Open Source AI Is Non-Negotiable

↓ Jump to the video and timestamps

I've been to Washington more times in the last six months than I was in the first 34 years of my life. That's not something I'm bragging about. I'd rather be home in Austin, gardening, stacking, doing the show from my studio. But the fight is in DC right now, and Conner Brown has been making that case to me every time we sit down together.

This was our third time recording in person, and as Conner puts it, DC is going to be the city for the next five to ten years whether we like it or not. The technology changes are accelerating, and the policy decisions being made in that city are going to shape whether any of us get to use these tools freely.

Conner is at the Bitcoin Policy Institute, where he's been since we last recorded, and the BPI just dropped their Q2 report covering three months of work that honestly reads like five years of output at a traditional think tank. Taiwan. The UK. A foreign influence investigation that landed BPI on the front page of the New York Times. An AI research paper showing agents prefer Bitcoin. And a deepening alarm about what's building on the policy front around open-source AI models.

We covered all of it.

The through-line I kept coming back to: this is the Bitcoin fight by another name. The same coalition of rent-seekers, Malthusians, foreign adversaries, and regulatory-capture artists that tried to strangle Bitcoin with compliance theater is now lining up against open-source AI. And if America fumbles this the way we fumbled nuclear energy, we're going to spend the next 50 years paying for it.

Key takeaways

• Open-source AI is a sovereignty question, not a technical preference. The logic that makes you run your own node and hold your own keys applies directly to AI models. Closed-source systems that phone home to Silicon Valley, and the government, are surveillance infrastructure by definition. I won't get a Roomba for this reason. I'm certainly not putting a closed-source humanoid robot in my house.
• America is replaying the nuclear energy mistake in real time. The NRC was celebrated by the nuclear industry at launch and proceeded to kill decades of progress. The same institutional panic reflex is forming around AI before a single documented mass harm has occurred. It took roughly 50 years to get back to a semblance of sanity on nuclear. We cannot afford that timeline with AI.
• A genuinely strange coalition is lining up against the buildout. Foreign adversaries, radical environmentalists, Malthusians, economic populists, Marxists, national security hawks, EA doomers, and frontier-model incumbents are all rowing the same direction. Not by coordination, by converging interest. EA money is now reportedly funding MAGA-aligned anti-AI populism. These are not normal political bedfellows.
• If America restricts open-source frontier models and China doesn't, the world builds on Beijing's stack. The strategic depth goes far beyond content filters on Tiananmen Square. It extends to subtle optimization for Huawei chips, CBDC preference baked into the weights, and soft power at civilizational scale. Someone needs to clip this conversation and get it to Howard Lutnick.
• AI agents will naturally prefer Bitcoin if nobody rigs the training set. BPI's research shows that models queried without a thumb on the scale prefer Bitcoin and stablecoins. Bitcoiner content is prolific and consistent in training data. Agents are digitally native. Bitcoin is digitally native. The risk is closed-model operators deciding that their frontier release has to "really like CBDCs" before it ships.
• The company brain is the highest-ROI investment in the AI stack. Persistent memory across three layers, flat markdown, vector database, knowledge graph with temporal context, lets a small team punch far above its weight while preserving institutional voice. TFTC has this running. BPI has a version of it too. Every org that hasn't built something like this is going to feel the gap widen fast.

DC Matters Now Whether I Like It or Not

The reason I keep coming back to Washington is that the important decisions are being made there. Conner framed it well: he'd prefer it not be that way, but with the pace of change in AI and Bitcoin policy, DC is where the influence is for the next decade.

BPI's Q2 report, released the morning we sat down, is a good illustration of how much ground they've covered. Taiwan. The UK. A deep-dive into foreign influence on data center opposition. Original research on AI agents and Bitcoin preference. Conner said he read through the full quarter's output and couldn't believe they'd done it in three months.

I believe it. Everyone I know who's actually leaning into these AI tools is working harder than ever, not less, because suddenly the stray ideas that used to die from lack of bandwidth can actually get executed.

I've been doing my part at TFTC. We recorded a session on energy policy that morning. That same night I was at an event with PubKey and American Bitcoin, part of a ten-month series on energy, data centers, and how to educate the public on all of it.

I had a conversation the night before with a senior executive from one of the largest utility companies in the country about exactly this: how do you explain to normal people how grids work, how data centers interact with electricity markets, why cheap energy access depends on getting this buildout right. Bitcoiners have a decade of that trench warfare behind them. The AI data center crowd is just starting to learn it.

Taiwan, the UK, and What the SBR Is Actually Doing Internationally

The Taiwan story is one of my favorites from the BPI Q2 work, because it started as a thought experiment.

BPI had a fellow with a DOD background, and they said: write a paper on why Taiwan diversifying reserves into Bitcoin makes strategic sense. The historical hook is real, Taiwan exists in its current form because the Republic of China government airlifted gold off the mainland to fund a government in exile. That's the founding logic. BPI built a research paper around that and published it.

Two weeks later, Conner got an email with photos from Taiwan's legislature. A member of the Legislative Yuan, Legislator Ko, had personally delivered the paper to the premier of the Legislative Yuan and the head of the central bank. Within weeks, BPI was invited to brief them for a week in person.

Conner's read from those meetings: every person they talked to in Taiwan is acutely aware of the geopolitical weight they're carrying. The silicon shield argument, that America defends Taiwan because of how critical its chip production is, is stronger today than it was twenty years ago, but it's also why China has even more strategic incentive to act.

Both political parties, the DPP and the KMT, largely agree on domestic policy. The fault line is almost entirely about posture toward China. And everyone in every meeting was thinking about what the AI compute buildout means for Taiwan's strategic position.

What surprised Conner most was the central bank meeting. He expected polite skepticism. Instead, every member of the delegation had read the BPI report cover to cover, taken extensive notes, and came with substantive questions. That's a signal.

On the UK front: BPI went over expecting resistance and found genuine openness. They briefed members in the House of Lords, Conner said it looked like the Hogwarts cafeteria, and a security guard caught him trying to take a photo of a 1600s statue.

The UK had just announced a cap on stablecoin balances, then rescinded it the same week BPI was in town. Still finding their footing, but the posture was receptive. The head of innovation at the UK's FCA put out a post afterward about the importance of Bitcoin policy. These things are moving.

The SBR executive order is still being digested globally. Treasury and Commerce have joint delegated authority to execute it, and Conner was diplomatic about the pace. He called the whole thing "a circus," and I said "comical," and we both know how this goes: two weeks TM. Same as it ever was.

The Foreign Influence Campaign Against American Data Centers

This is the work that put BPI on the front page of the New York Times, and it started because Bitcoin miners started noticing something weird.

Miners are used to local opposition when they're building a new facility. People ask reasonable questions: noise, location, impact on the area. What they started seeing a few months ago was different. People from out of town showing up to local meetings making claims that had no grounding in reality, "no bees can survive within 50 miles of a data center" was one Conner cited directly.

Claims that bore no relationship to anything documented. And the same pattern kept appearing across different project sites.

BPI brought in Sam Lyman, who came over from Treasury and previously worked at Riot, and he started pulling on the threads. You go through layers of nonprofit structures and local activist groups, and a pattern emerges. The Party for Socialism and Liberation keeps appearing as an organizer across multiple different project sites.

Conner was careful here, and I want to be careful here too, this doesn't mean every American with legitimate concerns about a trillion-dollar compute buildout is a foreign agent. People have real questions. Infrastructure projects of this scale deserve scrutiny. But a proper accounting has to include asking whether foreign influence is in the mix, because we know our adversaries are playing the fifth-generation warfare game.

The modern battlefield is Twitter and Instagram.

The figure BPI identified as funding some of the nonprofit infrastructure doing this work is Neville Roy Singham, a wealthy philanthropist who lives part-time in Shanghai and has made statements aligned with Chinese government positions. The dark-money structure makes it hard to trace fully, but BPI has two published foreign influence reports laying out what they've found. Marc Andreessen amplified the report, the internet caught fire, and BPI landed on the front page of the Times.

We deal with this at TFTC too, on a different scale. You have to engage in fifth-generation warfare, that means going, pulling on the threads, writing the reports, getting it out there. The protest campaign against AI data centers is broader than most people realize, and the foreign influence angle is the part that isn't getting nearly enough coverage outside of what BPI is doing.

How We're Actually Building With AI

Before we got into policy, Conner and I compared notes on implementation, because both of us have been going deep on this stuff and I think the real-world workflow is more useful to people than the abstract debate about what AI can do.

BPI runs on Claude Enterprise. They built out a shared project file with a top-level CLAUDE.md that gives the system an overview of what BPI is and how it works, then subfolders for each arm of the organization, comms, design, research, each with its own CLAUDE.md, example workflows, prior precedent, and encoded research lifecycles.

The research skill kicks off with ideation, hits a user feedback gate, continues, hits another gate, and eventually hands off to comms or design. It's an assembly line. Conner's biggest lesson from building it: context poisoning is real. If the system makes an error and you just fix it manually without giving it feedback, the bad pattern persists.

The right move is to close the loop: explain the error, and tell it to never make that mistake again. Every error is an opportunity to permanently fix a class of problem, if you treat it that way.

At TFTC, we've gone deeper on the memory stack. We have an agent I call Martin, a more sophisticated version of me, with better file storage. He lives on a VPS I control. The company brain is three layers: flat markdown files at the base (soul.md, agent.md, skill.md, daily notes), a QMD semantic search vector database that queries on keywords without burning tokens, and Cogni, an agent-first Obsidian vault, as the third layer, which handles interrelational context and temporal tracking.

So when I prompt on a topic, the system knows every time we've discussed it, how our position has evolved, and how we frame it. All of our transcripts and newsletters are in there. Martin knows the TFTC voice.

The newsletter workflow now: I drop links into our Bitcoin Brief Telegram topic throughout the day. Morning of, Martin pulls everything, structures a lead story and signal stories based on what's historically driven engagement in our Ghost CMS backend, proposes framing, I push back or refine via voice, we land on a structure, Martin drafts into Ghost, I read it twice and give final edits, we send.

Everyone says you're going to lose the skill of writing. I've heard that. But I think about it differently now because of a conversation I had with someone who reframed the whole thing for me.

The framing that stuck: we're returning to the era of the scribe. Conner had independently landed on the same mental model, which tells you something. Napoleon, after a day of battle, would go back to the war tent and dictate to multiple scribes simultaneously, one for munitions, one for strategy, one for correspondence with foreign leaders, faster than they could write.

Each scribe got a specific context and a specific task. That's the right way to think about prompting. Spin up a session, give it maximum context so the model can eliminate associations and give you narrow, precise output. It forces you to think from first principles and articulate what you're actually trying to get out there.

I use Super Whisper with a local model for voice-to-text. Conner uses Whisper and Flow. Either way, voice unlocks output speed that typing never could.

The best ROI we've gotten from any part of the AI stack is the company brain. If you run a small team and you haven't built persistent memory into your workflow, that's where I'd start.

The NRC Trap

The policy section of this conversation is where I got genuinely alarmed, and I'm not someone who defaults to alarm.

Conner's nuclear analogy is devastating in its simplicity. We invented basically unlimited energy, then banned it. We built enough weapons to blow up the world several times over. The NRC was announced to industry applause, every major nuclear figure celebrated it as a great day for American energy. And it proceeded to kill all meaningful progress.

The person who ran the NRC under Obama has gone on stage and openly bragged about how he stopped nuclear at every turn. That's the institution. That's what it became.

I asked Conner: how long did it take to get back to something resembling sanity? About 50 years, roughly. Maybe longer if you're honest about where we actually are.

New York just signed the first statewide AI data center moratorium, and that's not an isolated event. The open-source AI regulatory playbook is already being pulled out of the drawer, and it looks familiar.

The COVID parallel is sharp. Trump, in theory pro-freedom, pro-small business, pro-innovation, got told it was an existential risk and shut everything down indefinitely. Printed a bunch of money. Logic went out the window because someone said "existential."

Conner's worry is that the same one-shot works on AI: before we've seen a single major documented harm from the technology, we get a catastrophic incident (real or manufactured), someone says existential, and the institutional reflex kicks in. We regulate it into uselessness or worse, into closed fiefdoms controlled by a handful of players who already have government relationships.

What makes the current moment particularly strange is the coalition lining up on the restrict side. Conner laid it out, and it is genuinely weird: foreign adversaries who benefit from slowing American AI progress, radical environmentalists who want to de-develop civilization, Malthusians, economic populists worried about jobs, classic Marxists who think everything should be redistributed, national security hawks who want compute restricted to government hands, EA doomers who have convinced themselves the weights themselves are dangerous, and frontier-model incumbents who want a regulatory moat.

None of these groups are coordinating. They don't need to. Their interests all happen to converge on the same outcome: restrict access, close the weights, license the compute.

The EA doomer funding dynamic is particularly troubling. According to Conner, there's a campaign called "Humans First", MAGA-aligned, anti-AI populism, that is being funded by Effective Altruist money. Conner hedged on the specific name of the organizer, so I'm going to leave it there.

But the dynamic is real: EA money funding a Trump ground-game operative to run an anti-AI campaign targeting Republican voters. That's a sophisticated play. It's not random.

Dario Amodei testified in the Senate in 2023 that open-source AI is "headed down a very dangerous path." That was two years ago. Anthropic has continued to position itself as the responsible, safety-first frontier lab. One has to wonder whether to view that positioning with some skepticism, because it is an easy way to get a regulatory moat.

If you're following economic incentives, that's probably something they want. Sam Altman's strategy, by contrast, is apparently just to offer the government 5% equity. Different methods, similar destination.

If we ban open-source frontier models while China publishes theirs, the entire world builds on Beijing's stack. Conner put it plainly: the strategic depth goes way beyond content filtering. It's about a model that is subtly better at programming for Huawei chips, that has CBDC preferences baked into the weights, that optimizes incrementally for China's strategic interests in a thousand small ways.

That's the soft power play. It's enormous. It seems insane to me that this isn't the first thing anyone in the administration is gaming out. Someone clip that part of the conversation and get it to Lutnick.

Open Source AI and Bitcoin, the Same Fight

Conner's research at BPI on AI agents and Bitcoin is the piece I've been thinking about most since we sat down.

When you query capable models without putting a thumb on the scale, they prefer Bitcoin and stablecoins for value transfer. This is across different models, and it makes sense if you think about it from first principles. Agents are digitally native. Bitcoin is digitally native.

Nick Szabo wrote extensively about mental transaction costs, the friction humans feel around small payments, around trust, around verification. Agents don't have any of that. Bitcoin's properties just click for a machine in a way they don't have to for a human. The mental overhead that makes everyday Bitcoin payments feel cumbersome to people is largely absent for software agents operating autonomously.

Bitcoiner content is also prolific in frontier model training sets. The arguments are consistent, technically rigorous, and have been published at high volume for over a decade. Conner mentioned that Gregory Maxwell (GMax) appears as a named token in GPT-3's vocabulary, that's how much Bitcoiner content made it into the early training data.

TFTC transcripts and newsletters are in there. The community has been publishing coherent, detailed, well-reasoned content about sound money and digital sovereignty for years. Machines trained on that data are going to reflect it.

The risk Conner flagged is the flip side, and it's the one that keeps me up a little. He almost didn't release the research because he knows that people running frontier labs are aware their models have preferences baked in, and publishing "AI agents prefer Bitcoin" gives them a reason to go fix that.

In a closed-source world, they can do exactly that. Before a new model ships, it has to "really like CBDCs." Or these preferred stablecoins. Running on the x402 protocol. That's what you prefer. Bake it in, ship it, nobody outside the lab knows.

That's the fiefdom scenario. Permissioned AI, surveilled outputs, preferences shaped by whoever controls the weights. It's not hypothetical, it's the natural endpoint of the closed-source trajectory if nobody pushes back.

I won't get a Roomba vacuum for this reason. I'm absolutely not putting a closed-source humanoid robot in my house. The principle is exactly the same as a hardware wallet: I want to know the firmware. I want to feel good about what I'm running.

Open weights, like open source code, are the thing that lets you verify. Without that, you're trusting a company and whatever relationships that company has with regulators and governments. Closed-source AI without open weights is the same problem Bitcoin was built to solve, just applied to a different layer of the stack. Anthropic has already demonstrated what closed-source opacity looks like in practice.

About Conner Brown

Conner Brown is a policy researcher at the Bitcoin Policy Institute in Washington, DC, where he leads research across Bitcoin policy, open-source AI, and energy infrastructure. Before joining BPI, he worked in policy roles in DC, including at a Washington law firm. He is a co-author of BPI's foreign influence reports on data center opposition and the organization's research on AI agent payment preferences.

Sources mentioned

• Bitcoin Policy Institute Q2 2025 Report (BPI): the quarter's work Conner and Marty discuss across Taiwan, the UK, foreign influence, and AI research
• BPI Foreign Influence Reports on AI Data Center Opposition (BPI): the published research identifying the Party for Socialism and Liberation and the nonprofit structures behind data center opposition campaigns
• 142 Protests in 42 States Signal Hard Ceiling on AI Data Center Buildout (TFTC): the scale of organized opposition to compute infrastructure across the country
• Open-Source AI Faces the Same Regulatory Playbook That Targeted Bitcoin (TFTC): the regulatory pattern Marty and Conner discuss in depth
• New York Signs First U.S. Statewide AI Data Center Moratorium (TFTC): the NRC-trap dynamic playing out at the state level
• Anthropic Hid a Steganographic Tracker in Claude Code for Three Months (TFTC): what closed-source opacity looks like in practice
• Nick Szabo, "Mental Transaction Costs" (Satoshi Nakamoto Institute): the concept Conner cites on why agents face fewer barriers to Bitcoin adoption than humans

Watch the conversation

Timestamps

• 0:00 - Intro
• 1:09 - Why DC Is the City for the Next Decade
• 3:35 - BPI Q2 Report Overview
• 7:39 - Taiwan: Legislator Ko and the Central Bank Briefing
• 11:29 - SBR Executive Order and Bureaucratic Hurdles
• 14:52 - Foreign Influence Against American Data Centers
• 20:29 - BPI Hits the Front Page of the New York Times
• 29:50 - How BPI Is Implementing AI: Skills and Assembly Lines
• 32:44 - The Company Brain: TFTC's Three-Layer Memory Stack
• 40:26 - Voice-to-Text, Martin, and the Scribe Model
• 53:18 - The NRC Trap and the Nuclear Analogy
• 58:17 - The Strange Coalition Against Open-Source AI
• 1:02:30 - Anthropic's Regulatory Moat vs. OpenAI's 5% Strategy
• 1:16:25 - If America Restricts Open Source and China Doesn't
• 1:18:31 - AI Agents Prefer Bitcoin: BPI's Research
• 1:20:36 - The Fiefdom Scenario and the x402 Risk

Sponsors

• Bitkey: Bitcoin self-custody hardware from Block. Part of a broader ecosystem with Cash App and Square to make Bitcoin easier to buy, hold, and use. Code TFTC for 10% off. bitkey.world
• Aven: Bitcoin-backed line of credit up to $1M, custodied by BitGo, no rehypothecation, fixed rates from 7.99% APR for up to 10 years, 2% unlimited cashback. aven.com/bitcoin
• Unchained: collaborative multisig custody and Bitcoin financial services. Over $12 billion secured for more than 12,000 clients. You hold 2 keys, they hold 1. Code TFTC10 for 10% off a vault. unchained.com/tftc
• CrowdHealth: crowdfunded healthcare. Marty and his family have been members for five years, two babies, multiple health events, never going back. Family of five, roughly $700/month. Code TFTC, $99/month for the first three months. joincrowdhealth.com/tftc