AIAM

EPISODE 2026-06-17

AI:AM LIVE — June 17, 2026 — Math, Biosecurity, and World Models: Carina Hong, Doni Bloomfield, Sam Pasupalak

The open tracked the model layer commoditizing — OpenAI reportedly dropping below 50% share, the AI buildout outrunning cash flow, and AI starting to run physical research labs. Then three guests on AI's hard problems: Carina Hong of Axiom Math on formally verified mathematical AI; Fordham law professor Doni Bloomfield on whether export-control law has become America's de facto AI licensing regime; and Skyfall AI's Sam Pasupalak on enterprise world models as the answer to what comes after LLMs.

𝕏 Live broadcast

Wednesday's show paired a market-and-infrastructure open — commoditizing models, the AI buildout's financing wall, and AI agents running physical research — with three guests working on AI's hard problems: formally verified math, the law of AI risk and biosecurity, and enterprise world models.

Note: this record is published from the show plan reconciled against the live broadcast's actual timings. Per-segment timestamps, deep-links, and the full as-aired recap will be added once the recording posts.

Episode timeline

  1. --:--Opening30 min plannedOpening: AI's Hard ProblemsThree threads from the morning's feed: the model layer commoditizing as OpenAI's share slips and Google surges, a warning that the AI buildout can't fund itself much longer, and AI agents starting to run research in the physical world. (Per-segment timestamps and the full as-aired recap will be added once the recording posts.)

    OpenAI drops below 50% — is the model layer commoditizing? The claim going around: OpenAI's share fell below 50% for the first time, with Google eating in because regular users can't tell ChatGPT from Gemini and Google owns their ecosystem. If the model is a commodity, distribution wins, not benchmarks.

    The end of the self-funded AI buildout? Epoch AI flagged that hyperscaler cash capex is growing far faster than cash inflows — and on current trends they can't fund the buildout from operations by year-end. The moment the buildout stops being self-financed: prelude to a bigger boom, or first crack in the bubble.

    AI runs the lab now — robot researchers and wet-lab-validated design. Two announcements pushed AI past 'research on a screen': NVIDIA's ENPIRE handed eight Codex agents a fleet of robots, GPUs, and a token budget to run physical experiments, and Boltz shipped protein and small-molecule design models with real wet-lab validation and an agent-callable API.

  2. --:--Interview25 min plannedMathematical Superintelligence: Can Proofs Make AI Reliable? — Carina HongCarina HongCarina Hong is the founder and CEO of Axiom Math, building 'Verified AI' — models that produce formally checkable proofs in Lean. Fresh off a $200M round, she argues formal verification is the missing reliability layer for non-deterministic AI.

    We pressed on where formal verification genuinely extends AI capability versus where the specification bottleneck stays irreducibly human, and put the sharpest public critique to her directly — that the system doesn't eliminate hallucinations so much as launder them through a compiler if the natural-language-to-Lean spec is wrong. From there: whether autonomous proving generalizes to inventing new mathematics or only executes existing math faster, why a sub-million-parameter model can beat frontier LLMs on structured reasoning, and her claim that Verified AI accelerates superintelligence.

    Axiom Math ↗Carina Hong on X ↗

  3. --:--Interview25 min plannedBiosecurity and AI: Law as a Risk Control System — Doni BloomfieldDoni BloomfieldDoni Bloomfield is an associate professor of law at Fordham, working at the intersection of AI, biosecurity, and export-control law. He wrote one of the first papers on whether the government can legally treat model weights as an export.

    With the government pulling a frontier model over national security, we asked whether export-control law has quietly become America's de facto AI licensing regime — and whether it holds up legally given the First Amendment problems with controlling model weights and outputs. We also dug into his counterintuitive thesis that export controls actively undermine biosecurity by blocking timely safety evaluations, his Biosecurity Data Level framework for gating dangerous datasets, and whether chip export controls have helped or hurt the U.S. position.

    Doni Bloomfield — Faculty page (Fordham Law) ↗Biological data governance in an age of AI (Science, 2026) ↗

  4. --:--Interview25 min plannedEnterprise World Models: What Comes After LLMs? — Sam PasupalakSam PasupalakSam Pasupalak co-founded Maluuba (acquired by Microsoft in 2017), stepped away through the transformer wave, and is back building Skyfall AI. His contrarian bet: not LLM-agents, but an 'enterprise world model' grounded in how a specific company actually works.

    We asked where pure LLM-agents break in production and what has to be structurally true to fix it — grounded in Skyfall's own benchmark showing frontier models fail to predict dynamic enterprise state. From there: what's actually inside an 'enterprise world model' versus a knowledge-graph-plus-agents stack, whether the moat holds against Microsoft and ServiceNow's data gravity, and what a deep-learning founder who sat out the scaling era thinks the LLM-native crowd is missing.

    Skyfall AI ↗

  5. --:--ClosingCloseThree of AI's hardest problems in one morning — reliable reasoning, governable risk, and reliable enterprise autonomy — each with a guest betting their company on a different answer.

The open — commoditization, capex, and AI in the lab

OpenAI reportedly slipping below 50% share as Google surges raised the question of whether the model layer is commoditizing and distribution now wins. Epoch AI's warning that the buildout can't fund itself from cash flow much longer set up the bubble-versus-debt-phase debate. And NVIDIA's ENPIRE and Boltz's wet-lab-validated design models showed AI agents moving into physical research.

Verified math and the law of AI risk — Carina Hong and Doni Bloomfield

Carina Hong made the case for 'Verified AI' — machine-checked proofs in Lean as the reliability layer for non-deterministic models — and faced the critique that the specification bottleneck stays human. Doni Bloomfield argued that export-control law is becoming a de facto AI licensing regime on shaky legal ground, and that the same controls can undermine the biosecurity they aim to protect.

After LLMs — Sam Pasupalak

Skyfall AI's Sam Pasupalak argued that reliable enterprise autonomy needs a grounded 'world model' of the organization rather than raw prompting, drawing on a benchmark showing frontier models fail to predict dynamic enterprise state — and on the perspective of a founder who built grounded language AI before the scaling era.