Featuring
No guest this week. The besties move from frontier-AI safeguards to Amgen’s $27.8 billion Horizon deal, with quick stops at Apple’s coming headset and frozen real estate. The spice sits in the gap between spotting a real danger and writing a rule that can survive contact with reality. Friedberg has the best episode because he keeps dragging both debates back to that test.
Spice rack
Would licensing and compute stage gates make frontier AI safer, or mainly protect incumbents while open models route around them?
Original point: AI has overwhelmingly beneficial uses, but a small set of catastrophic applications justifies slowing the largest training runs with licensing or customer-identification stage gates before dangerous capabilities spread.
What everyone argued
Chamath Palihapitiya
Because a tiny number of malicious uses could cause extraordinary harm, cloud providers should identify customers seeking massive GPU clusters and ask what they intend to build before dangerous models become cheap and portable.
Jason Calacanis
Jason split the difference: reject a general licensing agency, but establish a commission and monitor high-risk uses because unsupervised agents could automate fraud, hacking, or other damaging activity.
David Sacks
Altman was playing regulatory chess: concede Washington's risk narrative, help write the standards, and turn compliance into an incumbent moat. Existing criminal and civil laws should punish harmful conduct rather than making entrepreneurs seek permission to build software.
David Friedberg
Model regulation is a scope mismatch: open models get smaller, cheaper, frequently modified, embedded in applications, and run at the edge. Auditing every server or model revision would become a fool's errand, so government should focus on tractable risks and uses rather than pretend it can approve the whole software layer.
Winner circle
Friedberg wins by separating a tractable frontier-compute control from the impossible promise of auditing the whole model ecosystem. Chamath was right that catastrophic risk deserves action before harm, and later reporting policy showed his narrow choke point was not fantasy. Sacks was right about capture but too confident that existing law was enough; Friedberg gave the cleanest implementation test without dismissing the risk itself.
Commentary
Chamath Palihapitiya
Assumptions and fact checks
A small set of catastrophic AI uses justifies intervention before the relevant capabilities become cheap and widely distributed.
Why it mattersFor high-consequence biological, cyber, or autonomous capabilities, waiting for realized harm is a poor safety strategy. The intervention still needs capability-based thresholds and safeguards against incumbency protection.
Customer identification at AWS, Google Cloud, and Azure would cover the training runs most likely to create dangerous frontier capabilities.
Why it mattersIt can cover conspicuous commercial runs, and later federal policy used similar reporting logic. It misses smaller models, fine-tuning, private clusters, foreign compute, stolen weights, and dangerous applications built from already released systems.
Jason Calacanis
Jason's middle path was directionally sound. He would have been stronger had he tied each feared use to a specific intervention instead of stacking vivid examples.
Assumptions and fact checks
A standing expert commission can improve preparedness without becoming a licensing bottleneck.
Why it mattersEvaluation capacity, incident reporting, and technical advice can improve governance without requiring prior permission for every model. Its influence and accountability still need clear limits.
Sam Altman proposed licenses for AI models generally to be trained and used.
CheckAltman proposed licensing development and release only above a capability threshold and explicitly said smaller models and open-source work should remain outside the regime. Jason's summary erased that limiting condition.
David Sacks
Sacks supplied the best warning about who writes and survives a licensing regime. He turned a strong anti-capture argument into a weaker claim that punishment after the fact solves every risk.
Assumptions and fact checks
Altman's primary motive was to build a regulatory moat around OpenAI.
Why it mattersOpenAI plainly could benefit from standards it helps shape, but incentive is not proof of primary motive. The proposal also imposed audits and revocation risk on OpenAI itself.
Existing law can handle the material harms from advanced AI without model-level oversight.
Why it mattersExisting law covers fraud, theft, violence, and many downstream harms, but it does not by itself provide pre-deployment capability testing, incident reporting, or prevention where attribution is hard and damage may be irreversible.
Altman proposed a new agency that could license frontier efforts, revoke licenses, set safety tests, and require independent audits.
CheckThe official hearing transcript records all four elements, limited to models above a scale or capability threshold.
David Friedberg
Friedberg kept the debate attached to implementation and caught the central category error: regulating every model is not the same task as monitoring a few frontier training runs. That distinction carried the ruling.
Assumptions and fact checks
Broad model licensing would be unenforceable once capable open models were cheap, portable, and frequently modified.
Why it mattersThe enforcement surface grows far faster than an approval agency could inspect it. Targeted rules for frontier runs, deployments, or harmful conduct remain possible, but they are narrower than model licensing as a whole.
MosaicML trained MPT-7B from scratch for about $200,000 using public data and released it for commercial use.
CheckMosaicML reported a 9.5-day training run costing about $200,000 on one trillion tokens, with open weights and commercial-use terms. That did not make MPT-7B equivalent to the largest frontier systems, but it supported Friedberg's cost-diffusion point.
Should the FTC have blocked Amgen's Horizon deal, or targeted exclusionary drug bundling without killing the merger?
Original point: Blocking a large pharma acquisition threatens the M&A exit market that finances risky biotech development, so regulators should not mistake deal size for competitive harm.
What everyone argued
Chamath Palihapitiya
Large pharma relies on smaller biotech firms to carry early scientific risk, and those firms rely on acquisition exits. Blocking a mature $27.8 billion deal would chill the same crossover investors who finance young companies; PBMs, not the merger itself, were the better target for drug-price inflation.
Jason Calacanis
Regulators should police anticompetitive tactics rather than reflexively stop acquisitions: allow deals that expand investment and consumer choice, then prohibit bundling, self-preferencing, or monopoly leverage that blocks rivals.
David Sacks
If bundling is the problem, ban bundling instead of blocking M&A. Removing acquisitions as an exit suppresses already difficult venture returns, while a targeted conduct rule preserves investment incentives and addresses the alleged abuse directly.
David Friedberg
Horizon was not an early-stage speculative biotech but a mature company with about $4 billion in revenue and substantial earnings. Amgen's larger portfolio could bundle rebates and pressure payers to favor Horizon's monopoly drugs, raising entry barriers; the FTC therefore had a compelling competition theory.
Winner circle
Sacks wins the remedy question, while Friedberg wins the diagnosis. The FTC ultimately did almost exactly what Sacks proposed: it let the acquisition close and prohibited the bundling and exclusionary rebate tactics Friedberg had identified. Chamath was right that exit incentives matter, but he leaned too heavily on an early-stage biotech story that did not describe Horizon itself.
Commentary
Chamath Palihapitiya
Chamath was strongest on incentives and weakest when he let the general biotech-financing story substitute for the facts of a mature, cash-generating target.
Assumptions and fact checks
Blocking this mature-company acquisition would materially reduce funding for early-stage biotech.
Why it mattersExit expectations affect venture returns, but one challenged mature portfolio deal does not automatically close the acquisition market. The specific spillover was asserted through private conversations rather than demonstrated.
A Phase 3 program can require five to eight billion dollars of financing.
Why it mattersLarge drug programs are expensive and total capitalized development costs can reach billions, but FDA describes typical Phase 3 trials as 300 to 3,000 participants over one to four years. Five to eight billion for a Phase 3 trial itself is not a credible general benchmark.
The Stemcentrx acquisition was a roughly $10 billion biotech failure.
CheckAbbVie paid about $5.8 billion upfront with contingent payments that could lift the value toward $10 billion. It later recorded a $5.1 billion impairment after stopping a Phase 3 Rova-T trial, making 'dud' fair shorthand, though $10 billion was the potential rather than fixed upfront price.
Jason Calacanis
Jason captured the eventual remedy almost exactly. He could have acknowledged that conduct orders require monitoring and may fail when the competitive harm is structural rather than contractual.
Assumptions and fact checks
A conduct remedy can preserve competition here without blocking the transaction.
Why it mattersThat became the actual settlement: the deal closed while Amgen accepted a monitored ban on bundling and exclusionary rebate conditions involving Tepezza and Krystexxa.
David Sacks
Sacks won by making the remedy fit the alleged conduct. Unlike his broader anti-regulatory rhetoric elsewhere, this argument accepted enforcement and narrowed it to the mechanism that mattered.
Assumptions and fact checks
Prohibiting the named bundling tactics would be more proportionate than stopping the whole merger.
Why it mattersThe final consent order adopted that approach, added monitoring and prior-approval obligations, and allowed the transaction to close.
David Friedberg
Friedberg had the best factual diagnosis and forced everyone to argue the actual transaction. He lost only at the remedy stage, where Sacks's narrower answer proved sufficient.
Assumptions and fact checks
The credible bundling theory required blocking the acquisition rather than imposing a monitored conduct remedy.
Why it mattersThe eventual settlement addressed the theory with conduct restrictions, prior approval, reporting, and a monitor while permitting the deal.
Horizon generated about $4 billion of annual revenue and roughly $1 billion to $1.5 billion of EBITDA before the acquisition.
CheckHorizon reported $3.629 billion of 2022 net sales and guided to roughly $1.32 billion to $1.34 billion of adjusted EBITDA, matching Friedberg's rounded description.
The FTC's theory was that Amgen could use portfolio bundling and rebate terms to entrench Tepezza and Krystexxa and disadvantage rivals.
CheckThat was the agency's stated theory, and the final order prohibited bundling, conditional rebates, and terms that exclude competing products.
Chamath found a governable choke point and treated catastrophic risk with the right seriousness. He weakened it by using civilizational language before defining the capability test that would trigger oversight.