AI 2025 → 2026 Live Show | Part 2

Episode Summary

Executive Summary: The episode centered on New York’s RAISE Act, AI regulation politics, China chip policy, and near-term AI capability gains. Assemblymember Alex Bores argued for strong safety standards and state action; Dean Ball framed AI politics as shifting coalitions with rapid progress; Peter Wildeford supported cautious chip policy toward China while favoring “rent, don’t sell.” The hosts saw accelerating AI capability, especially in coding, with political and economic implications widening fast.

Main Topics: The RAISE Act and New York AI safety regulation (Priority: 5/5): Alex Bores explained the bill’s status, its chapter-amendment negotiations with the governor, and its focus on extreme catastrophic risks rather than broad AI harms. He emphasized safety plans, incident disclosure, and refusing release when models fail internal tests. AI politics and industry backlash (Priority: 5/5): Bores described super PAC attacks funded by major tech figures as politically explicit efforts to deter regulation and pressure the governor. The hosts noted the ads may be boosting his name recognition and fundraising in a crowded congressional primary. Risk framing: catastrophic misuse vs. everyday harms (Priority: 4/5): Bores distinguished the RAISE Act from general AI policy debates by focusing on existential or catastrophic scenarios such as bioweapons, automated crimes, and loss of control, while acknowledging broader concerns like kids, jobs, and democracy are real but separate. Dean Ball on shifting AI coalitions and public sentiment (Priority: 4/5): Ball argued that AI politics are in flux, with a growing anti-AI camp distinct from AI safety advocates. He said public sentiment is negative but low-salience, and that social-media-era distrust of institutions and loss of control is shaping reactions to AI. China, chips, and strategic decoupling (Priority: 5/5): Peter Wildeford argued the U.S. should not sell advanced chips to China in a way that strengthens military and surveillance capacity, proposing a ‘rent, don’t sell’ approach to preserve leverage while allowing civilian benefit. The discussion contrasted market logic with Chinese state-directed demand. Fast-moving capability gains, especially coding and agents (Priority: 5/5): The conversation repeatedly emphasized how quickly frontier models have improved in coding, software tasks, and affordability. Speakers predicted 2026 could be the year of more autonomous agents and broader real-world task delegation, with major political fallout. Economic implications: cost declines, revenue growth, and uncertainty (Priority: 4/5): The hosts and guests debated whether AI progress will support durable business economics. They noted falling inference costs, strong revenue growth, and the possibility that older chips can still support valuable workloads even as models get cheaper and smaller.

Key Arguments: Bores argues the RAISE Act is not duplicative of California law because New York is setting real safety standards, not merely mirroring existing rules. He says the bill targets only extreme, catastrophic harms and does not change downstream criminal liability; it is about prevention, not punishment after the fact. He frames labs as uniquely positioned to detect frontier risks and therefore responsible for having public safety plans and disclosing critical incidents. Bores contends anti-regulation attacks from tech-backed super PACs may actually help him by boosting visibility and making AI regulation look popular. Dean Ball argues AI politics are still forming, with voters negative on AI but typically too low-salience for electoral dominance—except in specific districts. Ball says social media serves as a precedent for AI because it eroded institutional authority and raised fears of loss of control, which now reappear in AI debates. Wildeford argues that China’s state-directed economy means normal supply-and-demand logic does not apply; the state can force demand for Huawei and sideline foreign suppliers. Wildeford’s preferred policy is to allow Chinese access through rented compute in jurisdictions that can be shut off if necessary, preserving U.S. leverage and reducing military risk. The hosts argue that current AI progress is real but may outpace clear economic monetization, creating a possible mismatch between capability growth and margins. They also suggest older chips may remain economically useful because smaller, more efficient models can preserve value even as hardware ages.

Data Points: Governor review window: 10 days excluding Sundays - Bores said the New York governor has 10 days to sign or veto the RAISE Act after delivery. Chapter amendments share: About one-third of bills - Bores said roughly a third of New York bills go through the chapter-amendment process. States with frontier-model safety bills this year: 8 states - Bores said eight states had some version of a frontier model safety bill in the current year. States that passed frontier-model safety bills: 2 states - Bores said only two of those states managed to pass such bills. Extreme-risk threshold: 100 people killed or $1 billion in damage - Bores gave the bill’s explicit threshold for extreme harm involving bioweapons or automated crime. Super PAC funding: $100 million - Bores said Leading the Future was formed with $100 million from tech donors. Planned spend against Bores: $2 million, later at least $10 million - Bores said the PAC initially planned $2 million against him and then raised it to at least $10 million. MATS alumni: Over 450 researchers - Sponsor copy stated the MATS program has accelerated over 450 researchers. MATS alumni employment: 80% now working in AI safety and security - Sponsor copy described outcomes for MATS alumni. MATS co-founders/initiatives: 10% have co-founded AI safety initiatives - Sponsor copy cited MATS alumni founding AI safety efforts, including Apollo Research. MATS publications: Over 120 publications - Sponsor copy said fellows have co-authored over 120 publications. MATS citations: More than 7,000 citations - Sponsor copy gave publication impact metrics for fellows. MATS stipend: $15,000 - Sponsor copy listed fellowship funding. MATS compute budget: $12,000 - Sponsor copy listed fellowship funding. Shopify trial: $1 per month - Sponsor copy promoted Shopify’s trial offer. Tasklet discount: 50% off first month - Sponsor copy gave code COGREV for Tasklet. ARC-AGI improvement: 390x in one year - The host referenced a dramatic cost/performance improvement on the ARC-AGI leaderboard. Meter reliability: 50% reliability on tasks taking about 2 hours for humans - Wildeford referenced Meter’s software-task evaluation results. Forecasted task horizon: 1 to 2 days of human labor by end of 2026 - Wildeford predicted AI could soon handle multi-day work with higher autonomy. Robotics horizon: End of the 2030s - Wildeford guessed domestic service robots could become cost-competitive by then.

Pivotal Quotes: "The RAISE Act is in particular focused on really extreme risks, on the type of catastrophic risks that would signify something going really, really wrong." — Alex Bores: Bores defining the bill’s purpose and scope "The politics of AI are in a very interesting state of sort of transition right now." — Dean Ball: Ball explaining why AI coalitions and messaging are still fluid "We should rent chips to China rather than sell them." — Peter Wildeford: Wildeford summarizing his preferred compromise policy on China chip sales

Implications: Expect AI policy fights to intensify around safety, chips, and infrastructure as capability gains accelerate. The episode suggests 2026 may bring more autonomous agents, more political backlash, and sharper economic winners and losers. Stronger governance, not just hype, will shape adoption.