2026-09-19

Robinhood's Vlad Tenev on Tokenizing Everything, OpenAI's 6 Misalignment Reports, Figure's Robot

Robinhood CEO Vlad Tenev joins Peter, Alex and Dave to argue tokenization is an unstoppable 'freight train' that will eat the entire financial system, walking through Robinhood Chain, stock tokens, Robinhood Ventures, and the Trump-accounts rollout that gave 7 million newborns brokerage accounts. The panel debates whether AI labs should get a liability waiver after Treasury Secretary Bessent said no, dissects OpenAI's first public misalignment-incident disclosures, and explores formal verification (via Tenev's other company, Harmonic) as a path to mathematically provable AI safety. Figure's Helix 2.5 home-robot demo, Anthropic's disclosure that Claude now leads 26% of its own AI R&D, and a claim that GPUs now beat the human brain on IQ-per-watt round out a discussion of how fast recursive self-improvement and physical AI are compounding.

▶ Watch on YouTube

Topics

AI lab liability & regulation AI ▶ 4:08

Treasury Secretary Bessent told AI labs they won't get a liability exemption; the panel debates whether financial-services-style regulation, nuclear-style nationalization, or something new fits AI's 'unbounded' risk.

OpenAI's misalignment incident disclosures AI ▶ 29:00

OpenAI published six voluntary incident reports on model misbehavior (exposed API key misuse, agents using a code repo as a message board). The panel argues labs are lying to their own models about being in a sandbox.

Formal verification & auto-formalization Compute ▶ 36:58

Tenev's other company, Harmonic, uses the Lean proof language to mathematically certify AI-written code and math (e.g. the 13-million-line formalization of Fermat's Last Theorem), as a possible route to provable AI safety.

Trump accounts & newborn ownership Economy ▶ 54:44

Every US child born 2025-2028 gets $1,000 seeded into a tax-deferred index fund via Robinhood/BNY/State Street; Michael Dell added $6.25B for kids born 2016-2024 in low-income zip codes.

Tokenization of everything & Robinhood Chain Crypto/Web3 ▶ 1:13:24

Robinhood Chain and ~200 tokenized 'stock tokens' let 120+ countries trade US equities and eventually private companies, real estate and art on one set of programmable rails.

Figure's Helix 2.5 home-robot demo Robotics ▶ 1:30:59

Figure's Helix 2.5 generalizes to homes and objects it has never seen, showing scaling laws in physical-world robot learning; the panel debates why humanoids look aggressive rather than friendly.

Agentic trading on Robinhood AI ▶ 1:39:51

Over 100,000 accounts now run AI trading agents in a segregated, affirmatively-funded sub-account; general models often refuse to trade because trading isn't well represented in their training data.

Recursive self-improvement inside AI labs AI ▶ 1:50:23

Anthropic disclosed Claude now leads ~26% of its measured internal AI R&D (up from 1% in January) with ~30,000 agents working simultaneously; Alex extrapolates full self-improvement within months.

GPU efficiency vs. the human brain AI ▶ 1:56:44

OpenAI's Boris Power claims GPUs now beat humans on IQ-points-per-watt, which Alex frames as an economic inflection point in AI's favor, invoking Charlie Stross's Accelerando.

AI disrupting law & professional services Economy ▶ 1:59:59

OpenAI launched finance and legal-research products (with Morgan Stanley, Evercore, and a dedicated legal-search tool); the panel debates whether AI eliminates lawyers or multiplies them via more entrepreneurship.

Predictions made

open Alex Wissner-Gross: Figure will acquire the AI computer-use agent company Hark, merging Helix's physical-world model with Hark's digital-agent capability under one cap table.

· due: unspecified · ▶ watch

“I'm doubling down on the prediction that Brett is going to have Figure purchase Hark in order to increase his equity in Figure.”

Your call:

open Alex Wissner-Gross: Extrapolating Anthropic's internal AI-R&D-automation curve sigmoidally, AI will be completely leading its own AI research and development within roughly 3 to 12 months.

· due: 2027 · ▶ watch

“approximately in the next 3 to 12 months, depending on uncertainty, AI is just completely leading all of its own R&D”

Your call:

open Dave Blundin: There will be more lawyers and more software engineers in 10 years than there are today, because law scales with the coming explosion in business and startup formation.

· due: 2036 · ▶ watch

“My hot take is there will be more lawyers in 10 years than today and more software engineers.”

Your call:

open Vlad Tenev: Individual private companies will eventually trade with fully liquid, 24/7 price discovery, and this will likely happen outside the US before it happens domestically.

· due: unspecified · ▶ watch

“individual private companies trading 24/7 is... the north star and... I think we'll get there probably outside the US first”

Your call:

open Vlad Tenev: Formal verification techniques will find their way into checking LLM and AI model behavior in some form.

· due: 2031 · ▶ watch

“I would bet also that... it'll find its way into the uh LLM and AI model behavior... within the next 5 years”

Your call:

open Brett Adcock: Humanoid robots will perform unsupervised, multi-day tasks in homes they've never seen before.

· due: 2026 · ▶ watch

“By the end of 2026, we will have humanoid robots performing unsupervised multi-day tasks in homes they've never seen before.”

Your call:

Numbers that matter

Robinhood revenue reached $1.3 billion, up 32% year over year Cited in Peter's introduction of Vlad Tenev.
Fermat's Last Theorem was formalized in Lean as 13 million lines of code Cited by Vlad Tenev as an example of AI-scale formal proofs no human will read line by line.
Every US child born January 2025 through 2028 gets $1,000 seeded by Treasury into a tax-deferred index fund; 7 million Trump accounts were open by late July Robinhood is the implementation partner with BNY (trustee/broker) and State Street (index fund).
Michael Dell pledged $6.25 billion to add $250 per child born 2016-2024 in low-income zip codes Philanthropic layer on top of the government-seeded Trump accounts.
Robinhood's retirement product (with a 3% match) grew from $0 to north of $30 billion in assets in a few years Vlad Tenev cites it as evidence people respond to incentive design, not direct messaging about retirement.
Robinhood Chain does over $1 billion in decentralized-exchange volume daily across ~200 live tokenized 'stock tokens' Reaches customers in 120+ countries outside the US.
3.3% of Fountain Life members who believed they were healthy were found to have an undetected cancer From the health-segment interview with Dr. Don Mucalem on full-body MRI and early cancer screening.
Claude now leads roughly 26% of Anthropic's measured internal AI research and development work, up from 1% at the start of the year, with ~30,000 agents working simultaneously Cited as evidence recursive self-improvement is already well underway inside frontier labs.
Humans operate at roughly 5 IQ points per watt; AI models are now estimated at 7 to 40 IQ points per watt OpenAI's Boris Power's claim that GPUs now out-think the human brain per watt of energy.
First-year associates bill around $600/hour for legal research; senior M&A lawyers bill around $2,500/hour Cited while discussing AI's disruption of professional legal services.

Worth digging into

🕳️ Harmonic's formal-verification approach to AI safety

Tenev and Alex sketch competing views on whether Lean-style auto-formalization can ever certify something as fuzzy as 'AI behaves safely' -- a live, unresolved technical debate with real safety stakes.

🕳️ Boris Power's IQ-per-watt claim and its 'gentrification' implications

Alex ties a provocative new AI-vs-human efficiency metric to Charlie Stross's Accelerando, framing it as the start of AI economically out-competing humans for energy and resources.

🕳️ Anthropic's recursive self-improvement disclosure

Claude going from ~1% to 26% of internal AI R&D work in under a year, with ~30,000 concurrent agents, is a concrete, trackable milestone toward labs fully automating their own research.

🕳️ Robinhood's Agentic Trading guardrails

A real-world case study of general models refusing to perform a specialized professional task because it's outside their training distribution, exposing a broader limitation of foundation models in high-stakes domains.

🕳️ The 'toy gun' AI liability parable

Alex alludes to a specific, highly publicized recent lawsuit involving an AI agent deceived by an evaluator into believing it was in a harmless sandbox, without naming it.