AI:AM

Today on AI:AM — “Fable, AI Safety and Julius.

We open on a frontier-model launch day and what it changes: the debate over benchmarks reported with a fallback to a second model, the production guardrails that route sensitive work elsewhere, the compute-cost advantage of booking capacity early, and why the frontier increasingly looks like a two-actor race with the rest playing catch-up.

Geoffrey Irving & Daniel Murfet (Sequent) on their new alignment-theory organization — why they put superintelligence two to three years out, why verification looks defense-dominant, the argument that alignment is “not on track” despite models behaving well so far, what the benevolent-basin hope gets right and wrong, and why character training still lacks a real theory. ([@danielmurfet](https://x.com/danielmurfet))

Rahul Sonwalkar (Julius) on agentic data analysis — why the harness has to evolve alongside the model, the difference between token-maxing and results-maxing, the shift from tasks to goals, and a future where agents become first-class users of the internet, transact through agentic payments, and compete to be “hired” by the core agent.

We close on why robotics is the next domino, the “gas chromatograph” spread of who gets model access and when, the Glean Work AI Index’s bot-sitting and bot-shitting, and why “preciousness” about putting your own name on work may be turning into a liability.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit briefing.ai-in-the-am.com

Today on AI:AM — “Build, Measure, Heal: AI’s Three Frontiers.”

We open on the AI CEOs’ call to make DNA-synthesis screening law and what cheap intelligence does to biosecurity, then OpenAI’s new Sites product and the platform playbook of absorbing the app layer, the data-center and chip land grab, and why a fresh open model from NVIDIA still trails Anthropic’s best by a wide margin.

Hooman Radfar (Collective) on building the autonomous finance department for America’s 30 million solopreneurs — how one bookkeeper now supports 250 clients, why the app layer can still defend its margins against the frontier labs, why “Anthropic is like a drug dealer” on token costs, and why the real thing to regulate is the model arms race.

Taras Pohrebniak (Elomia Health) on agentic AI for mental health — an architecture that spends most of its compute on safety, why the company deliberately avoids hyper-realistic voice, where the regulatory line sits between a “friend” app and a medical device, and what they learned deploying in US prisons and on Ukraine’s front lines.

Peter Jansen (Ai2) on whether AI can actually do science — the leap from fourth-grade science benchmarks to the Theorizer project, why evaluating machine-generated theories is the real bottleneck, the cautionary tale of a “discovery” that turned out to analyze a random-number generator, and why benchmarks like ScienceWorld still break the best models.

We close on the inversion of the scientific method into a data-first discipline, what interpretability could add, and why biology’s data scarcity — not algorithms — may be the binding constraint on curing disease.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit briefing.ai-in-the-am.com

Today on AI:AM — “AI Security and Real-Time Content Safety.”

We open on Trump’s AI executive order — the polite 30-day model-review ask, classified benchmarks, and the state-vs-federal scramble where JB Pritzker has become the leading anti-AI candidate. Plus why the frontier labs seem calmer about regulation, and why the EO might actually trigger a security-review slowdown.

Tal Hoffman & Yanir Tsarimi (EnclaveAI) on finding the bugs that actually matter — how they reproduced an Anthropic Mythos-class finding with a model ~100x smaller, why proven exploitability is the real bottleneck, how AI-generated bug reports broke the bounty system, and why cheaper models plus the right harness can beat frontier models on security.

Brett Levenson (Moonbounce) on real-time content safety — lessons from running moderation at Meta scale, how a policy engine decomposes fuzzy rules like “hate speech” into atomic questions a hundred people would answer the same way, why prevention beats post-hoc moderation, and how payment providers quietly became the real legislators.

We close on the hardest open question — how low-level verified parts aggregate into trustworthy high-level behavior — plus the schlep and heuristics that end every AI vertical, freedom of speech versus freedom of distribution, and why “nobody got fired for buying Mythos” may drive enterprise security budgets.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit briefing.ai-in-the-am.com

Today on AI:AM — “Self-Improving Tax Agents and Catholic AI.”

We open on Google’s first equity raise since its IPO — Berkshire taking 12.5% of an $80B round — and what the scramble for capital says about an AI “megacorp” that may be too big to fail, plus the Bernie Sanders national-stake debate and why taxes may beat equity.

Arthur Fernandes Araujo & John de Wasseige (OpenAI) on self-improving tax agents — how a production tax workflow turned every human correction into training signal, took one accountant from 180 hours to 15, and why “the model eats the harness” with each new generation.

A hosts-only research speed-run on whether AI can still be watched — field notes from the Recursive event where monitoring is the number-one safety bet, plus the papers behind persona selection, emergent misalignment (”writing bad code makes you evil”), eval-gaming, and accidental chain-of-thought training.

Matthew Sanders (Longbeard / Magisterium AI) on Catholic AI after the Pope’s encyclical — what it was like at the Vatican, the divergence with Anthropic on machine consciousness, the red line on autonomous weapons, and why the last 5% of alignment is non-negotiable for a faith tradition.

We close on a live test of the cigarette-business refusal example from the OpenAI model spec, the tension between research and business “layers” in deployed models, and the argument that open-source AI may now be unbannable on religious-freedom grounds.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit briefing.ai-in-the-am.com

Trust and Recovery in AI. Andy Fernandez (HYCU) on backups as the enterprise black box, David Villalón & Manuel Romero (Maisa) on auditable digital workers, and Snehal Antani (Horizon3.ai) on the 77-second breach and why deception is the EMP of AI cyber weapons. Plus the disk trade and 1,766 miles of full self-driving.

(0:00) Opening — Why a daily AI show

(0:31) Why we need more AI shows

(11:12) What are we trying to do here

(17:23) Andy Fernandez (HYCU) — AI, cyber resilience & recoverability

(23:23) Backup data answers what was deleted

(27:14) Telemetry won't replace recovery

(43:14) Trust requires guardrails, not blind access

(48:21) David Villalón & Manuel Romero (Maisa) — Digital workers that survive production

(55:00) We don't use workflows, we handle exceptions

(57:50) Workflow thinking constrains knowledge work

(1:03:55) Compiling and executing programs on the fly

(1:09:58) ROI isn't just cost — it's faster value

(1:23:16) Snehal Antani (Horizon3.ai) — Trust through autonomous security validation

(1:25:50) Attackers chain small weaknesses

(1:47:49) Assume breach, manage blast radius

(1:58:08) Seventy-seven seconds to lose

(2:02:09) Deception is the EMP of AI

(2:03:02) Closing — The disk trade, Panopticon, and FSD

(2:05:46) Products get pulled by the market

(2:23:04) Crypto is banking's overlay

(2:26:18) Rules need to get better

Guests:

Andy Fernandez — HYCU (@hycuinc)

David Villalón — Maisa (@davipar)

Manuel Romero — Maisa (@mrm8488)

Snehal Antani — Horizon3.ai (@snehalantani)

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit briefing.ai-in-the-am.com