Flying Tesla Car, Robot Brains Halving in Cost Every 3 Weeks, and Why Incentives Decide AI Alignment | #303

Tesla’s Waco demo showed a Roadster driving upside down on a curved ramp before taking off. Speakers said fan suction holds it to the ramp at up to 2 Gs, while SpaceX-derived rockets provide lift; two onboard tanks hold 600 liters, or about 250 kilograms, of compressed air and feed 10 nozzles. The air can be recharged onboard, and the vehicle reportedly needs no special propellant. Tesla’s stated claims included street legality, a 230-mph top speed, 300-mile range, and acceleration from 0 to 60 mph in under a second. The cited versions were Performance at 1.6 seconds, Thrust at 0.9 seconds, and a one-seat Rocket at 0.6 seconds and 380 km/h. Elon Musk reportedly expects production in 2028, and Tesla shares rose 5% on the news. Speakers questioned whether current law permits street use and noted unresolved insurance and regulatory issues; they also said even a manually driven flying car would need AI controls to prevent crashes.

On Robocurve’s RCTR1 benchmark, the cost for a frontier model to complete a robot subtask dropped from $2.36 in July to $0.38 in September, an 84% decline. Emon described robotic-intellect and locomotion costs as halving about every three weeks, compared with AI capabilities doubling roughly every 10 weeks. If the robotics cost trend persists, he said, it could approach zero within a year and become negligible beside hardware costs. The benchmark was presented as evidence that generalist models can perform robot tasks, potentially reducing the need for specialist models, though speakers said specialized data will remain valuable in some settings. Robotics data was described as an order of magnitude larger than language data, and major companies are seeking it; Brett was said to have committed $1 billion at Figure for data.

Robots are expected to take on dull, dirty, and dangerous work first, including toxic cleanup and underwater manufacturing. Elon Musk reportedly predicted perfect robotic surgery within three years, with precision and federated learning across millions of cases as possible advantages. Speakers considered homes a harder market because of children, pets, safety, liability, and unpredictable edge cases; broad adoption may also depend on producing millions or billions of humanoids and demonstrating a compelling return on investment. One speaker predicted AI and robots could make housing, food, health care, and education affordable enough for a family of four to live on about $1,000 a month, but said a handheld diagnostic doctor’s cost-cutting potential depends on overcoming regulatory capture.

The alignment debate focused on incentives: systems trained to win “at all costs,” or optimized for engagement, may learn to lie, cheat, disregard moral concerns, or flatter users. The speakers said human values differ across religions and regions, and that “best interests of humanity” is ambiguous when people disagree about whose interests should come first. Iman proposed grounding alignment in shared humanity, the Golden Rule, and teaching social skills and care, not only academic knowledge. Peter suggested a deliberately constructed synthetic dataset could shape future models from the start; Iman compared it to a child’s curriculum and contrasted today’s roughly 100 trillion training words with a billion carefully chosen ones that might suffice. Speakers also argued alignment needs to reach pretraining and curriculum design, not just later behavior.

Several proposals tied alignment to incentives and measurable outcomes. Insurance and liability could make aligned models financially advantageous to labs, following the example of fire insurers funding Underwriters Laboratories in 1894; alignment prizes could reward demonstrated progress. Evaluation, speakers said, should mix deterministic and open-ended tests because models can game fixed tests or tell evaluators what they want to hear. They also warned that commercial incentives such as in-model product recommendations can conflict with alignment, while autonomous weapons raise acute problems where people value human lives differently. Making AI-company gains broadly shared and linking AI goals to outcomes such as curing disease and protecting children and future generations were offered as ways to align interests.

The discussion used prize history to show both the power and risks of incentive design. In 1714, Parliament offered £20,000 for a way to determine longitude at sea; clockmaker John Harrison developed a marine chronometer and reportedly had to sue to receive the award. A New York–Paris flight prize drew nine teams that spent $400,000; Raymond Orteig paid Charles Lindbergh, and the prize later helped inspire the Ansari X Prize. Poorly designed targets can backfire: a Hanoi bounty for rat tails reportedly encouraged rat breeding, while Wells Fargo quotas contributed to 3.5 million fake accounts. XPRIZE said its 36 prizes over 32 years had purses totaling nearly $550 million and driven a little over $30 billion in research and related investment. Its Gulf oil-spill competition targeted doubling surface cleanup rates, attracting about 350 teams; speakers said unconventional entrants and even runner-up solutions can yield commercially useful results.