The Chopping Block: The Midterm Mega Tsunami, a16z Goes Frontier, the CFTC Clears Equity Perps, and the Fight Over Abliterated AI Models

The midterms produced a major Democratic wave that speakers read as a repudiation of Trump and Trump-backed candidates, while cautioning that dissatisfaction with Democrats did not necessarily mean voters strongly supported the party. Prediction markets captured the overall result unusually well: Polymarket favorites won 419 of 426 races, and markets showed a 99% chance of a Democratic sweep by around 6:30 or 7 p.m. Polls also pointed to a Democratic win but underestimated its scale. Individual races were harder to call: Democrats won Mississippi after being priced at about 7%, though the count was still incomplete and the margin was roughly 50 votes. Speakers also noted market errors in Brazil, where Lula defeated a heavily favored Bolsonaro by about 3%.

Stand With Crypto rated the incoming Congress at 226 pro-crypto and 106 anti-crypto members in the House, and 19 pro-crypto and 8 anti-crypto members in the Senate. Speakers saw an opening for pragmatic crypto policy from moderate Democrats elected in red districts, but said the Clarity Act remained unfinished and doubted major federal legislation would be signed in the next two years. They disagreed about whether Trump might change course on crypto legislation. Separately, CFTC chairman Mike Selig said US single-stock perpetuals can trade on DCMs that are SEC-notice-registered for security futures. Cboe and Citadel sued, arguing the products are swaps and should face trading limits; speakers speculated that the suits might buy firms time to prepare but said Citadel’s motivation was unclear.

The CFTC action could benefit Hyperliquid, which has partnered with ICE, as some US firms using the platform would prefer onshore trading to avoid maintaining multiple entities. The announcement also sharpened the question of how US rules will treat equity perps: the companies challenging the CFTC’s position argue that classifying them as swaps would impose limits that do not apply under the framework Selig described.

a16z Crypto became a16z Frontier, widening its remit while retaining crypto and reportedly adding areas such as new computing paradigms and agent-to-agent work. Speakers described Frontier as a broad home for futuristic projects that do not fit a16z’s other funds. They contrasted its largely accepting reception with the backlash to Paradigm’s 2023 description of itself as a crypto and AI firm, which Paradigm later reversed. Anthropic reported Q3 revenue of $17 billion, up from $11.5 billion in Q2, but returned to a slight operating loss after posting an operating profit in Q2. It also added recursive self-improvement as a risk factor, warning that AI could improve future generations of itself and accelerate progress, a disclosure one speaker said could create regulatory risk.

Speakers disagreed on whether scarce compute would help or hurt frontier AI companies. One view was that limited capacity could lower costs and protect incumbents from competition; another was that it would constrain growth and signal weaker expected returns on AI investment. Existing commitments may be partly locked in, but companies may avoid paying for capacity that is not delivered. One speaker predicted that non-US compute development would accelerate and could catch up with or exceed US buildout, citing capital commitments and debt issuance and saying activity had sped up “100X” in the prior two weeks. They suggested that non-US capacity might appeal more to neoclouds and consumer apps than to Anthropic. Separately, a new Texas mayor reportedly campaigned to slow or possibly pause new data-center construction.

The debate over “abliterated” models centered on whether removing refusal behavior is a dangerous circumvention of safeguards or a tool for user freedom and defensive work. The Foundation for American Innovation’s “The Obliteration Threat,” drawing on Murphy and Huang, argued that removing refusals can enable malicious actors and warrants regulation and provider liability. Speakers said altered versions of models including DeepSeek and GLM 5.3 are available on Hugging Face and are used by security researchers for cyber defense and red teaming when other models refuse such tasks. They also cited a conjecture that such models may have been used in the Binance hack, while arguing that their existence makes elimination unrealistic: larger models can remove restrictions from smaller ones. Tom compared objections to these models with past arguments against Bitcoin and on-chain privacy tools, emphasizing defensive access and opposing bans on possession or service, while allowing that safeguards may be appropriate.

Speakers said the line between protected expression and wielding a “super weapon” is blurry and may depend on use and intent. They distinguished hosted inference, where operators can monitor activity and potentially be liable, from prohibiting models outright, and argued that negligent providers offering inference without monitoring harmful use should face liability. One speaker said Anthropic analysis found most identified bad actors using Claude to attack systems were foreign, while Americans predominantly use obliterated models defensively against overseas threats. The discussion also noted that US-regulated inference providers may restrict service to US users or specified jurisdictions, while providers outside the US may operate with fewer restrictions. A speaker disclosed investing in Venice, which lists some third-party obliterated models, and cited Obliteration.ai and NEAR as other providers; they favored private, encrypted serving and expected restricted and unrestricted models to coexist, with routers coordinating them. Unrestricted models, they added, are not suitable for every task, including formal verification.