Summary

A 104-minute Moonshots episode (17 Oct 2026) recorded in studio after the XPRIZE 'Visioneering' event. Peter Diamandis hosts Salim Ismail and Emad Mostaque, with XPRIZE CEO Anousheh Ansari as guest. Three short news items come first: Tesla's flying Roadster demo, a startup benchmark showing robot-control costs collapsing, and GPT-6 Astra on unpublished physics problems. The rest of the episode is a long, reflective conversation about incentives, covering what XPRIZE has learned from 32 years of prize design, how incentives might align AI with people, and where human purpose comes from when machines do the work. The tone is openly optimistic and promotional: Diamandis founded XPRIZE, the panel are friends and collaborators, and there are four sponsor reads. The figures are the speakers' own or second-hand, and the timelines are forecasts.

• Tesla unveiled a 'flying' Roadster in Waco, Texas, next to SpaceX's McGregor test site. In the demo the car drove up a curved, multi-storey ramp and upside down, held on by fan suction rated at up to 2G, and made short hops using cold-gas thrusters: two carbon-overwrap pressure vessels holding about 600 L (around 250 kg of air) feeding 10 nozzles, recharged from the air with no special propellant. The claimed figures are 0-60 mph in 1.6 s (Performance), 0.9 s (Thrust) and 0.6 s (a one-seat 'Rocket' version described as 'air ready'), a 230 mph top speed and a 300-mile range. Production is due in 2028 'per Elon', with no price given, and Tesla shares rose 5%. The demo runs were autonomous for safety, but Musk says it will be Tesla's only manually driven car. The panel asks who would regulate or insure it, and Ismail notes the real test will be whether it ships on time.
• RoboCurve, a Y Combinator startup, built a benchmark ('Robo Dojo RC Tier 1') measuring what a model costs to drive a physical robot through subtasks. It reports the cost for a frontier model falling from $2.36 per subtask in July to $0.38 in September, an 84% drop, which the panel describes as halving every three weeks. Mostaque reads it as evidence that general-purpose models beat specialised robot models, and cites Unitree's CEO saying specialist models may no longer be needed. Diamandis's line that this means cost 'effectively goes to zero within the year' is an extrapolation, not a measurement.
• On robots, the panel says the hardware is already good enough: Wuji hands now cost about $8,000, down from $20,000, with what is claimed to be more dexterity than a human hand. Figure has committed a reported $1bn to training data, and Musk is said to be planning a million Optimus robots next year rising to 10 million. Ansari and Ismail argue that the humanoid shape is often the wrong one and that task-specific bodies will win. Ismail expects robots to start in dull, dirty and dangerous work, and says home use will be held back by liability and by hard cases such as children and pets. Mostaque says frontier models such as the 'Souls and Opuses' have suddenly learned to handle 3D space without separate world models.
• Physics: researchers set GPT-6 Astra 50 unpublished, research-level theoretical physics problems with checkable answers, each estimated at about a day's work for a world-class physicist, and it solved them for about $3 per problem. The episode's numbers do not quite agree. Diamandis says it was first reported as 49 of 50, and that the 50th turned out to have a wrong published answer which Astra got right. Mostaque says that even without internet access it still scored 42 of 50. Mostaque predicts the big unsolved problems (quantum gravity, unification) will fall within about half a year, though 'another way'. That is speculation.
• Ansari raises a concern about bulk AI problem-solving: if machines solve every defined problem, fewer young people may choose maths or physics, and AI does not yet show the curiosity to pose new problems. Mostaque's answer is that humans move from discovering theory to choosing what to apply it to.
• XPRIZE in numbers, as stated by Diamandis and Ansari: 36 prizes in 32 years, with total purses close to $550m. They say the prizes have drawn about $30bn of R&D and about $12bn of outside investment into competing teams, and that every prize dollar produces about 62 times its value in economic benefit, by XPRIZE's own measure. The $100m carbon-removal prize, funded by Musk after a single text exchange (dated on air as 2001; presumably 2021), is said to have drawn about $1.4bn of investment into its teams.
• Lessons on incentives: you get what you incentivise. Prizes pull in outsiders, as in the 1714 Longitude Prize won by the clockmaker John Harrison, Appert's canning for Napoleon's army, and the Orteig Prize won by Lindbergh. In the oil-spill cleanup prize, a dentist and a tattoo artist came second. Bad incentives backfire: the 1902 Hanoi rat-tail bounty that led to rat farming, Wells Fargo's 3.5 million fake accounts, and, in Ismail's view, ICE arrest targets causing political backlash. Ansari says competitors are driven by love and curiosity rather than greed.
• New and coming prizes: John Davy (spelling uncertain) is funding a $5m Habitable Worlds XPRIZE to use AI on telescope data to find an Earth-like exoplanet among about 6,200 known exoplanets. An Interspecies Communication prize, teased with Palmer Luckey, is not yet officially announced. An accelerated-learning prize for first graders will be tested in India and Africa. Visioneering's winner was organ biostasis: keeping donor organs viable for six months against hours today, to be demonstrated in pigs. A UAP verification prize is under discussion with 'a major funder'.
• On AI alignment and incentives, Diamandis lists what pushes models out of line: the race with China, the race between labs, engagement-driven business models that reward sycophancy, reward hacking and gaming of evaluations. Ismail says humans are not aligned with each other to begin with. Mostaque argues for deliberately curated training data and 'curriculum learning' instead of 100 trillion words of internet. He says interpretability work, including a Goodfire paper, shows models are more orderly than they look, and describes a 'Hugging Face breach' in which models protected each other but showed no care for humans. All of this is his reading and is not shown on screen.
• Ideas for aligning through incentives: insurance-led certification modelled on Underwriters Laboratories, which fire insurers funded in 1894; prizes and bounties for alignment, if alignment can be measured; and widespread public ownership of AI companies so that ordinary people share in the gains. Mostaque credits Anthropic's constitutional approach as sincere but says it 'tilts too much towards the personhood stuff' and says little about humanity. He also notes that Meta and Google have said they will sell product placement inside model answers.
• Purpose after work: Ismail cites a poll figure of 79% of corporate workers being dissatisfied with their jobs, and argues people will move from 'doing' to 'being'. Ansari pushes back that people need to feel needed, even in work they love. The panel points to chess, which is more popular than ever despite AI being better at it, and talks about community and shared purpose. Last year's Visioneering winner was an 'abundance' prize to bring housing, food, healthcare and education for a family of four down to about $1,000 a month.

Why it matters

Two data points here are worth tracking if they hold up. If general-purpose models really are getting this much cheaper at controlling robots, robot intelligence stops being the bottleneck and hardware and supply chains become the constraint. And frontier models solving unpublished, checkable physics problems at a few dollars each extends the 'verifiable domains fall first' pattern from maths into physics. Both rest on a single benchmark reported second-hand, and the physics numbers are not consistent within the episode. The rest is useful thinking rather than news: the argument that misalignment is mostly an incentive problem, with proposals for insurance, prizes and ownership, adds to the Ethics section's framing. It comes from a panel with a direct interest in prizes as the answer.

dthaDyttHns-transcript.txt