A method
How to read an AI benchmark claim.
What "ninetieth percentile on the bar exam" actually meant, why benchmark numbers systematically mislead, and the four questions to ask before believing any AI performance figure.
Read the methodA new podcast·launching summer 2026
A weekly reading of what is happening in artificial intelligence.
To think clearly about artificial intelligence — and to orient its trajectory toward use, not replacement.
The Aiomic Age is a podcast about artificial intelligence, in the company of a weekly editorial newsletter, an open library of methods for assessing it, and a single position on the long horizon.
Episodes publish in three rotating formats — a deep dive on a single technical or thematic question; a conversation with an expert, followed by a brief audit of where they were strong and where they overreached; and a journal club, a structured critique of a recent paper in the tradition of medical journal clubs.
Every conversation closes the same way. We ask the guest two questions: what would have been different in their career if AI had existed at the start of it, and what specific change to the way AI is being used would they most like to see in their field. The orientation is constructive — you cannot stop the wave, but you can orient its trajectory. Augmentation, not replacement.
I. Coming on the show
Questions we are preparing to examine.
A few of the conversations being readied for the first season. They will publish on this page as they become available.
-
A deep dive
Episode one
The evaluation problem.
Why benchmarks systematically mislead, what "ninetieth percentile on the bar exam" actually meant, and what a clinical-trial standard for AI evaluation might look like.
-
A journal club
In preparation
Reading METR's task-length paper.
The single most-cited empirical claim about agent progress, examined the way a methods reviewer would. Seven beats, plainly.
-
A conversation
In preparation
On deployment, with someone who has done it.
A long-form interview with a senior leader who has scaled a production AI system inside a regulated industry — and who has also pulled one back.
-
A deep dive
In preparation
Why agent demos succeed and deployments stall.
The unforgiving mathematics of multi-step reasoning under noise. Three nines becomes seventy-three. The reason a five-step demo looks magical and a twenty-step deployment fails.
II. First Principles
Durable methods for assessing any AI claim.
A growing library of short, evergreen essays on how to think about artificial intelligence. Each one teaches a single transferable method or mental model — no news, no opinions, only the questions a thoughtful reader should ask before believing the next AI headline. The first essays are publishing now, ahead of the podcast.
A model
Why agent demos succeed and deployments stall.
The unforgiving multiplicative mathematics of multi-step reasoning under noise. Three nines becomes seventy-three. The reason a five-step demo looks magical and a twenty-step deployment fails.
Read the modelIII. The Signal
A weekly reading of what is changing.
A weekly editorial newsletter, delivered on Sunday mornings. Each issue is eight or ten things we read this week — each with a sentence on its wider significance, and a closing paragraph on the pattern they collectively trace. Written by one host in rotation. About twelve hundred words.
The Signal launches alongside episode one. Subscribe now and we will send you issue one when it publishes — and nothing before.
IV. The AGI Clock
Our reading of how close we are to a meaningfully transformed world.
Not a forecast. A position, argued in public against a published methodology, moved only when the evidence does — and only with an essay to explain the move. The methodology is being written, in the open, before the first reading is taken.
- Narrow tools
- Useful assistance
- Substantive autonomy Awaiting first reading
- Broad transformation
- Beyond
Methodology in public draft · first reading at episode one.
The first reading will publish alongside episode one, after the methodology proposal has been read and argued with in public. Until then, the spectrum is uncalibrated — a placeholder for the work being done. We are taking the time to define the threshold precisely because the threshold matters more than the number that follows from it.
Once the reading is taken, the Clock will move only on events from a published taxonomy, and only with a published justification. We expect two or three movements a year. Each one will be an essay. The number is the summary; the prose is the work.
Methodology proposal in draft · Read the current version
V. The hosts
Three readers, from three vantage points.
A founder
Sven Jungmann
Builds and deploys artificial intelligence in real-world settings. The question he brings to any AI claim is whether it would survive contact with a production environment — what works, what fails, and what the gap between a demo and a deployment actually looks like.
An investor
Karim Galzahr
An investor with an economist's discipline. The question he brings is whether the numbers add up, and what it means for the wider economy if they do.
A scientist
Moss Trestman
Works on artificial intelligence for science. The question he brings is whether the science under the claim is actually real, or whether it is something else dressed in scientific language.