The Fight Over Open Source AI, Anthropic's $1.5B Payout, NYC Socialists: Evictions = Violence?

2026-07-24 - 93 min - source - Read full transcript

Jason Calacanis (host)Chamath Palihapitiya (host)David Sacks (host)David Friedberg (host)

Key insights

Sacks and Chamath argue Anthropic's push to ban Chinese open-source models is regulatory capture dressed up as national security.

Chamath's proof point: if Anthropic actually wanted to stop distillation, the fix is to block Chinese access to Anthropic's own models (or KYC its API customers), not to lobby the US government to ban American developers from using Chinese open-weight releases. He calls it 'the tell' that reveals the national-security argument is pretextual.

ai-open-source-policy

The panel says a US ban on foreign open-source models would function as a hidden tax on every American company using AI, and would tank markets.

Chamath's Coca-Cola example: if US firms are legally restricted to closed-frontier models priced 25-50x above open alternatives, that cost shows up in every enterprise's margins, forcing a re-rating - while simultaneously proving Anthropic/OpenAI's revenue is regulatory-protected rather than market-driven, which craters their own valuations once investors see it.

ai-open-source-policy

Distillation - training on another model's outputs - is framed as a decades-old, cross-industry practice, not IP theft.

Friedberg compares it to Google in its early days submitting queries to Yahoo and Microsoft search engines to benchmark results, or one carmaker studying a competitor's car. The panel's dividing line: copying model weights (the file of trained parameters) would be theft; learning from published outputs is not, legally or historically.

ai-copyright-litigation

Anthropic settled its AI book-piracy lawsuit for $1.5 billion - the largest copyright settlement in US history - specifically because it pirated ~7 million books from sites like LibGen rather than buying even one legal copy of each.

Lawyers reportedly took $100-101M of the settlement; authors get roughly $3,000 per book across 500,000 covered titles, with 91% of eligible authors having already claimed a share. The panel notes that if Anthropic had purchased one copy per book, it likely could have preserved a fair-use defense - the underlying fair-use question for AI training remains unresolved in the courts.

ai-copyright-litigation

Sacks calls Anthropic and OpenAI hypocritical: both insist they can train on all creators' output under fair use, yet call it 'IP theft' when Chinese labs distill from their own models.

He argues this contradiction is dangerous for the labs themselves - if training on others' output without consent is theft, content owners (New York Times, music industry, book publishers) can turn the same argument back on Anthropic and OpenAI and claim a much bigger share of AI revenue than a modest licensing pool would cover.

ai-copyright-litigation

Anthropic's own blog post never actually used the phrase 'IP theft' - it coined 'industrial-scale distillation attacks' instead - because saying 'theft' explicitly would undermine its own fair-use defense in pending litigation.

Sacks argues that once Anthropic and allies started using the IP-theft framing publicly (rather than the more defensible 'deceptive business practice' angle, e.g. fake accounts violating terms of service), it activated the startup ecosystem against them - citing Gary Tan and roughly 200 startups signing a letter, because IP-theft framing would taint every American product built on Chinese open weights (e.g. Thinking Machines' model and Cursor's Composer 2, both derived from Kimi K2.5).

ai-copyright-litigation

Chamath argues frontier AI models are commoditizing on a multi-week cycle rather than the traditional five-to-ten-year cycle, which is why closed labs are racing for government protection now.

He points to open-weight models matching or beating newly published closed-model benchmarks within weeks, arguing capital allocators will inevitably conclude foundation models aren't a durable moat, shifting real value to the application layer and to infrastructure/chip providers instead.

ai-open-source-policy

Sacks disputes that frontier labs are actually struggling, pointing to accelerating revenue as evidence they don't need government protection.

He cites Anthropic's ARR growing from about $10B in January to over $70B by mid-2026, and OpenAI's ARR reportedly rising from $33B in May to $41.3B in July with an internal forecast near $75B exit ARR - framing any perceived Anthropic revenue dip as a data artifact, not evidence of distillation-driven harm.

ai-open-source-policy

Friedberg frames China's long-game strategy as commoditizing the global knowledge and services economy via open-source AI so that value shifts back to physical production, where China already holds a large capacity edge.

He cites the US at roughly 1 terawatt of electricity capacity versus China's build-out toward 8 terawatts, and about 10 billion square feet of US manufacturing space versus 200 billion in China - arguing that if AI flattens the value of 'bits,' the scarce resource left is 'molecules' (manufacturing and energy), which China is positioned to dominate.

ai-open-source-policy

The panel reads Google's record capex guide and first-ever negative free cash flow as a buy signal, not a red flag, given its 25-year average return on invested capital of roughly 32%.

They argue Google is uniquely positioned to profit regardless of which AI model wins because it is model-agnostic (Google Cloud can serve any model), owns meaningful stakes in Anthropic, SpaceX and Waymo, and would still have world-class infrastructure to monetize even in a worst-case scenario where its own models and application-layer bets fail.

big-tech-capex-earnings

Friedberg argues NYC's eviction and rent-control 'violence' framing echoes an early-stage tyranny playbook: moralizing an owner as evil before revoking their property rights.

Citing John Quincy Adams's 18th-century writing that property rights are foundational to liberty, he argues every step toward removing landlords' rights - background-check bans, eviction bans, rent freezes - follows the same pattern of first framing the property owner as the aggressor to justify taking control of the property.

nyc-rent-control-property-rights

Sacks argues NYC's proposed tenant-screening and eviction restrictions will make housing worse for the tenants they're meant to protect, not just landlords.

He predicts landlords facing an inability to vet tenants or evict non-payers will respond by tripling or quadrupling asking rents and demanding large prepayments to self-insure, or by leaving units vacant entirely - citing roughly 50,000 reported vacant 'ghost apartments' in New York City tied to renovation-cost and rent-control dynamics, plus a parallel Airbnb ban limiting alternative uses for empty units.

nyc-rent-control-property-rights

Media referenced

Stratechery (Ben Thompson) - Kimi K2 cost analysis - article - Sacks cites Ben Thompson's post-launch cost breakdown of Kimi K2 to argue it is not actually much cheaper to run than claimed, undercutting the panic-narrative.
Anthropic blog post coining "industrial-scale distillation attacks" (Feb 2026) - article - Sacks argues the post never uses the phrase "IP theft" - deliberately, since claiming that would undercut Anthropic's own fair-use defense in its copyright lawsuits.
Axios report on White House considering a Chinese open-source model ban - article - Monday report that kicked off the week's open-source policy debate; Wired later reported Commerce Secretary Howard Lutnick opposes a ban.

Companies

Anthropic - Center of the episode: accused of pushing a Chinese open-source ban as regulatory capture, and separately settled a book-piracy copyright suit for $1.5B.
OpenAI - Named alongside Anthropic as part of the closed-frontier duopoly; also being sued by the New York Times over training-data scraping; ARR reportedly re-accelerated to $41.3B by July.
Moonshot AI - Chinese lab behind Kimi K3/K2.5, the open-weight model release that triggered the week's distillation and open-source-ban debate.
Google / Alphabet - Q2 earnings covered: record ~$195-205B capex guide, first-ever negative free cash flow, 32% 25-year average ROIC, and equity stakes in Anthropic, SpaceX and Waymo.
Tesla - Capex surged 140% YoY to a projected $25B; stock down ~14% on the earnings reaction alongside Google's ~7% drop.
SpaceX - Now trading around $1.5T, down ~30% from its IPO-day close of $2T, with staggered lockup expirations adding pressure.
Thinking Machines - Mira Murati's startup; its top American open-weight model was bootstrapped by distilling from the Chinese Kimi K2.5 model, cited as proof the IP-theft framing would also taint American derivative work.
Cursor - Its Composer 2 model was post-trained on Kimi K2.5 plus proprietary coding data, another example of an American product built on a Chinese open-weight base.
8090 - Sponsor read for a software-implementation consultancy; promo code J-CAP.
New York Times - Suing OpenAI over training-data scraping; cited throughout as the model case for content-owner leverage against AI labs.

Techniques and frameworks

Distillation (model training technique) - Training a new model on another model's outputs rather than its weights; panel frames it as equivalent to benchmarking or reverse-engineering, not theft of the underlying software.
KYC (know-your-customer) verification - Chamath's proposed real fix for industrial-scale distillation - require identity/payment verification on API accounts - which he says Anthropic avoids because it would slow growth.
Regulatory capture - Sacks and Chamath's central charge against Anthropic: using national-security and IP framing to lobby for a government-protected duopoly rather than competing on merits.
Molecule economy vs. knowledge economy - Friedberg's framework for China's long game: if AI commoditizes the knowledge/services economy, value shifts to whoever can convert molecules (manufacturing, energy) cheapest - where China already holds a large capacity edge.

Summary

The episode opens on the week's dominant AI story: Moonshot AI's release of the open-weight Kimi K3 model, which reportedly matches frontier closed models like Opus and GPT on some benchmarks at roughly half the cost, reigniting "DeepSeek moment" anxiety and pushing a Chinese open-source-model ban onto the White House's radar (Polymarket priced a 2026 ban at 45%, up from 22% days earlier). Sacks and Chamath spend the first third of the show arguing this is regulatory capture: Anthropic, they say, is the fastest-growing tech company in history (roughly $10B to $70B+ ARR in a matter of months) and does not need government protection from competitors. Their central claim is that if Anthropic genuinely wanted to stop Chinese labs from distilling its models, the fix is on Anthropic's side - KYC verification, blocking Chinese-origin accounts - not a US ban on American developers using anything built on Chinese open weights. Chamath extends the argument economically: forcing US enterprises onto closed models priced far above open alternatives would function as a hidden tax and would, paradoxically, prove to markets that closed-lab revenue is regulatory-protected rather than market-driven, tanking their valuations. Friedberg largely agrees, framing distillation (learning from a competitor's output, as opposed to stealing its underlying weights) as an old, cross-industry practice rather than theft, and argues open source has historically diffused value across an entire ecosystem (his Netscape/Mozilla/Apache browser-wars analogy) rather than concentrating it in a few billionaire-owned labs. Sacks pushes back moderately on the "frontier labs are struggling" framing, citing accelerating ARR for both Anthropic and OpenAI as evidence neither needs protection - the group lands somewhere between "open source is winning on cost and adequacy for ~95% of tasks" and "closed frontier labs will survive by moving up the application-layer stack."

The second segment covers Anthropic's $1.5 billion settlement of its AI book-piracy lawsuit - the largest copyright settlement in US history, stemming from Anthropic downloading roughly seven million pirated books to train Claude rather than purchasing even single copies. The panel's sharpest point is about hypocrisy: Anthropic and OpenAI maintain it's fair use to train on all of the world's creative output without consent, yet frame Chinese labs' distillation of their own model outputs as "IP theft" - a claim Sacks notes Anthropic itself never made explicitly in its blog post (it coined "industrial-scale distillation attacks" instead), because doing so would undercut its own fair-use defense in pending litigation. This inconsistency, the panel argues, has both activated the startup ecosystem (Gary Tan and roughly 200 startups signed a letter opposing the framing, since it would taint American products like Thinking Machines' model and Cursor's Composer 2 that were built on the Chinese Kimi K2.5 base) and handed ammunition to content owners like the New York Times to argue AI labs owe them far more than a modest licensing pool.

A shorter markets segment covers Google's and Tesla's earnings: Google guided to $195-205B in 2026 capex and posted its first-ever negative free cash flow, sending the stock down ~7%, while Tesla's capex rose 140% YoY with the stock down ~14%. The panel reads both as a buy signal rather than a warning sign, especially for Google, whose 25-year average return on invested capital is roughly 32% and which benefits from being model-agnostic across any AI provider while also holding equity stakes in Anthropic, SpaceX, and Waymo.

The final segment, labeled "Socialism Corner," turns to New York City politics under Mayor Zohran Mamdani, including new rules restricting landlords' ability to run credit and income checks on prospective tenants and rhetoric describing evictions as "violence." Friedberg delivers an extended argument - anchored in John Quincy Adams's writing on property rights as the foundation of American liberty - that framing property owners as morally culpable is the first step in a broader erosion of property rights that ends in tyranny. Sacks adds a more practical critique: restricting landlords' ability to screen or evict tenants will primarily hurt the other residents in a building (through deteriorating conditions and problem neighbors who can't be removed) and will push landlords toward defensive behavior - tripling asking rents, demanding large prepayments, or simply leaving units vacant, citing roughly 50,000 reported vacant "ghost apartments" in the city.

Notable Quotes

"The tell on this, the way that you know that this whole distillation thing is fake, is because if stopping distillation was their primary objective, Anthropic would push to ban Chinese access to American models, not American access to Chinese models." - David Sacks

"IP for we, but not for thee." - David Sacks

"All of technology ultimately leads to that simple equation: molecule conversion... The value of the U.S. and the Western economy has been largely degraded [if the knowledge economy compresses], and what's left is the value of the molecule economy." - David Friedberg

"The moment the idea is admitted into society that property is not as sacred as the laws of God and that there is not a force of law and public justice to protect it, anarchy and tyranny commence." - David Friedberg, quoting John Quincy Adams

"If you were forced to live under these rules as a landlord, what would you do? The obvious answer is you'd start rent three or four times higher... So run the experiment and let's just observe what happens." - Chamath Palihapitiya