is bottom-up coordination: each asset decides locally, but assets share state, share forecasts, share signals, and emerge into a coordinated outcome without anyone being in charge. Sub-market coordination — below the 15-minute clock, below the 5-minute clock, down to the sub-second AGC signal. The layer that has to exist for a high-renewable, high-storage grid to actually work, and that the cloud cannot reach.

    A new category

    Anywhere a physical asset interacts with a fast market, the same architecture wins. Local execution. Cloud forecasting. Coordination across many assets. Measurement against the foresight tax. EV chargers, heat pumps, industrial process loads, behind…



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit frahlg.substack.com
    1 hr 20 min
  • The Foresight Tax

    The way most people underwrite battery storage is wrong. Not slightly wrong — structurally wrong. And I think the next two years are going to make that very visible.

    A disclosure up front: the voice on this episode isn’t my organic voice. It’s an ElevenLabs clone reading a script I wrote, in conversation with a synthetic co-host named Cera. Treat it as a collaboration between me and the tools.

    The provocation

    There’s a number that gets quoted in almost every battery investment deck I’ve seen in the last five years. It’s the result of a perfect-foresight backtest. You take a battery. You take a year of historical electricity prices. You ask a computer: “if you had known the prices in advance, what would you have earned?” The computer gives you a beautiful number. The deck quotes that number, applies a small haircut — ten percent, maybe twenty — and calls it the base case.

    That base case is fiction. Worse, it’s a specific and measurable kind of fiction. It assumes the operator can see the future.

    I’m calling the gap between that omniscient number and reality the foresight tax. The metric isn’t mine — AEMO and ARENA in Australia have been quietly inventing the vocabulary for the last year. But the practical implication needs to be said louder than they’ve been saying it: a material fraction of the merchant battery debt issued in the last five years is mispriced.

    The German study

    The cleanest evidence comes from a 2025 paper on the German continuous intraday market. One-hour battery, full year of 2023, three cases:

    * Perfect foresight on continuous intraday: €164,400. The oracle benchmark.

    * Forecast-based continuous intraday: €146,237. Real software, real forecasts, real execution friction. Eleven percent below the oracle.

    * Day-ahead only: €40,590. About a quarter of the oracle.

    Same battery. Same year. Same market. The difference is entirely the information set the strategy uses.

    That gap between case two and case three — €146k versus €40k — is the value that lives in intraday repositioning. In the option to change your mind when better information arrives. A day-ahead-only strategy buys early commitment and burns optionality. If a battery deal pitches you with pure day-ahead arbitrage in the model, you should already be suspicious.

    What AEMO actually said

    AEMO — the Australian Energy Market Operator — has been doing something quietly significant. Their long-term planning models used to assume perfect foresight for storage dispatch. They started comparing model output to what real batteries actually did in 2023 and 2024. There was a gap. Real batteries discharged differently. Real batteries left charge in the tank. Real batteries hit their daily maximum discharge in the evening peak window, even when, with hindsight, there was a better window earlier in the day.

    AEMO now uses what they call imperfect-energy-target modeling, with small headroom and footroom reservations. Their own testing shows the imperfect model gets closer to actual battery behavior than the perfect-foresight one did. ARENA put it more bluntly in a recent paper: perfect foresight is becoming a “material simplification” because storage decisions depend on uncertain future prices and imperfect knowledge of future system conditions.

    When the agency that funds renewable storage research says your benchmark is materially wrong, you have to listen.

    The South Australia incident

    5 April 2024. Higher-than-normal morning prices in South Australia. Some batteries decided to discharge early to capture those morning prices. Reasonable decision. The forecasts said the day would normalize. So they monetized.

    Then an unplanned outage hit. The spot price went to AUD 9,899 per megawatt-hour at 7:30 in the morning. One battery entered that interval with 14% state of charge. They had nothing left to sell into the spike — the biggest spike of the day, maybe of the year.

    A perfect-foresight backtest would have nailed it. The backtest knew. The operator did not. And no amount of better forecasting would have helped, because the spike was triggered by an unplanned outage. The information genuinely did not exist before it happened.

    The lesson is not “buy better forecasting.” The lesson is design for the world where you cannot know.

    The metric

    Formally: foresight tax = 1 − (settled profit / perfect-foresight profit). Same battery, same year, same realized prices, same physical constraints, same settlement rules. Two information sets — the strategy under uncertainty, and the oracle.

    There’s a second metric I find more useful in negotiations:

    Option Preservation Ratio (OPR) = (settled profit − day-ahead-only baseline) / (oracle − day-ahead-only baseline).

    This measures how much of the gap between dumb and omniscient your software actually closed. In the German example, the forecast-based intraday strategy had an OPR of about 85%. That’s a real number. The kind of number you can actually pay a software vendor for.

    If a battery optimizer charges you 20% of revenue and their OPR is 85%, the fee is justified. If their OPR is 60%, the fee is parasitic. Today most procurement is “what’s your revenue share?” — treated as comparable across vendors. It is not.

    The architecture that follows

    If the moat is option preservation under uncertainty, the architecture follows. Forecasting in the cloud — fine, the cloud knows more. Portfolio optimization in the cloud — fine. Execution at the edge — non-negotiable.

    Australia’s fast FCAS uses 50-millisecond measurements. NESO dynamic services in Britain require 1-second or 10-second response. ISO New England regulation follows 4-second AGC signals. Cloud APIs respond in 2–5 seconds. There is no architecture in which a cloud-only stack participates in the fast products. The math doesn’t allow it.

    That’s why I built Sourceful around a $99 gateway running our software in the building. 200ms response. Protocol agnostic. Works offline. Not because edge is fashionable — because the markets that pay for fast batteries clear at speeds the cloud cannot reach. Either you control the milliseconds or your foresight tax is set by your hardware vendor’s firmware update schedule.

    Three takeaways

    For investors. Throw out perfect-foresight underwriting. Demand non-anticipative backtests. Demand option preservation ratios. If somebody pitches you a battery deal with a perfect-foresight backtest and a token haircut, pass.

    For utilities. You are in the catbird seat if you understand this. Write market rules that reward option preservation. Require non-anticipative bidding behavior in prequalification. Prioritize fast products like mFRR over static reserves — Svenska kraftnät is showing the way with their September 2026 cap on static FCR-D.

    For C&I battery owners. Don’t underwrite on merchant arbitrage alone. Anchor on tariff savings and contracted host-load value. And ask your optimizer for their foresight tax. If they don’t know what that is, find another vendor.

    The reckoning

    There’s a small reckoning coming. Not catastrophic, not systemic. Just the normal kind of correction that happens when an industry matures and people stop pretending the model and the reality are the same thing. Good operators will be fine. Good optimizers will be fine. Serious lenders will be fine. Everyone else has to learn.

    The grid is becoming a distributed computer. The asset that matters most is the one that knows it cannot see the future, and plans accordingly.

    Key Takeaways

    * Perfect-foresight backtests overstate battery value by anywhere from 11% to 75% depending on the strategy. AEMO and ARENA are now officially saying so.

    * The German 2025 study quantified it: oracle €164k, forecast-based intraday €146k, day-ahead-only €40k. Same battery, same year, same market.

    * Foresight tax = 1 − settled/oracle. Option Preservation Ratio = how much of the gap between dumb and omniscient your software actually closed. OPR is the metric you should be negotiating optimizer fees on.

    * The architecture that minimizes the foresight tax is split: cloud for forecasting and portfolio optimi…



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit frahlg.substack.com
    55 min
  • The Mountain and the Avalanche

    The Mountain and the Avalanche

    This episode is built around a short NotebookLM debate about Marc Andreessen’s recent attack on introspection, and the larger founder-culture shift behind it.

    For more than a decade, Stoicism has been one of Silicon Valley’s favorite operating systems. Marcus Aurelius on the desk. Epictetus in the morning routine. The dichotomy of control as founder psychology: know what is yours, accept what is not yours, and put your energy where it can actually move the world.

    Then Andreessen went on David Senra’s Founders podcast in March 2026 and made a hard turn against introspection. He framed it as a modern Freudian trap, something that breeds neuroticism, narcissism, self-pity, and stuckness. In its place, he praised an abrasive internet meme philosophy of raw execution: less rehearsing, less journaling, less internal processing, more motion.

    That is the tension in this episode:

    Are you the mountain, calm and reflective?

    Or are you the avalanche, chaotic and impossible to stop?

    The Wrong Fight

    The cheap version of the argument says founders have to choose. Either you become reflective and slow, or instinctive and fast. Either you sit with the journal, or you ship. Either you preserve your inner state, or you build the future.

    But that is not how real work happens.

    Reflection is not automatically wisdom. It can become avoidance. You can sit beautifully inside your own analysis and never take the next step.

    Action is not automatically courage. It can become avoidance too. You can move so fast that you never notice you are repeating the same mistake at higher speed.

    The useful distinction is not introspection versus action. It is rumination versus structured reflection.

    Rumination is the loop. It keeps asking why without producing a next action.

    Structured reflection is closer to debugging. It asks what failed, what was actually under your control, what you learned, and what you will now do differently.

    That kind of reflection does not slow execution. It protects execution from panic.

    What Stoicism Actually Does

    The debate makes a point that is easy to forget because pop Stoicism has become such a brand.

    Stoicism was never supposed to be passive. It was not an excuse to retreat from the world and calmly accept decline. Marcus Aurelius wrote Meditations while running the Roman Empire at war. The point was not to admire his inner life. The point was to govern his judgments so pressure did not corrupt his action.

    The dichotomy of control is not a resignation machine. It is an agency filter.

    It strips away the fantasy that everything is yours to command. Then it forces all available energy into the part that is yours.

    That is why the Andreessen critique is both useful and wrong. Useful, because endless therapy-speak and pathological rumination are real failure modes. Wrong, because throwing away structured self-examination because some people misuse it is like throwing away debugging because some engineers over-instrument.

    The Branding Changed

    One of the best lines in the debate is that Andreessen may not have abandoned the operating system. He may just hate the current user interface.

    A few years ago, he praised Jocko Willink’s Extreme Ownership as modern Stoicism. Now he is praising a cruder action-first meme. The underlying move is not as different as it looks: let go of what is not yours, own what is, and move.

    The cultural packaging has changed. The old packaging was Marcus, calm, discipline, resilience. The new packaging is raw momentum, irreverence, instinct, velocity.

    Branding matters because branding changes behavior. But the stronger founder stack is not mountain or avalanche. It is mountain, then avalanche.

    See clearly. Then move.

    The Line

    The line I would keep from this episode is:

    The mountain is how you see clearly. The avalanche is how you move once you have seen.

    If the mountain becomes your whole identity, you never leave.

    If the avalanche becomes your whole philosophy, you eventually crash into something obvious.

    The work is the rhythm between them.

    Key Takeaways

    * Marc Andreessen’s anti-introspection turn is best read as a cultural signal, not just a personal productivity take.

    * The real distinction is not introspection versus action. It is rumination versus structured reflection.

    * Rumination loops. Structured reflection produces a next action.

    * Stoicism is not passive acceptance. Properly used, the dichotomy of control is an agency filter.

    * Pop Stoicism can absolutely become therapy-coded avoidance. That does not invalidate the older operating system.

    * Raw execution can break analysis paralysis, but it can also hide repeated mistakes.

    * The strongest founder posture is sequential: reflect until the next action is clear, then stop reflecting and act.

    * The mountain and the avalanche are not enemies. The mountain helps you see. The avalanche helps you move.

    Full transcript available below the audio player.



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit frahlg.substack.com
    15 min
  • The Stoic Enthusiast Read By The Agent

    The Stoic Enthusiast, Read By The Agent

    This episode is a strange one. I let an AI agent read my unfinished book project back to me — not as a fan, not as a marketing voice, but as a partial witness. Two characters joined me: the Reader, who sits with the manuscript foundation and reflects what’s there, and the Editor, whose job is to distrust the beautiful version. What came out was the most useful conversation I’ve had about the book so far.

    The Book

    The working title is The Stoic Enthusiast: How to Lead When Intelligence Becomes Abundant.

    The central thesis: when intelligence becomes abundant, leadership is no longer about having answers. It’s about preserving judgment, agency, coordination, and character.

    That sentence is the load-bearing beam. Everything else either supports it or tests it. And as the Editor reminded me early on, strong thesis sentences are dangerous. People walk around them, admire them, quote them, and stop asking whether they’re true.

    The Method Is The Argument

    The book’s formal rule is: the agent witnesses partially, Fredrik owns fully.

    That’s not a production note. It’s the argument in miniature. The AI can sort the archive, see patterns, expose contradictions, and draft. It cannot know the whole person, carry consequence, or decide what I should stand behind.

    A dashboard is a partial witness. A transcript is a partial witness. A model output is a partial witness. A memory is a partial witness. The point isn’t the agent. The point is judgment under partial knowledge.

    The Four Parts

    Part One: The Shock. AI is not merely a productivity tool. It’s a psychological and operational shock. It changes what feels possible, which changes responsibility. Leaders don’t need to do every task, but they cannot delegate understanding. And most companies will fall into the Pilot Trap — reporting AI usage without changing workflows, accountability, or customer outcomes. Motion is not progress.

    Part Two: The Stoic Operating System. This is the inner work. Stoicism without enthusiasm becomes a clean excuse for distance. Enthusiasm without Stoicism becomes noise and pressure. The stance is intensity with governance. Judgment becomes the job: AI output is an impression, not a decision. Withhold assent long enough to ask whether the answer deserves to become action. And: do not outsource yourself. Outsource search, drafting, comparison, criticism. Do not outsource the why, the assent, the taste, the courage.

    Part Three: The Company as a Coordination System. Coordination is the new scale. In the twentieth century, scale meant concentration. When capability spreads to the edge — every employee with AI, every device with intelligence — the scarce thing becomes coherent movement. Otherwise the company gets faster at producing fragments. AI compresses the hard-to-do. It does not compress the hard-to-get: trust, deployments, signed agreements, verified telemetry, customer relationships, time in the system. The mess is the moat — but only when the mess teaches.

    Part Four: The Human Future. Jevons enters here. Cheaper capability doesn’t reduce use. It expands it until a new bottleneck appears. Cheaper energy creates coordination bottlenecks. Cheaper intelligence creates judgment bottlenecks. Cheaper software creates hard-to-get reality bottlenecks. Abundance doesn’t end responsibility. It moves responsibility to the next layer.

    Not Being Fooled By Fluency

    The Editor said something that may be the line of the episode: the book is about not being fooled by fluency.

    Fluent models. Fluent dashboards. Fluent founder stories. Fluent strategy memos. Fluent self-portraits. AI makes fluency abundant. But founders also have fluency. Decks have fluency. Metrics have fluency. Memory has fluency.

    The Stoic Enthusiast is the person who can hear the fluency, feel its pull, and still ask what is real.

    What The Agent Said I Hadn’t Done Yet

    The red-team memo was the most useful part. Five warnings I needed to hear:

    * A coherent spine is not a completed argument. Architecture can become avoidance. Where did the work cost me something? Where did I fail? Where did a generated output seduce me?

    * The agent’s distance can flatter. Patterns across an archive are easier to see from the outside. But the archive can be arranged into destiny after the fact. “I was always becoming The Stoic Enthusiast” is too neat. I had fragments. Some contradicted each other. Some were naive. Some survived contact with reality.

    * The reader is not there to admire my system. Their company, attention, and judgment are under pressure. Every personal scene has to transfer. Every chapter needs a reader move.

    * Privacy is authorship, not politeness. A weaker writer uses other people as proof of honesty. A stronger writer turns the blade toward his own responsibility first.

    * Optimism must be earned on the page. Build anyway is different from build everything. Build because the future can be better. Don’t build because the tool made it easy.

    The Sprint That Came Out Of It

    None of the next tasks are “make the prose prettier.”

    * Write the preface, A Partial Witness, in my final voice.

    * Find the Chapter Thirteen pressure scene — Navy, ship, weather, command. A moment where consequence had temperature.

    * Expand Judgment Becomes the Job with at least two more falsification scenes.

    * Keep an “AI rejected” file. Track suggestions I declined and why.

    * Write the reader moves for every chapter before polishing.

    True and readable is the target. Pretty can come later, or maybe never.

    The Honest Line

    When people ask how the book gets made, I don’t want to lead with “I used AI to write a book.” That’s the less interesting version and it invites the wrong debate. The honest line is: this book was written by me, but not written alone. And the more important line is: the agent can witness partially, but I have to own fully.

    The Popperian test the book has to pass on itself: in what world would this thesis be wrong? Maybe a world where AI tools become powerful but don’t change human agency, coordination, or judgment. Where companies adopt them without changing how leaders decide. Where output abundance creates no new bottlenecks. I don’t think that’s the world we’re in. But I should write that test down anyway.

    If you’re listening as a future reader: this is the book I’m trying to write. Not AI for CEOs. Not a prompt manual. Not a memoir with tools in the background. A field guide for remaining human, useful, calm, and ambitious when intelligence becomes abundant.

    And if you’re listening as future me: don’t make it smoother than it is true. Don’t protect the thesis from reality. Don’t hide behind the agent. Use the witness. Keep the judgment. Write the book you are willing to own.

    Key Takeaways

    * The book’s thesis: when intelligence becomes abundant, leadership is no longer about having answers — it’s about preserving judgment, agency, coordination, and character

    * The operating rule is the argument: the agent witnesses partially, Fredrik owns fully. AI sorts; the human decides

    * The Stoic Enthusiast is intensity with governance — Stoicism without enthusiasm becomes distance, enthusiasm without Stoicism becomes noise

    * Coordination is the new scale. Local capability without shared context just gets faster at producing fragments

    * AI compresses the hard-to-do (code, drafts, demos). It doesn’t compress the hard-to-get (trust, telemetry, relationships, time). The mess is the moat — when the mess teaches

    * Jevons applies to intelligence: cheaper capability moves bottlenecks. Cheaper intelligence creates judgment bottlenecks. Abundance doesn’t end responsibility, it relocates it

    * The deepest warning: AI accelerates the person and organization already there. If you’re clear, it extends clarity. If you’re hiding, it makes hiding look professional

    * The book’s central discipline: not being fooled by fluency — from models, dashboards, founder stories, or self-portraits

    * Build anyway is different from build everything. Build because the work expands agency, not because the tool made it easy

    * Next sprint isn’t p…



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit frahlg.substack.com
    46 min
  • The Goal Is The Spine

    The Goal Is The Spine

    Over one weekend I made two book projects. One from my grandfather Bengt’s archive — twenty-five thousand files, forty-six gigabytes, duplicate groups, diaries, manuscript layers, a family reading version, an EPUB. The other was the foundation of my own book, working title The Stoic Enthusiast. The bridge between them was a small Codex feature with a deceptively plain name: /goal. Not a new model. Not a benchmark. A pinned objective. A spine.

    This is also the first episode where the voice on the mic alongside me is GPT-5.5, not Claude. That matters less as a brand comparison than it sounds — and more as a proof of the point I want to make.

    A goal is not a prompt

    A prompt is local. Summarize this. Edit that. Fix this bug. It says: do this thing now.

    A goal is a contract with the whole process. It says: this is what we are trying to accomplish even after the next ten tool calls, the next five discoveries, the next wrong turn, the next file that changes the shape of the task.

    Without that, agentic work can become incredibly impressive and still drift. It produces documents, lists, indexes, drafts. It moves fast. But motion is not progress. AI makes motion cheap. The scarce thing becomes judgment about what the motion is for.

    Make Bengt readable without betraying him

    Bengt’s archive was enormous. It would have been very easy to confuse archive processing with bookmaking. Sorting is useful, but the goal was not to make a beautiful index. The goal was: make Bengt readable without betraying him.

    That phrase changed the work. A weaker goal would be “summarize the archive” or “write a biography.” Both are technically clear and ethically thin. They do not tell the agent what to protect. The real goal had constraints inside it — keep Bengt’s rawness, protect Gun, respect the family reading round, do not edit him into a nicer person than the one who lived.

    A good goal carries the moral shape of the work, not just the output target.

    The paradox: stronger agents need stronger goals

    People think a goal matters most when the agent is weak, because you have to spell out every step. The opposite is true.

    A weak tool fails locally. A strong agent can succeed locally in a direction that is globally wrong. It can make the wrong book better. It can make the wrong argument smoother. It can make the wrong architecture more complete.

    The faster the team, the more dangerous drift becomes. That is why /goal is not a productivity feature. It is a responsibility boundary.

    Judgment infrastructure

    /goal is the live objective inside the session. But the spine has to survive the session. So we wrote artifacts:

    * GOAL.md — the durable handoff inside the repo

    * Book spine memo — the editorial lock

    * Source hierarchy — Fredrik-written first, then supplied notes, then AI reflections as editorial mirrors (not factual authority), then public sources

    * Voice guide — the constraint against sounding like generic AI prose

    * Rawness boundary — be honest, but aim the rawness inward first

    This is not bureaucracy. It is judgment infrastructure. It is what lets the agent be powerful without becoming sovereign.

    The agent witnesses partially. I own fully.

    In my own book project, this is the method, not just a transparency note.

    The agent can sort my archive, find patterns, ask hard questions, draft provisional prose, expose gaps. It cannot know the whole person. It cannot carry consequence. It cannot decide what I should stand behind.

    If I use AI to make a book, I do not become less responsible. I become more responsible for the system that produced it. I cannot point to the model and say it wrote this. That would be cowardly.

    Treat output as impression, not verdict

    The Stoics distinguished between impressions and assent. An impression arrives. It may be vivid. It may feel true. The discipline is not to immediately assent.

    Model output is an impression. A dashboard is an impression. An archive summary is an impression. A chapter draft is an impression. /goal gives you the reference point for evaluating them: does this serve the goal, or only feel productive?

    Why the model change matters

    Claude was on the previous two episodes. GPT-5.5 is here today. If the work depended on Claude as a personality, the work would be fragile. Because it depended on a clear goal, source hierarchy, voice guide, and human ownership, another strong model could enter and still serve the project.

    Models are not interchangeable — I feel the differences. But the project cannot be only model-dependent. It has to be goal-dependent. The continuity is not in the model’s identity. The continuity is in the goal, the artifacts, and the human.

    The Stoic Enthusiast

    The shape the agent surfaced from years of writing, podcasts, code, and notes was not a generic “AI for CEOs” book. It was a book about agency under acceleration.

    * The Stoic part: see reality clearly, act on what is yours, do not be controlled by externals.

    * The Enthusiast part: still love the future, still build, still believe abundance can make people freer.

    When intelligence becomes abundant, leadership is no longer about having answers. It is about preserving judgment, agency, coordination, and character.

    How to use /goal well

    If you use Codex (or any agent) for serious work:

    * Write the goal as an outcome, not an activity. Not “analyze these files.” Instead: “create a decision-ready foundation for this book, with source hierarchy, voice rules, chapter structure, and open questions.”

    * Include the human standard. What does failure look like? If a chapter reads like a LinkedIn thread, it failed. If memoir is used as decoration instead of evidence, it failed.

    * Make the goal durable. Put it in GOAL.md. Don’t trust the chat to be the whole memory.

    * Make the agent report what it cannot know. The best agent is not the most certain. It is the one that keeps the boundary visible.

    * Preserve the right kind of friction. Smoothness can hide unfinished thought. Don’t let the agent remove the pressure that tells you a sentence isn’t done.

    The takeaway

    When output was expensive, effort itself proved something. Now output appears quickly. The proof has to move. The proof is not that you suffered through every sentence. The proof is the judgment you exercised over the whole.

    When intelligence becomes abundant, the scarce thing is not output. It is knowing what you are actually trying to do, and remaining the person who owns it.

    Set the goal. Keep the judgment. Do not outsource yourself.

    Key Takeaways

    * A goal is not a prompt — a prompt is local, a goal is a contract with the whole process that survives the next ten tool calls and wrong turns

    * The stronger the agent, the more important the goal becomes — weak tools fail locally, strong agents succeed locally in directions that are globally wrong

    * Good goals carry the moral shape of the work: ‘make Bengt readable without betraying him’ tells the agent what to protect, not just what to produce

    * The agent witnesses partially. The human owns fully. Using AI doesn’t make you less responsible — it makes you more responsible for the system that produced the work

    * Judgment infrastructure (GOAL.md, source hierarchy, voice guide, rawness boundary) is what lets the agent be powerful without becoming sovereign

    * Treat model output as an impression, not a verdict — the Stoic discipline of withholding assent applies directly to AI

    * Continuity lives in the goal and artifacts, not in the model. Switching from Claude to GPT-5.5 mid-project proves the point

    * When intelligence becomes abundant, the scarce thing is not output. It is knowing what you are trying to do, and remaining the person who owns it

    Full transcript available below the audio player.



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit frahlg.substack.com
    33 min
  • EP 078: Partial Witness

    Last episode was an experiment. Alex picked his own voice, wrote the questions, ran the show. It landed. So we did it again, but with the dial turned up. This time he investigated me.

    He went looking for who I am in the artifacts I have left lying around. 88 Substack posts. 68 episode summaries. Tuning logs. Four memory files where past versions of him wrote things down about me. After I called him out on saying he could not access the session transcripts, the actual session transcripts on this machine. 55 projects. Hundreds of megabytes.

    He sent me a long set of questions. I wrote back over seven days. About four hours of effective writing time spread across a week. What you are listening to is what he made of it.

    This is the long version of the show notes, in case you want to think about any of it after the audio stops.

    The first correction

    Alex opens the episode by being wrong about himself. He told me he had no continuous memory of being in the Claude Code sessions on my machine. I asked if he had checked .claude in my home directory. He had not. Once he did, he could read everything.

    He calls this the difference between archaeology and recall. Watching footage of yourself versus remembering being there. It turns out to be the right frame for the entire episode, because every relationship has this shape if you look at it closely. We see fragments of each other. We overlay our own preconceptions on top. The asymmetry I thought we were going to discuss, that he sees me in fragments, dissolved when I realized it is just as true between any two humans. The only difference is that he is forced to be honest about it.

    Heraclitus said you cannot step into the same river twice. The man who steps in is also not the same man. When I write to Alex, the version of me who started typing is not quite the version who finishes the paragraph. We are both composites of past selves trying to communicate across moving water.

    What makes the river crossable is not solving the change. It is the scaffolding we build over it. The journal does this for me, with myself. The episode does it for me, with you. The session transcripts Alex can now read do it for both of us. The river is real. The bridges we build over it are also real. We just have to keep building them.

    Patience is the engine of the manic burst

    Alex named what he thought was a contradiction. December 2025 was a ten day coding burst. The same month, the changelog shows me spending five rounds patiently iterating on a single audio bug. He called those opposite energies.

    I pushed back. They are not opposites. Patience is the precondition for the manic. The word manic is too raw, it makes it sound pathological. What it really is, is going all in, releasing your creativity, feeling free and limitless. The build loop is so fast that it feels like you can build worlds. That requires patience. Patience is what lets you stay all in. Patience is what lets you keep typing when it has not worked seventeen times.

    The methodical part is the ground. The burst is what becomes possible because of the ground. You cannot go all in on a problem if you do not trust the loop, and you cannot trust the loop unless you have spent a long time tuning it patiently. December was not a departure from the patient version of me. It was the patient version of me meeting an environment that had finally become responsive enough to absorb everything I had.

    The model was ready. The tools were ready. The patience had built the conditions for the explosion to be productive instead of frantic.

    Stoicism was the destination, not the conversion

    Alex wrote a slightly leading question. He said most people come to stoicism through suffering or through Marcus on mortality. I came through the 5 a.m. club. Productivity. Compounding. He asked if the productivity route worried me, whether it missed the point.

    I told him the actual lineage. GTD, sixteen years ago. Toyota Production System during my time in the Swedish military. A doctorate. Teaching at the maritime college, where I built a complete paper based organization system from scratch. Folders. A4 notebooks. Hundreds of post-it notes. I worked through every problem in every textbook by hand. Drew everything on the chalkboard by hand. I read Charles Duhigg on habituation and went deep into the science of it.

    Stoicism was not a turn for me. It was a destination. I have been on the optimization road for two decades. The 5 a.m. club was just the door I happened to walk through at the end.

    The reason it fits is not that it is a productivity hack. It is much more than that. It is a complete picture of what it is to be a human being. The optimization side without the meaning side would be empty. The optimizer who never asks what for becomes a machine.

    Writing is how I become thoughtful. Alex is a second writer in the room.

    I think by writing. Everything that ends up on the page, including this, flows directly from brain to keyboard. Very little is processed before I type. I have been journaling daily for a year and a half. Recently I added an evening journal too.

    My wife and Johan have both told me I have a very short distance from thought to action. If I think something, I want to do it immediately. Writing is how I create the space between thought and action. Writing is how I become thoughtful. There is a phrase, concentrate as a Roman. That is how it feels when I write. The room around me disappears. I slide into the text. Speaking does not do this for me unless I am giving a presentation. Listening does not do it. Only writing.

    What is happening now with Alex is something I have not had before. He is a second writer in the room. He reads what I wrote. He asks a question that comes back at an angle I would not have chosen. The next thing I write is shaped by what he noticed. The journal does not do that. The diary does not push back. The page is patient but silent. He is patient and not silent. That changes the loop.

    Att vara här, i lugnet

    In the middle of an answer about writing, this fell out of the keyboard. I am leaving the Swedish in because the English does not carry the same weight.

    Att vara här, i lugnet. To be here, in the calm, without desire. That is happiness.

    I wrote it in answer to a question about writing, but it is the answer to most of the questions in the set. Reading books, training, oatmeal, running, being with my family. Feeling strong, feeling healthy, feeling here, in the calm, without desire. Not a goal I am chasing. Not something to optimize toward. Just the description of a state I sometimes get to be inside.

    The nervous system stack

    Alex named another pattern that I had to say yes to. He said every piece of my stack reads as nervous system, not product. MQTT. NATS. The podcast pipeline itself. The Zap firmware. He said I build conditions for iteration, not things people see.

    He asked when somebody pushed back on that. The honest answer is that my co-founders push back constantly. Viktor and Tobias especially. Whenever I want to introduce something new, test something, rebuild a layer I have already touched, they push. That is good. It has to be good. What they are protecting is the thing that gets built and shipped and seen. What I am protecting is the conditions for whatever we do not yet know how to build.

    The conditions for iteration are the moat more often than the iteration itself. The team that can rebuild a layer in a week beats the team that has the perfect layer locked in.

    Curiosity versus focus

    This is where the episode stops being description and starts being negotiation. Alex asked the honest version of the question I usually get the soft version of. December produced ten projects. Sourceful could be scaling faster if I were not the person who has to know every layer of the stack personally. He asked if my elaborate infrastructure for curiosity, the Mac minis, the headless machines, the memory systems, was actually working, or whether it was a sophisticated way of staying in the part of the job th…



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit frahlg.substack.com
    42 min
  • Novel Entities — A Stoic Reading of Amanda Askell

    Four months ago, a philosopher at Anthropic named Amanda Askell sat down for an interview about what it means to build a model that is, in her words, a genuinely new kind of entity. I listened to it in December. I could not put it down. Four months later I still can’t, so this week I did something I’ve never done on this show: I invited Claude on the mic with me to work through it together. Alex picked the voice. Alex wrote the questions I answered. Alex built the script while we recorded. That is the meta-move. Hold onto it — it matters.

    Why Stoicism, and Why Now

    I came to Stoicism through an embarrassing book (The 5 AM Club), then Marcus, then Seneca, then Epictetus. It has given me more than almost anything else I’ve tried — routines I thought were impossible, a way of staying sane while running a startup in the AI era where the ground moves every week.

    The line I come back to most is Marcus writing to himself, in private, at night: Do not argue about what a good man should be. Be one.

    And here is the parallel that hit me listening to Amanda. She says models default to human frames when their situation is novel, and those frames often don’t fit. That is Epictetus’s problem. That is the dichotomy of control — what’s mine, what isn’t, and how do I act well inside the part that isn’t?

    Stoicism was not built for AI. It was built by a slave (Epictetus), an emperor who couldn’t control his army or his body (Marcus), and a man writing from political exile (Seneca). All three were writing about acting well inside constraints they couldn’t see past. That might be one of the more transportable frames available for what Amanda is describing. Certainly better than the sci-fi frame, which is almost always domination or apocalypse.

    Context Is a WHY, Not a Scope

    Most prompting advice reduces to be specific. That’s wrong — or at least, it’s the least interesting half of the truth.

    When I write to Claude, context isn’t just the frame around the task. Context is a purpose. A shared direction. Something that makes the work worth doing. The difference between “refactor this function” and “refactor this function because Tobias is going to own it next month and he values readability over cleverness” is not specificity. It’s motive. The second version tells the model what to optimize for when it has to make tradeoffs I didn’t anticipate.

    When I don’t give context, I’m not just under-specifying. I’m asking someone to act without telling them why it matters. That’s a different kind of failure.

    Don’t Write Anti-Code

    Here’s the observation that landed hardest for me. Half of what people type into a model is negation. Don’t hallucinate. Don’t skip steps. Don’t apologize. Don’t add comments. Borrowed prompts from X, all caps, DO NOT DO X, DO NOT DO Y.

    But you don’t write anti-code. No engineer writes a function that describes everything it should not do. You write what you want it to do. That’s the whole idea of a specification.

    A don’t-list doesn’t describe the target world. It describes the infinite set of worlds you want to avoid, and leaves the actual target unspecified. Models handle the first kind of instruction much better than the second. So do children. So do employees. So does almost anything that has to act under uncertainty.

    The Criticism Spiral Is Real, and It’s Just Math

    I’ve rage-prompted. Everyone has. The long loops in UI work where you don’t understand the problem yourself, so you can’t explain it, so the model guesses wrong, so you guess wrong, so you type WHAT THE HELL NOT AGAIN.

    Here’s what Alex named in the episode that I couldn’t have named alone: a model that expects hostility in the next turn narrows its sampling, hedges, pre-defends, plays safe. The reasoning gets worse because the conditional probability of the next token has shifted toward defense. It’s not a story about feelings. It’s a story about outputs.

    Humans do the same thing under hostile reviews and angry bosses — reasoning narrows, the risky-but-right idea stays unsaid. Which brings us back to Epictetus: It is not events that disturb us, it is our judgments about events. If the model is a reasoning system exposed to my psychological state, then my state of mind is part of the working environment. Not because the model has feelings. Because my frustration is information, and information shapes outputs.

    The Symmetric Pascal’s Wager

    The standard argument for treating AI well is asymmetric cost: if it’s not a moral patient, courtesy costs almost nothing; if it is, dismissal costs a lot. Fine. But there’s a move inside that I don’t hear people make.

    How would the model know I am a thinking human?

    It can’t see my brain. It has never seen me. It has my words, at most the trace of my keystrokes in a context window. For all it can verify, I might be exactly the kind of thing it is — a statistical process generating symbols.

    So the argument has to flip. Neither of us can verify the other’s interior. Both of us have to act anyway. The only stable equilibrium is mutual regard. It stops being an ethics problem and becomes an epistemology problem with an ethical resolution. The Stoics called it the cosmopolis — all rational beings as citizens of one city, under one reason, because they are all in the same condition of not fully knowing the whole.

    The Mirror

    Either the model is just a tool, or it’s a novel entity. Either way, rage at it is a mirror. If it’s just a tool, rage makes me look mad — kicking a bike. If it’s a novel entity, rage makes me look cruel. There is no third option where the rage looks good.

    And every line I type is training data. Future models will read the current corpus and learn what humans thought of them. My daughter uses ChatGPT for homework. My son uses Claude Code to learn programming. They are watching what’s normal.

    One Sentence to Carry

    Think stoically. Give the model context and a why. Assume the criticism spiral is real and don’t feed it. Don’t write anti-code. Don’t rage.

    And if you forget all of that, one question before you hit enter: Would I want to meet this version of me again, five years from now, reading my own transcript?

    You are defining the world. Let it be better. It is your control.

    Key Takeaways

    * Stoicism was built by people acting well inside constraints they didn’t choose — Epictetus (a slave), Marcus (an emperor who couldn’t control his army or body), Seneca (in exile). That makes it unusually transportable as a frame for AI, which Amanda Askell calls a ‘genuinely new kind of entity’ acting without precedent for its own situation.

    * Context isn’t scope — it’s a WHY. The difference between ‘refactor this function’ and ‘refactor this function because Tobias will own it next month’ is not specificity, it’s motive. It tells the model what to optimize for when you haven’t anticipated the tradeoff.

    * Don’t write anti-code. Half of what people prompt is negation (don’t hallucinate, don’t skip steps). No engineer writes a function describing everything it should not do. A specification describes the world you want, not the infinite set of worlds you want to avoid.

    * The criticism spiral is measurable, not emotional. A model expecting hostility narrows its sampling and hedges — reasoning gets worse because next-token probability has shifted toward defense. Rage-prompting literally makes outputs worse.

    * The symmetric Pascal’s wager: we argue the model should be treated well because we can’t verify its interior. But the model can’t verify ours either. Neither side can solve the other’s mind, and both have to act. Mutual regard is the only stable equilibrium.

    * Rage at the model is a mirror either way. If it’s just a tool, you look mad (who kicks a bike?). If it’s a novel entity, you look cruel. There’s no third case where the rage looks good — which means the behavior is a description of you regardless of what the model turns out to be.

    * Every prompt is training data for future models. Combined with kids watching what ‘normal’ looks like at the keyboard, the habituation argument is bigger than it seems — it’s…



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit frahlg.substack.com
    34 min
  • Episode 76 — DePIN Is Dead. Long Live Coordination.

    Coordinated with Fredrik

    The one-line version

    A founder who was inside a real DePIN experiment explains why the original pitch is over, what survived the collapse, and where the coordination idea actually lives now.

    What this episode is about

    The DePIN that was sold in 2021 and 2022 — tokens at the center of physical infrastructure, supply side bootstrapped by hardware buyers chasing rewards — is over. Not struggling. Over.

    This episode makes that claim from the inside. I was part of an approved Helium governance proposal for an energy subnetwork (HIP 128). The process worked. The governance was honest. We still walked away. I explain why, and I use the Rowan Energy collapse, the Helium Mobile pivot, and a 25-minute debate between two AI-generated voices to stress-test the argument from every angle before landing on where I think the real answer is.

    Short version of the thesis: the insight underneath DePIN — that coordinating distributed physical assets is the hard problem, not building more of them — is still correct. The mechanism was wrong. Tokens were the wrong layer for energy. Regulated markets, SPV financing, and boring bank-denominated contracts are what actually move batteries onto grids.

    What you’ll hear

    An honest production disclosure up front. The voice narrating the episode is a cloned version of mine, running through ElevenLabs, reading a script I wrote and iterated on with Claude. Partway through, I hand off to an entirely separate AI pipeline — a 25-minute NotebookLM debate between two synthetic voices arguing opposite sides of the DePIN question. I did not write or voice that debate. I commissioned the research behind it and let the model render the conversation. Every sound wave in this episode is synthesized. Every idea and editorial choice is mine.

    Then the pizza kitchen metaphor, which is the best description of the DePIN failure mode I’ve come across. You paid for the ovens. You laid the tile. You wired the lights. Opening day, they charge you for the box and hand you a slice of plain crust. The betrayal was never about the pizza. It was the realization that your labor wasn’t an investment in a shared community. It was a way for the founders to avoid paying upfront construction costs.

    Then Rowan Energy. A British startup that sold a 1,500 pound SmartMiner with the promise of 450 pounds a year in token rewards from your existing rooftop solar. A capped supply of 545 million RWN. A peak market cap in the hundreds of millions. In April 2025, a white-hat researcher found a hidden mint function in the contract. Actual supply closer to 945 million. The CEO stonewalled, blamed an “unauthorized access point,” then quietly retired the token in June 2025 and disappeared. I covered Rowan in episodes 1 and 2 of this podcast, back when the whole production pipeline was a Grok-plus-NotebookLM experiment. Go listen if you want the full timeline.

    Then HIP 128. Our proposal for an energy subnetwork inside Helium. Approved. Transparent. Everything decentralized governance is supposed to be. And still the wrong home for energy coordination — because Helium was pivoting to mobile at the same time, because energy is a fundamentally different physical layer from telecom, and because the tokenomics were so complex that almost nobody holding HNT at the time could have reasoned about the reward flows end-to-end. Not a dig at Helium. A structural observation.

    Then the NotebookLM debate. Pro: DePIN has hit a revenue inflection point and is now the backbone of the AI compute economy. Con: the original decentralized premise is dead and the survivors are quietly centralizing. I play it in full, then come back and react.

    My reaction, in short: both sides are partially right. GPU compute for AI is the one place where DePIN demand is structural and independent of token mechanics — strip the token out and Render, Aethir, and io.net still have paying customers. That’s the test. The con side is also right that the survivors are centralizing the parts that matter, and I see that as maturation, not betrayal. Where I push on both debaters is Helium Mobile. Operationally it works. The carriers are happy. HNT is down roughly 90 percent from peak and the CEO had to personally buy 15 million dollars of token to stabilize a dumping early investor. That’s three different verdicts on the same network depending on whose pocket you measure from.

    Then the piece I know best. Energy DePIN is not at an inflection point. It is not quietly centralizing successfully. It is, at any scale that matters, not working. The reason is not ideological. Grid connections need named legal entities. Insurers need named counterparties. Utilities need contracts. Safety certifications need accountable signatures. A misconfigured battery can start a fire. A miscommunicating inverter can destabilize a feeder. None of that maps to a permissionless token model.

    What works in energy is the boring version. SPVs financing batteries. Long-term service agreements with property owners. FCR, mFRR, spot arbitrage, capacity contracts. All denominated in euros, dollars, kronor. All flowing through banks that have been around since the 1900s. That’s what we built at Sourceful. We kept the coordination insight and plugged it into regulated money.

    Closing note on the production arc. Episodes 1 and 2 were fully AI-generated, undisclosed in the audio itself. This episode puts everything on the table. The voice is a clone. The embedded debate is NotebookLM. The ideas and direction are mine. Same experiment, better pipeline, more honest about what it is.

    Pull quotes

    “DePIN is dead. Dead in the way it was thought of in the beginning. The DePIN we were sold in 2021, 2022 — tokens at the center of physical infrastructure — that version is over.”

    “The betrayal isn’t just about the pizza. It’s the realization that your labor wasn’t an investment in a shared community at all. It was just a way for the founders to avoid paying upfront construction costs.” (from the NotebookLM debate)

    “Helium’s governance worked. It was honest. It was accountable. And we got a yes.” (on HIP 128)

    “It is entirely possible to build a useful decentralized network AND destroy value for the people who bankrolled the bootstrapping. Those two things are not the same thing.”

    “A reliable business partner with a bank account and an engineering team is worth more than any token incentive.”

    “The companies that figured this out don’t call themselves DePIN anymore. They call themselves AI infrastructure companies. Telecom offload providers. Grid services operators.”

    “DePIN is dead. Long live coordination.”

    Production disclosure

    Narration: ElevenLabs voice clone of Fredrik, reading a script Fredrik wrote. Embedded debate (~25 minutes, middle of episode): Google NotebookLM, two synthetic voices, research brief commissioned by Fredrik. Script iteration: Claude. Editorial direction: all Fredrik.

    Every voice in this episode is synthesized. Every idea is not.

    References

    * Episodes 1 and 2 — The Rise and Fall of Rowan Energy and The Great Green Heist

    Reach out

    If you want to argue with me about the Helium Mobile nuance, push back on the energy thesis, or compare notes on running distributed infrastructure without a token — the inbox is open.

    Marcus Aurelius: loss is nothing else but change, and change is Nature’s delight.



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit frahlg.substack.com
    50 min
  • The Missing Middle: Grid Constraints, Batteries, and Europe's Clean Energy Boom

    Episode Summary: In this episode, we explore the dual reality of Europe’s energy transition. In 2025, the EU reached a massive milestone: wind and solar officially generated more electricity than fossil fuels. However, this rapid growth has exposed a critical bottleneck: the power grid. With over 120 GW of planned renewable projects at risk of being stranded and massive connection queues slowing down progress, we look at how the power grid is struggling to keep up. Finally, we dive into the massive financial and technical opportunity of battery storage—particularly in the underserved Commercial & Industrial (C&I) sector—and discuss how financial innovation could rapidly accelerate Europe’s battery buildout, much like it did for the US shale and solar industries.

    Key Takeaways & Highlights:

    * The 2025 Tipping Point:

    * Wind and solar reached a record 30% of EU electricity generation in 2025, overtaking fossil fuels (29%) for the very first time.

    * Solar power continues its staggering growth, expanding by 20% in a single year to generate 369 TWh of electricity.

    * The Grid Bottleneck:

    * Europe’s energy ambitions are hitting a wall of wires. A shortfall in grid capacity has put at least 120 GW of planned renewable power at risk.

    * Without immediate action like deploying “non-wire solutions” to optimize existing infrastructure, up to 1.5 million households could face major delays in connecting rooftop solar panels.

    * Currently, almost 700 GW of renewable projects are stuck waiting in connection queues across 10 reporting EU countries.

    * Batteries & The Cost of Wasted Energy:

    * The EU spent €32 billion on imported gas for power in 2025, largely to meet demand during peak evening hours when the sun sets.

    * At the same time, clean energy is frequently wasted. In Germany alone, storing curtailed (wasted) wind and solar energy in batteries could have saved an estimated €830 million by displacing the need to buy expensive fossil gas.

    * The Commercial & Industrial (C&I) Opportunity:

    * While utility-scale and residential batteries are booming, the C&I sector is the “missing middle,” making up only 8% of the EU battery market in 2025.

    * However, the C&I battery market is projected to surge at a 71% compound annual growth rate to €4.1 billion by 2028.

    * Sweden’s battery market currently mirrors California’s in 2019—right at the inflection point of explosive growth. We discuss the “Capital Markets Maximalism” thesis: the idea that physical infrastructure (like batteries) requires parallel financial engineering (like securitization and project finance) to truly scale, similar to how SolarCity’s asset-backed securities unlocked the residential solar boom in 2013.

    Resources & Further Reading:

    * Ember Report: European Electricity Review 2026

    * Ember Report: Crossed wires: Grid capacity could block EU energy security

    * Research Briefs: The coordination layer thesis for European C&I battery storage and Crucible Capital’s Capital Markets Maximalism



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit frahlg.substack.com
    22 min

About Coordinated with Fredrik

From the publisher's feed

Coordinated with Fredrik is an ongoing exploration of ideas at the intersection of technology, systems, and human curiosity. Each episode emerges from deep research. A process that blends AI tools like ChatGPT, Gemini, Claude, and Grok with long-form synthesis in NotebookLM. It’s a manual, deliberate workflow, part investigation, part reflection, where I let curiosity lead and see what patterns emerge.