AI 2025 → 2026 Live Show | Part 2

NY Assemblymember Alex Boris discusses the RAISE Act, which targets catastrophic AI risks (100+ deaths, $1B+ damage from CBRN weapons or autonomous crimes). The bill is in chapter amendment negotiations with the governor, facing pressure from a16z-backed super PACs running attack ads. Boris argues the bill sets meaningful safety standards beyond California's SB 53 transparency requirements.

• •RAISE Act requires labs to have public safety plans, disclose critical incidents, and not release models that fail their own tests
• •Super PAC 'Leading the Future' ($100M from a16z, Greg Brockman) targeting Boris with ads, but he claims it's boosting his name recognition
• •Governor has until Saturday to sign or veto; negotiations focus on whether to match California's weaker SB 53 standards
• •Boris left Palantir in 2019 over ICE deportation contract disagreement, showing willingness to take principled stands
• •Attack ads may be designed to discourage other candidates from making AI regulation part of their platform

Former White House AI adviser Dean Ball maps emerging political coalitions: traditional AI safety advocates, pro-AI industry groups, and a new anti-AI populist coalition spanning left and right. He argues Trump personally cares about federal preemption and that the president will be a 'great man of history' figure shaping AI policy regardless of one's political views.

• •Three main factions: AI safety advocates, pro-AI groups (not just industry), and emerging anti-AI populists (Bernie Sanders to right-wing critics)
• •Key question: will middle groups (kid safety, consumer protection advocates) align with safety or anti-AI camps?
• •Trump's preemption push is personal conviction, not just tech industry influence - he consistently cares about this issue
• •Social media was an 'authority destroyer' - AI policy debates reflect fears of similar loss of institutional control
• •Copyright battles are really about protecting large IP portfolios (Disney, News Corp) not individual creators

Forecaster Peter Wildeford argues against selling H200 chips to China, proposing a 'rent don't sell' model using Malaysian data centers. He contends China's civil-military fusion means commercial chips enable military AI, and that China will push out NVIDIA once Huawei scales up, just as they did with Tesla and Apple.

• •China isn't a capitalist market - government will ensure Huawei has unlimited demand regardless of NVIDIA availability
• •Huawei currently supplies only 1-4% of Chinese demand; NVIDIA is 'useful idiot' filling the 96% gap temporarily
• •Rent-don't-sell via Oracle-controlled Malaysian data centers allows economic benefit while maintaining kill-switch capability
• •Military won't use foreign data centers (espionage risk), but commercial sector would - separates economic from military benefit
• •Historical pattern: Tesla, Apple, Uber all got pushed out for domestic alternatives once China developed capacity

Discussion of the Arc AGI-1 leaderboard showing 390x improvement in cost-efficiency over one year. Wildeford emphasizes that while capability increases get attention, the dramatic reduction in costs for existing capabilities is equally transformative - old frontier capabilities now run on laptops or phones.

• •Same capabilities that were barely possible are now achievable by open source models and consumer hardware
• •Cost reduction multiplied by capability increase creates compounding opportunity
• •Trend applies across domains: coding, math, computer use, legal analysis - all improving on similar trajectories
• •Enables new use cases through affordability even without capability breakthroughs
• •Open source models catching up to old frontier performance democratizes access

Analysis of why AI company revenues massively exceeded forecasts while benchmark scores came in slightly below expectations. Wildeford admits his mistake: company revenue projections use conservative economic models that can't account for 'unprecedented' AI progress, making them systematically underestimate growth.

• •Companies must use standard backward-looking economic models for projections - can't factor in unprecedented AI progress
• •Revenue explosion driven primarily by coding models (Claude Code, Cursor, Windsurf) becoming breakout use case
• •Expected 'year of the agent' for general computer use didn't materialize, but coding agents delivered massive value
• •Wildeford now rarely writes code himself - Claude Code is faster and better for first drafts
• •CFOs face different constraints than researchers - must maintain credibility with investors, can't promise 10x growth without signed customers

Wildeford shares his information diet for AI forecasting: Twitter remains primary source, supplemented by Zvi's aggregations, Epoch AI, Meter, and Semi Analysis. He emphasizes ignoring first 24 hours of model releases due to hype from early access recipients, and values organizations with track records of honest assessment.

• •Twitter primary source, supplemented by Understanding AI, Nathan Lambert's substack, Semi Analysis, Epoch AI, Meter
• •Ignore first 24 hours of model releases - early access goes to people who will write glowing reviews regardless
• •Values organizations focused on reliability metrics, not just capability benchmarks (Meter's work on consistency)
• •Professional domain experts (mathematicians, policy analysts) provide ground truth on real-world utility
• •Forecasting value chain: asking right questions and taking action matter more than just putting numbers on predictions

Wildeford predicts 2026 will see AI capable of 1-2 days of autonomous human labor (vs current 2 hours), with reliable computer use becoming transformative. He expects this to create a 'ChatGPT moment' for agents, viscerally demonstrating unemployment potential and triggering political backlash heading into midterms, with populist attacks on data centers intensifying.

• •Meter evaluation: AI currently 50% reliable on 2-hour tasks, could reach 1-2 day tasks by end of 2026
• •Computer use capabilities improving across all domains on similar trajectories, just from lower starting points
• •Even inefficient agents (5 tries to succeed once) valuable if cheap and can run while you sleep
• •Waymo and robotaxis becoming commonplace, creating visceral sense of automation impact
• •Political ramifications: populist backlash against data centers, potential measurable unemployment tick-up, midterm election dynamics
• •Repeating 'year of the agent' prediction after 2025 didn't deliver - but foundation is stronger now