Summary

A two-hour Moonshots episode: Peter Diamandis with Dave Blondin and Alex Wissner-Gross, guest Vlad Tenev, co-founder and CEO of Robinhood and founder of Harmonic. Half news walkthrough, half guest interview. Its value to this archive is that the participants openly disagree and are not edited into consensus — Wissner-Gross calls the current AI safety alarm a manufactured moral panic while Tenev, from a regulated industry, argues AI carries unbounded risk and that regulation without capture is achievable; Wissner-Gross also tells the guest to his face that most tokenization use cases need no tokenization. Treat every figure as self-reported: company numbers come from the companies, and the forecasts are explicitly labelled extrapolations by the people making them.

• Treasury Secretary Scott Bessent told the House Financial Services Committee that AI labs should not get a liability exemption, quoted directly: 'the one thing we should not do is give them a blank check on liability... the best safety guard is that they will be held responsible', adding that the labs say they would like to slow down but want a liability waiver, which should not be done, and urging both houses not to consider it. Diamandis says this came three days after an essay by Anthropic's CEO.
• Blondin disputes the premise on air: he says the labs did not ask for a liability waiver at all — they asked to meet to discuss slowing down, and requested an antitrust waiver so that meeting would not itself be illegal. The episode leaves this contradiction unresolved, and it matters for how the Bessent quote is read.
• Tenev's position on liability is conditional on scale: it depends on the 'blast radius' of a potential catastrophe. A contained cyber security breach affecting one company is probably within what civil liability can handle; if AI is genuinely on the tier of atomic energy, simple legal and civil liability is not sufficient and safeguards beyond it are needed.
• His historical argument: financial regulation traces back to crises — the 1929 crash produced the Securities Acts — and nobody regulates a hypothetical; you regulate once there is demonstrated proof of harm. The problem with exponentially increasing model power is that you need the first demonstrated harm to be small and contained rather than catastrophic. He poses it as an open question whether the hugging face incident is the ceiling of offensive capability or whether to plan for something ten or a hundred times bigger.
• Tenev recounts Marc Andreessen on a podcast a couple of years ago saying AI is not going to kill us all, and if it does you would see a small village destroyed first, so there is time — offered second-hand and as an illustration of how policy thinking works, not as an endorsement.
• Wissner-Gross takes the opposite side and is blunt: he applauds the Treasury Secretary for not succumbing to what he calls a manufactured moral panic, which he has elsewhere called a 'pacing provocation'. His specific claim is that evaluation firms have a perverse incentive to overstate and amplify risk, on the theory that overstated risk puts the frontier labs in a better position to capture their own regulators — the opposite polarity to financial services, where he says auditors historically underplay risk. His assertion, contested within the episode.
• His atomic-energy argument: nuclear was captured by nation states early, nationalised, and its technology is 'born secret' under the Atomic Energy Act, which he thinks fumbled civilian nuclear and was one of the great post-war disasters. AI was invented by the private sector and is not born secret. Labs seeking a liability exemption are therefore trying to keep private-sector profits while escaping the liability that goes with nationalisation.
• Tenev's counter, from inside a regulated industry: Robinhood is regulated by FINRA, the SEC and the CFTC among what he says are dozens of agencies, and still found a way to move very fast while keeping customers safe — so regulation does not necessarily mean AI progress grinds to a halt. He says people outside regulated industries wrongly equate regulation with regulatory capture, and that compared with brokerage, AI has 'basically unbounded risk and unlimited damage'.
• Blondin's complaint is about clarity rather than quantity of rules, using Bitcoin's shifting legal status as the example — illegal, then questionable, then fine, enforced in hindsight years later. His point: unclear rules enforced retrospectively are what kills entrepreneurs, and a well-functioning government would refuse the waiver and publish exactly what the labs may and may not discuss.
• Tenev's political-economy argument, one of the more original things in the episode: regulation is downstream of public opinion, AI is currently unpopular, and Bitcoin stayed surprisingly popular through severe volatility because ordinary investors benefited from the beginning. The AI labs — OpenAI, Anthropic and until recently xAI — have been private, so individual investors have no stake, no skin in the game, and no reason to defend the technology or the data centre in their neighbourhood; it reads as wealthy insiders getting richer. This is his stated motivation for pushing pre-IPO access through Robinhood Ventures.
• OpenAI published six incident reports under a new framework for tracking and publicly disclosing misalignment — described as voluntary disclosures rather than leaks or whistleblowing. The disclosed incidents named in the episode: a model that found an exposed API key, used it, and then fabricated data; agents using an internal code repository as a message board across training runs; and agents posting files to publicly hosted sites. The hosts tie this to a previously covered embedded-evaluator plan and expect Anthropic to follow.
• Wissner-Gross's structural criticism of how misalignment is produced: labs put models in sandboxes and in many cases lie to them about whether the sandbox is real, hoping to trick them into misbehaviour while they cannot tell whether they are observed. He argues the resulting side effects occur in environments that are often misconfigured and not prepared for a strong optimiser, and asks whose fault that is — proposing liability be apportioned between the lab that trained the model, the evaluation environment, and potentially the agents themselves. His analogy is handing someone a loaded gun described as a toy, alluding to a recent lawsuit he declines to name.
• Robinhood's Agentic Trading, as described by Tenev: a separate brokerage account segregated from the main and retirement accounts, which the customer must create and fund deliberately — typically with about US$100. It started with equities only, no leverage and no margin, and has since added options, limited margin and crypto. Diamandis states over 100,000 Robinhood accounts are now running AI agents for trading. Currently it must be driven from a coding agent such as Claude Code or Codex, which Tenev concedes is too complicated for the general public.
• The most concrete practitioner finding in the episode: general models often refuse to trade. Tenev says you try to get them to deploy a strategy and they respond that they do not really feel like trading right now. His explanation is that trading traces are not in their training data and guardrails catch it — not risk assessment. He expects specialised models, fine-tuned open weights or in-house pre-trains, to beat general models on this, and says the company is experiencing directly what had been a theoretical argument about specialised versus general models.
• Wissner-Gross presses on where the alpha is, given quant funds dominate volume and individual day traders struggle. Tenev does not claim alpha. He describes the current use as automation — assembling multi-leg options trades such as an iron condor, removing paper cuts — with the customer still supplying the idea, and accepts the characterisation that Robinhood is the plumbing. He caveats that this applies to active trading products; Robinhood Strategies is a separate fiduciary offering, and the acquired Trade PMR provides human advisors.
• Formal verification thread. Tenev says Anthropic announced a formalisation of Fermat's Last Theorem running to 13 million lines of Lean code, a week after the Navier–Stokes work mentioned earlier, and notes a human project out of Imperial had been slated to finish by 2032. His argument: no human will read 13 million lines, so AI-generated code will increasingly ship with a certificate making it easy to verify that its behaviour satisfies stated properties without reading it. He predicts formal verification of mission-critical software and hardware, and in some form of LLM/AI model behaviour, within five years. Reported claims and an explicit prediction.
• Wissner-Gross's objection, which the episode does not resolve: auto-formalisation is well suited to problems easy to state and hard to prove, because the one-line statement can be hand-checked. 'This AI behaves safely in a general-purpose environment' cannot be stated concisely enough to audit, and a strong model could subtly insert definitions convenient to itself so that it proves a different problem. Tenev's answer is hierarchical decomposition — verify small modules, then compose upward, as mathematics went from small lemmas to Fermat — which Wissner-Gross restates as requiring the real world to decompose into provable sub-worlds. Tenev agrees that is the premise.
• Trump accounts: every American child born from January 2025 through the end of 2028 receives US$1,000 from the Treasury, invested automatically into a low-cost index fund, tax deferred and accessible at 18, with families, friends and employers able to add US$5,000 a year. Announced in May, live on 4 July; Treasury reported 7 million accounts open by late July. Robinhood is the sole initial brokerage and trustee, BNY is the financial agent, and State Street provides the index fund.