Featuring
Spice rack
Did the reaction to Charlie Kirk's killing reveal a specifically left-wing assassination culture?
Original point: Kimmel's failure to correct his earlier framing avoided a necessary reckoning with a poisonous left-wing ideology that increasingly accepts political violence.
What everyone argued
Jason Calacanis
Jason rejects the jump from ugly posts and one killer to 'the left' as a strategy. He says the Democratic figures he knows condemned the murder, points to January 6 as clear right-wing political violence, and argues that both parties need leaders who lower the temperature.
David Sacks
Sacks says Kimmel effectively framed the shooter as MAGA, then replaced that error with a 'random nut theory' that hid ideological motivation. He cites large-scale celebration on TikTok, Bluesky, and Reddit plus polling he says shows roughly three times more acceptance of political violence on the left, arguing that this asymmetry must be confronted without both-sides framing.
Winner circle
Jason wins the evidentiary question. Sacks identifies real and ugly celebration, and Kimmel's wording plainly invited a MAGA interpretation, but a feed full of examples is not a denominator and an online fringe is not a party strategy. Representative polling available after the episode shows strong, nearly symmetric rejection of political violence, so the confident claim of a specifically left-wing assassination culture fails its burden of proof.
Commentary
Jason Calacanis
Assumptions and fact checks
Condemnation by Democratic leaders shows that political assassination is not a strategy of the political left.
Why it mattersA party-level strategy would require endorsement, organization, or tolerance from institutions and leaders, none of which Sacks established. Elite condemnation does not disprove a violent subculture, but it defeats the broader strategy claim.
Treating political violence as a bipartisan problem is more accurate and constructive than assigning it to one side.
Why it mattersThe record contains ideologically varied attacks and strong cross-party rejection of violence. Different periods and definitions can produce asymmetries, but a universal one-side diagnosis outruns the evidence presented here.
January 6 rioters beat police officers.
CheckThe Justice Department reported that approximately 140 officers were assaulted at the Capitol and documented numerous convictions for assaulting, resisting, or impeding officers.
All January 6 defendants were given pardons for that conduct.
CheckPresident Trump granted full pardons to most convicted defendants and ordered pending cases dismissed, but commuted the sentences of 14 named defendants rather than pardoning them. Jason's larger point about sweeping clemency is correct; 'all were pardoned' is not.
David Sacks
Sacks is right to reject the comforting fiction that ideology becomes irrelevant whenever an attacker is unstable. He loses the broader ruling because his strongest evidence establishes a disturbing fringe and a misleading Kimmel implication, not a representative left-wing assassination culture.
Assumptions and fact checks
The volume of celebratory posts visible on major platforms measures how common pro-assassination views are on the left.
Why it mattersViral content is selected for outrage and cannot supply a population denominator. It proves that celebration existed and spread, not how representative it was.
Calling the killer a disturbed individual necessarily strips ideology from the explanation.
Why it mattersMental disturbance and ideological motivation can coexist. Kimmel's phrasing could minimize ideology, but the categories are not mutually exclusive and the transcript does not establish the killer as representative of a larger movement.
Jimmy Kimmel said that Kirk's shooter was MAGA.
CheckKimmel did not make that direct declarative statement. He said the 'MAGA gang' was trying to characterize the killer as 'anything other than one of them,' wording that strongly implied MAGA affiliation and reasonably invited the interpretation Sacks gives it, but the literal claim is inaccurate.
Polling showed acceptance of political violence was three times greater on the left than on the right.
CheckNo identified survey in the episode supports this broad claim with a stable definition. A large September 2025 Harvard CAPS/Harris poll instead found similar overwhelming rejection of partisan violence among Democrats and Republicans, and later Pew polling found nearly equal concern about left- and right-wing extremism.
Should dramatic AI research papers change forecasts before they become working products?
Original point: If analog in-memory attention can deliver even part of its reported efficiency, sophisticated models could run locally and force a rethink of projected data-center and power demand.
What everyone argued
David Sacks
Sacks calls the hardware result 'a little too good to be true' and says he gives far more weight to products than papers because research may not reproduce or scale. He expects smaller local models, but still predicts large electricity demand as AI use and robotics expand.
David Friedberg
Friedberg highlights PDDL-Instruct's large symbolic-planning gains and an analog attention architecture reporting orders-of-magnitude speed and energy improvements. He repeatedly says 'if this scales,' but argues the work could move powerful inference to robots and other edge devices and radically lower infrastructure needs.
Winner circle
Sacks wins the forecasting standard, while Friedberg wins credit for spotting important work early. Both headline results are genuine, but PDDL accuracy on constrained domains does not establish general reasoning, and a huge gain in the attention block does not equal the same gain for a complete frontier service. Until manufacturability, system integration, and rebound demand are measured, product evidence is the sounder basis for changing power and data-center forecasts.
Commentary
David Sacks
Sacks wins the narrow forecasting burden by asking for a product, not by proving the papers unimportant. The strongest version of his case is that component gains can be real and still fail to reduce system-wide energy or aggregate demand.
Assumptions and fact checks
Product launches are a more reliable basis for infrastructure forecasts than striking paper benchmarks.
Why it mattersInfrastructure forecasts depend on manufacturability, yields, system integration, software, reliability, workload mix, and demand response—constraints a benchmark or circuit simulation cannot settle alone.
Efficiency gains will not prevent total AI electricity demand from rising because usage and robotics will expand.
Why it mattersRebound demand is plausible and historically common, but its size depends on price elasticity, deployment pace, grid constraints, and whether computation shifts from data centers to devices.
David Friedberg
Friedberg presents the research with unusually useful caveats, but then lets the headline component numbers carry a full-system implication. Naming which fraction of inference energy the design replaces—and what remains digital—would have made the forecast much tighter.
Assumptions and fact checks
The analog gain-cell architecture will scale economically from the demonstrated attention design to full, manufacturable frontier-model systems.
Why it mattersThe mechanism is credible and peer reviewed, but analog non-idealities, model mapping, precision, fabrication, integration, and the energy cost of non-attention operations remain material.
A large reduction in energy per token would reduce projected data-center electricity demand.
Why it mattersIt lowers the cost of a fixed workload, but cheaper inference can expand token demand and robotics adds new workloads. Per-operation savings do not determine aggregate consumption.
PDDL-Instruct achieved up to 94% planning accuracy, a 66 percentage-point absolute improvement over baseline models.
CheckThe paper reports up to 94% validity on standard symbolic-planning benchmarks and a 66-point absolute improvement over its untuned baselines. The maximum comes from constrained benchmark domains, not general reasoning across arbitrary tasks.
The analog attention architecture achieved roughly 100 times lower latency and 70,000 times lower energy use than an Nvidia H100 for the same attention outcome.
CheckThe published paper reports up to two orders of magnitude lower attention latency and four orders lower attention energy versus GPUs, including the cited H100 comparison. The result applies to attention computation in the proposed architecture, with text performance comparable to GPT-2; it is not a measured 70,000-fold reduction for an entire frontier LLM service.
Did the COVID censorship backlash create real safeguards against another speech crackdown?
Original point: Exposure of COVID-era platform pressure created a new sensitivity to government censorship and clearer boundaries that should make a repeat less likely.
What everyone argued
David Sacks
Sacks says there is little evidence the political left learned that lesson: California lawmakers passed SB 771, politicians pressured affiliates over Kimmel, and past platform-pressure disclosures show a persistent appetite to suppress disfavored speech. He predicts Newsom will sign the bill to satisfy a presidential primary base even if courts later strike it down.
David Friedberg
Friedberg says disclosures of Biden-era pressure made platforms and the public more alert to censorship, then raises SB 771 as a dangerous test of those boundaries. He worries vague civil-rights enforcement could chill lawful speech, hopes Newsom will veto the bill, and argues creators need transparent notices and reason codes when platforms restrict distribution.
Winner circle
Friedberg wins narrowly in hindsight. Sacks was right that censorship incentives persisted—the Legislature passed SB 771—but wrong that Newsom's national ambitions would compel him to sign it. The veto shows a real guardrail, while the bill's progress shows it is neither durable nor consensual; the lesson was learned by enough decision-makers to stop this measure, not by the whole political system.
Commentary
David Sacks
Sacks is strongest when he points to passage of the bill: backlash had not erased the demand for regulation. He loses ground by calling a civil-liability theory a generic hate-speech ban and by treating a speculative account of Newsom's base as more decisive than the governor's competing incentives.
Assumptions and fact checks
Passage of SB 771 proves that the political left learned nothing from COVID-era censorship controversies.
Why it mattersLegislative passage shows the restrictive impulse remained influential, but the bill targeted algorithmic participation in civil-rights violations rather than simply banning offensive viewpoints, and the governor's veto shows a meaningful internal check.
Newsom's presidential incentives made signing the bill more attractive than vetoing it.
Why it mattersThe observed decision went the other way. His incentives evidently included constitutional risk, technology-policy positioning, or broader-electorate concerns that outweighed any pressure from the bill's supporters.
The First Amendment has no general hate-speech exception.
CheckThe Supreme Court has repeatedly held that government may not suppress ideas merely because they are offensive; threats, incitement, discrimination, and other conduct can be regulated under separate doctrines, but 'hate speech' is not a freestanding exception.
Governor Newsom would sign SB 771.
CheckNewsom vetoed SB 771 on October 13, 2025.
David Friedberg
Friedberg earns the hindsight edge because he treats backlash as a possible check, not a guarantee, and the veto supplied a real check. His case would be stronger if he separated offensive speech from actionable threats, coercion, discrimination, and platform conduct instead of accepting the episode's 'hate speech law' shorthand.
Assumptions and fact checks
Public exposure of COVID-era platform pressure created durable resistance to future government censorship.
Why it mattersThe veto and greater platform transparency support some learning, but legislative passage and continuing pressure campaigns show no settled cross-party norm. One successful check is not a durable system.
Unclear algorithmic liability would push platforms to over-remove lawful content even if the statute formally targets civil-rights violations.
Why it mattersLarge per-violation penalties and uncertainty over aiding or algorithmic amplification can rationally induce over-compliance. The magnitude depends on judicial interpretation, enforcement practice, and available safe harbors.
Alphabet acknowledged that Biden administration officials repeatedly pressured it to remove some COVID content that did not violate YouTube policy.
CheckAlphabet's counsel told the House Judiciary Committee that senior administration officials conducted repeated and sustained outreach about user content, including content that did not violate YouTube policy, while also stating that YouTube made its own moderation decisions.
SB 771 would fine a social network merely for allowing content that California deems hate speech.
CheckThe enrolled bill authorized civil penalties when a large platform, through its algorithms, violated or aided specified existing civil-rights laws involving violence, intimidation, coercion, or discrimination. It did not create a general state list of banned words or impose liability merely because officials called content hateful, though its algorithmic-aiding theory could still create a chilling risk.
Jason wins by separating condemnation of real celebratory posts from collective guilt. He would have been stronger with representative polling in hand rather than 'everyone I know,' but his demand for a denominator and a consistent standard is the key analytical move.