Featuring
Spice rack
Does public-web access permit unlicensed AI training when the model can substitute for the source?
Original point: Finding copyrighted work through an open-web crawl does not supply a license, especially when the source copy was pirated and the resulting model substitutes for the copyright owner's market.
What everyone argued
Jason Calacanis
Jason argues that technical access is not permission and centers fair use's market-effect factor. He predicts a Spotify-like licensing settlement for closed models, using his own Microsoft book license and his replacement of a Wirecutter subscription with ChatGPT as evidence that models can substitute for paid content.
David Friedberg
Friedberg says publicly viewable information should be readable by an LLM just as a human can read and remember it. For him, the meaningful line is output: copying or repeating a protected work is different from learning patterns and producing a changed synthesis, much as artists and composers learn from predecessors.
Naval Ravikant
Naval sides with Jason on closed models that crawl the open web and then replace the source, but rejects licensing as a stable global solution because foreign developers can crawl anyway. His compromise is reciprocity: if a company trains on open data, it should release the model weights, while owners who refuse crawling will need stronger technical protection.
Winner circle
Jason wins the narrow question because public access alone is not permission, and later law treated acquisition and substitution as facts that matter. Friedberg was right that training can be transformative and that changed, non-substituting outputs strengthen fair use; Bartz and Kadrey confirmed as much. But his blanket human-reader analogy cannot explain away corpus copies or pirated sources, while Jason's ROSS and Spotify extrapolations remain much broader than the precedent supports. The durable rule is annoyingly lawyerly and correct: source, purpose, retention, output controls, and market harm all count.
Commentary
Jason Calacanis
Assumptions and fact checks
Closed AI models will ultimately pay content owners a large, Spotify-like share of revenue.
Why it mattersLicensing markets are growing and the Copyright Office recommends letting them develop, but rights concentration, work value, attribution, transaction costs, fair use, and model economics vary sharply by sector. The New York Times case also remains unresolved.
A model that replaces visits or subscriptions creates legally decisive market harm.
Why it mattersSubstitution is highly relevant to the fourth fair-use factor, but it is weighed with purpose, source, amount used, and output controls. One canceled Wirecutter subscription illustrates the mechanism without proving the aggregate effect or the outcome of a particular case.
The February 2025 Thomson Reuters v. ROSS order rejected ROSS's fair-use defense over copied Westlaw headnotes.
CheckThe district court granted Thomson Reuters summary judgment on fair use and infringement of selected headnotes. The case involved a non-generative competing legal-research tool and copied headnotes, so its reach to general-purpose LLMs is limited; the Third Circuit heard the interlocutory appeal in June 2026 and had not resolved it by this review date.
Spotify pays roughly 65 cents of each music-revenue dollar to music rights holders.
CheckSpotify says roughly two-thirds of every dollar generated from music goes to artists' and songwriters' rights holders. That validates the rounded percentage, but streaming's established collective-rights system does not prove the same revenue split is workable for heterogeneous AI training data.
David Friedberg
Friedberg is right that learning is not the same as replay and later rulings vindicated that distinction. 'A human could read it' is still a metaphor, not a license: it does not answer who supplied the copy, what the developer retained, or which market the output displaces.
Assumptions and fact checks
If a work is freely viewable on the internet, a commercial model should be allowed to train on it without permission.
Why it mattersPublic access does not waive copyright or reveal whether the online copy was authorized. Source legality, commercial purpose, retention, safeguards, and substitution remain relevant even when no paywall blocks a crawler.
The human-learning analogy captures the legally important features of model training.
Why it mattersThe analogy explains statistical learning and non-identical outputs, but a person reading a page and a company making corpus-scale machine copies differ in scale, acquisition, retention, commercial deployment, and market effect. It illuminates transformation without completing the legal analysis.
Fair use means copyrighted content cannot be copied or repeated, but changed output is permitted.
CheckFair use can permit copying, even of an entire work, and changed output does not automatically make a use fair. Courts balance purpose, nature, amount, and market effect, while AI cases also distinguish training, source acquisition, retained libraries, and outputs.
Naval Ravikant
Naval contributes the debate's best policy constraint: a rule that only compliant U.S. firms obey can backfire. His proposed bargain is incomplete because public benefit from open weights does not by itself satisfy the creator whose work supplied the input.
Assumptions and fact checks
Requiring payment for training cannot be stable because foreign developers can crawl the same works and release weights.
Why it mattersCross-border enforcement and open releases are real constraints, but they do not make domestic rights meaningless. Market access, hosting, distribution, contracting, and remedies can still shape developer behavior, while international rules may converge or diverge.
Open-sourcing a model is adequate reciprocity for training on public copyrighted material.
Why it mattersOpen weights can benefit researchers and competitors, but that benefit is not directed to the copyright owner and may intensify substitution. Reciprocity and permission answer different questions.
Will AI eliminate jobs faster than it creates new opportunities?
Original point: Millions of driving jobs will disappear within a decade, just as millions of cashier jobs already did, so automation should affect how the United States thinks about immigration and labor supply.
What everyone argued
Jason Calacanis
Jason treats large gross displacement as close to certain: self-driving will remove millions of driving jobs, cashier automation supplies the precedent, and a country shedding that much work may need fewer new workers. He repeatedly presses the others to admit that visible progress in autonomy makes a large employment shock more than abstract doom talk.
David Sacks
Sacks rejects the premise that millions of net jobs are about to vanish. He separates productivity gains, changes within jobs, and actual job loss; says history makes the first two larger; and argues that policy should not pretend it can predict agentic AI's labor effects before the evidence exists.
David Friedberg
Friedberg says the jobs frame mistakes faster work for redundant workers. In his company, analysts use AI to complete hours of work in minutes and then raise organizational output; he expects capital and labor to rush into newly feasible industries faster than old roles deflate.
Naval Ravikant
Naval argues that technology replaces jobs with better opportunities and that natural-language interfaces make AI unusually accessible. He points to platform work and creator businesses as categories that did not exist before and predicts AI job creation will be at least as fast as destruction.
Winner circle
Sacks wins the question as framed because he keeps gross task displacement separate from net employment and applies the right burden to a confident forecast. Later evidence supports his caution: AI exposure has begun to affect hiring patterns, but it has not produced the systematic unemployment shock Jason assumed. Friedberg and Naval add plausible creation mechanisms, while Jason deserves credit for highlighting concentrated transition risk; that risk calls for measurement and adjustment support, not a claim that the net outcome is already known.
Commentary
Jason Calacanis
Jason is right to force the optimists to name who bears transition risk. He weakens that useful challenge by treating gross losses in selected jobs as proof of a net national shortage of work and by using a false cashier number as the bridge.
Assumptions and fact checks
Autonomous vehicles will eliminate millions of driving jobs within ten years.
Why it mattersThe exposed workforce is large and deployment is advancing, but regulation, capital replacement cycles, geography, remote supervision, fleet growth, and partial automation determine the pace. The episode offers no adoption curve that turns technical capability into a dated employment forecast.
Expected automation losses should lead the United States to admit fewer immigrants.
Why it mattersThat depends on which occupations shrink, which skills immigrants supply, regional shortages, labor-force aging, and how quickly workers can move. A national headcount rule does not follow automatically from displacement in driving or retail.
The United States had already lost millions of cashier jobs.
CheckBLS counted about 3.40 million cashiers in 2014 and 3.16 million in 2024, a decline of roughly 241,000 rather than millions. BLS projects another 313,600 decline by 2034, which supports continued pressure but not the historical magnitude Jason stated.
David Sacks
Sacks wins by distinguishing task disruption from net employment and by keeping uncertainty explicit. His case would be stronger if it paired 'wait for evidence' with concrete triggers for when labor evidence should change the policy response.
Assumptions and fact checks
The historical pattern of productivity growth and job creation will remain larger than AI-driven job loss.
Why it mattersThat remains the better base case in current projections: BLS forecasts 5.2 million net new jobs from 2024 to 2034 while explicitly including AI-related gains and losses. It is still a forecast built on gradual historical adjustment, not a guarantee against a faster shock.
Waiting for clearer evidence is safer than regulating around predicted labor harms now.
Why it mattersRigid limits based on speculative forecasts can misfire, but worker training, portable benefits, measurement, and transition support can be prepared without freezing the technology. Evidence humility need not mean policy inactivity.
There was no evidence that AI had cost any jobs after roughly two and a half years.
CheckNo aggregate data series can prove that AI caused zero individual job losses. Later research found limited evidence of an employment effect and no systematic unemployment increase in exposed occupations, but it also found suggestive evidence of slower hiring for younger workers. The categorical wording outruns what can be measured.
David Friedberg
Friedberg gives optimism a real mechanism—more throughput—not merely a slogan. His company example cannot carry the macro forecast by itself, and he should price the mismatch between the people displaced and the people hired into new industries.
Assumptions and fact checks
Most AI time savings will become additional output rather than lower headcount.
Why it mattersThat is plausible in growing firms with unmet demand and less so in mature businesses with fixed output. Product demand, competition, wages, and management choices determine who captures the productivity gain.
Capital and workers will fill AI-created opportunities faster than displaced roles disappear.
Why it mattersCurrent labor data have not shown mass displacement, but the speed, location, and skill requirements of new work can differ from the jobs that shrink. Net creation does not guarantee a painless or well-matched transition.
Naval Ravikant
Naval is strongest when he attacks the zero-sum job frame and weakest when he turns an easy interface into an easy labor transition. Talking to a model is simple; becoming competitive in a newly reorganized occupation may not be.
Assumptions and fact checks
AI will create opportunities at least as fast as it destroys jobs.
Why it mattersThe aggregate evidence has not contradicted this yet, but no current measure can establish the relative future speeds. New opportunities may also arrive in different regions and require different skills.
Because AI accepts ordinary language, displaced workers will need little retraining.
Why it mattersInterface fluency is only one requirement. Domain judgment, verification, workflow redesign, credentials, access to capital, and the ability to move industries can demand substantial learning and support.
Uber and DoorDash driving jobs did not exist ten years before the episode.
CheckDoorDash began operating in 2013, and Uber reported more than 400,000 active U.S. driver-partners by December 2015. Those platform categories were relatively new, but they existed roughly a decade before this February 2025 recording.
Jason spots the two facts Friedberg's philosophy skips: how the copy was acquired and whether the product competes with the work. He should have kept ROSS in its lane and treated Spotify as one licensing analogy rather than the inevitable destination.