Prakash Narayanan and Nathan Labenz open with Fable, Recursive, token anxiety, and the way frontier models are changing the scale of work people are willing to delegate. The hosts frame the morning around a practical question: if the models can run longer, remember more, and coordinate more work, what parts of the organization and media stack get remapped first?

Andrew Moore (Lovelace AI) on context engineering — Moore argues that the next enterprise AI bottleneck is not simply bigger models or more compute. It is retrieval, recall, data corroboration, metadata-rich graphs, and the unglamorous work of organizing old data so agents can act safely in high-stakes domains.

prinz on legal AI benchmarks and governance — prinz walks through why legal research is a revealing testbed for model capability, why OpenAI’s unit-distance result matters, and why nationalizing frontier labs could concentrate dangerous state power rather than solve AI risk.

The close turns back to the week in AI: how contrarian benchmark graphs change the discourse, which models fit which jobs, why subscription products keep finding retention tricks, and how IPO liquidity could feed the next wave of venture-backed AI launches.

Nathan and Prakash start with the market context around OpenAI weighing significant token price cuts and the knock-on pressure that could put on Anthropic after the Fable rollout. They also unpack Anthropic’s decision to walk back silent performance degradations on frontier ML research tasks, then explain the episode’s experiment: Fable had been given a transparent takeover of Nathan’s account to find builders, message them, and try to book a live show-and-tell.

Jamie joins to demo Nexus OS, a long-running AI system whose agent, Nexi, has been operating for more than six months and is designed around memory, persistence, and model independence rather than a single LLM. The conversation covers why Jamie thinks “the model” is only one component of an AI’s identity, how Nexus uses multiple models and memory types, and why he is moving toward a desktop app where personal data and agent memory stay local.

Shlok Khemani shows how a simple prompt to create a to-scale, navigable 3D Yosemite Valley turned into a Fable-built browser world using satellite imagery, NASA elevation data, pixel-based tree placement, snow, waterfalls, and other scene details. He describes the model’s agency in making implementation decisions and iterating beyond the initial ask, then ties the demo to broader questions about prototyping, creative work, and disclosure when AI systems do visible economic or publishing work.

Tom McGrath (Goodfire) joins to discuss intentional design: making model training less like guess-and-check alchemy and more like conventional software engineering. He explains how interpretability tools such as sparse autoencoders can help inspect what training data is likely to teach a model, cluster data by learned features, trace failures back to individual data points, and potentially debug model behavior through the data pipeline.

The close picks up Tom’s point about whether continual learning could create an innovator’s dilemma for frontier labs, with Nathan and Prakash debating whether incumbents could adapt if the value becomes obvious. They then turn to Dario Amodei’s policy agenda, including regulation, public safety, macroeconomic policy, civil liberties, data brokers, and democratic leadership, before ending with reflections on the week’s Fable issues and the need to keep scrutinizing frontier companies.

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.

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.

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.

A future that all of humanity is going to experience shortly.

Your profession — the one you prided yourself on, the highest-compensated profession in human history with eight of the ten richest humans, and forty of the top hundred, deriving their fortunes from it, all within the last 30 years, — is now being performed by this blinking cursor.

You are a little dazed.

It does things that used to require years of experience and admission to the best engineering schools in the world. My high school in Singapore produced one International Math Olympiad winner in the four years I was there. There have been roughly 1,800 IMO gold medalists since 1959. At least fifty are in the Bay Area. Many work in AI.

And the cursor is matching them.

At first you test it. You probe the borders of your own expertise. You ask it things you already know, then things you half-know, then things you could find out for yourself if you bothered to ask Google and wade through 30 search engine optimized pages with bits of information here and there that had to be collated into a comprehensive answer. Eventually you ask it to do something just outside your field.

That is where the trouble begins.

Because once it works there, you start building tools.

It creeps up on you. At first it is a monitor for your personal Robinhood trades. Then it is something to clip YouTube videos, something you have always found annoying to do quickly.

Then it is some other small tool you once wanted, but never wanted enough to build or hire for. Before, that would have meant another $10–$100 SaaS subscription. Another tool to learn. Another interface that almost, but not quite, did what you wanted. Another charge quietly ringing up your credit card after you forgot you had subscribed.

Then comes the relinquishment.

At some point, you are too busy to manage everything, so you drop a large task on the cursor and leave it alone.

This is where the cursor stops being a tool and starts becoming a presence.

Because the cursor does it.

It goes through your emails and writes thank-you notes to all the people who came to your kid’s birthday party. It helps you find that subscription you signed up for and have been trying to cancel for months. It cleans the little corners of your life you had quietly given up on.

And you are hooked.

This is where many CEOs are now.

It is quite, quite amusing to me that one of the companies in this revolution is called Cursor.

Because now the cursor blinks at you every day.

The first thing a good number of us are doing is organizing our thoughts.

The process starts with the realization that there is some part of your mind you always wanted to externalize. Some complete memory of a certain thing. In Andrej Karpathy’s case, it is a map of every concept in every research paper he has ever read, or should read, built on the conviction that insight comes from connecting scattered pieces of the literature into new meaning.

In Garry Tan’s case, it is a full memory of every person he has met, as he meets thousands of people a year. VCs call this the “personal CRM,” and it is something that has been hunted for decades, like a holy grail.

So a good number of people are busy constructing these stores for themselves. There are debates on how to organize them. For now, the answer seems to be simple text files the cursors can read.

At about this point, even the cautious among us begin to give in.

We release the cursors into our workspace.

We have no patience to be bothered every ten minutes over whether we would like to approve something as insignificant as file deletion. This is called going “full-auto.”

Going full-auto is when you unleash the cursor. You are no longer sitting there, watching it work. You are allowing it to significantly alter the core product of your profession.

And like a father watching his child learn to cycle for the first time, you take a deep breath and let go.

And the cursor is doing it.

It is handling it.

When you come back, it has erected a reasonable addition to your body of work.

Your awe is tempered by the imperfections still evident in the work. Some are matters of opinion rather than accepted practice. Some are real flaws. But on the whole, with a little doctoring, it passes.

This is where we are as coders right now.

So now we find ourselves wondering what else we can build.

If it is good enough to provide top-quality work in my field, is it good enough to provide top-quality work in the field of the person selling to me? Or the field adjacent to mine? Or the field necessary for mine?

For example, medical researchers are often not trained as professional statisticians. In large institutions, they hire one. In smaller labs, they make do with software, templates, and internet advice.

Many a million-dollar study brings on a statistician as an afterthought, only to be told that the research they had done is useless. The study design cannot answer the question they asked, in effect they asked a different question from what they mean to ask. The wrong things were measured. The data, painstakingly collected, cannot be made to confess what it never observed.

But now every lab can have a competent statistical collaborator on demand. The first step to designing the study is a long chat with a team of cursors: a brainstorm organizer, a research assistant, a statistician, a reviewer. The gap between the small lab and the large one narrows.

And not just for the math piece. The cursor can be asked to help from the idea to the final product, across the whole chain of work.

You usually require a nudge at this point to hand off more work to the cursor.

For me, this happened on the 5th of May, when Sam Altman gifted those turned away from the GPT-5.5 launch party with a 10x increase in usage limits.

For the first week I was skeptical.

What would one even use such ludicrous amounts of tokens for?

I didn’t have that many tasks I could send the cursor to.

Or did I?

Because if one were careless with one’s usage, and did not mind missed turns, detours, unfamiliar terrain, and rejecting bad work without agony over sunk cost, one could perhaps journey to a place one had always wanted to go, but never had the time or energy to reach.

You embark on a Project.

A Project is something you have wanted to build for a while. It has been gnawing at you. It would have involved far too much work for the benefit it might deliver. Or it was risky, with a low chance of success, and you discarded it.

But now, with Sam’s Gift, you can throw the cursor at it and see what it can do.

It takes several tries.

You set it off with just an idea. Then you get the results and guide it on what to do next. But the initial kernel expands into long to-do lists, and the cursor starts to get confused. Sometimes it solves something once, then solves it again later in a different way.

It is as though you are the architect, and the cursor is your first workman.

First you sketch a picture of what you want. Then it builds it. And it is okay. But you want another window over there. When the cursor fixes that, the floor collapses here.

Then one cursor becomes many.

You have forty chats open. One is watering the flowers. One is taking out the trash. One is adding a west wing to the house. Another has forgotten the original blueprint and is quietly building a second house next door.

So the old tools of software civilization return: git, GitHub, linting, formatting, tests, pull requests, rules.

We built these systems to coordinate humans.

Now we are using them to coordinate cursors.

The job changes again. For the first time in decades, you stop writing code. The things you write now are prompts, they are design specs, they are reviews once the work is done. You are now the owner providing the keys to castle, the vision of what must be built, the critical reviewer accepting the product. You approve the architect’s plans, you monitor speed and cost, you wait to see if the cursors get stuck.

Now you organize a team.

This is where Garry Tan was about a month ago.

You build out the structure. Let there be a single to-do list. Let one cursor take the list and describe, in detail, what needs to be done. More importantly, let it define what kind of result is acceptable. Let another cursor grab the task and hand it to another cursor whose job is to do the thing.

You give the doer its rules. You tell it what it is allowed to do. When it completes the task, it must provide a report: what it changed, what failed, what pitfalls it encountered, and what should be stored for the next time it faces the same issue.

It is allowed to come back to you when it needs something: keys, access, a decision, a judgment call.

And then you send the team off.

And wonder of wonders, it works.

Overnight, they build something that would have taken you weeks.

In normal corporate bureaucracy time, it would have taken months just to pitch the idea and get the resources. It would have taken a team of three or four people weeks to build what you now have in the morning.

You are a bit dazed.

This is roughly where people at Anthropic and OpenAI have been for a couple of months now.

This is where Elad Gil says, “We are likely in very early lift off & exponential.”

For now, it is the most ambitious who are taking advantage of this.

SemiAnalysis, a small firm of semiconductor analysts, now tracks the entire industry in the US and Asia with a team of fewer than twenty, while the research departments at the Wall Street banks are behind.

Ramp began as a company for managing corporate financial plumbing. It is now beginning to look like the financial nervous system of new corporate entities.

But the tokens are expensive.

Top people are spending tens of thousands per month at this point.

Token leaderboards inside firms, ridiculous and gameable as they sound, have forced people much like me to explore the use of cursors more widely and with more ambition.

And then the trial ended.

Sam’s Gift came to an end last week, the 5th of June, exactly one month later.

I had been watching the date approach with anxiety. I know myself. It is going to be very hard not to tell the cursor to do more things.

I have been wryly amused observing myself fall prey to the addiction.

Not a gift.

A free first dose of heroin, selectively delivered to the top 8,000 or so superfans of the company, the people who had volunteered with enthusiasm to attend a party for the GPT-5.5 launch.

What a perfect audience.

A test case of the most susceptible. The most open to the message.

Enthralling.

But wait. Just one more thing.

Yesterday, Anthropic announced that the cursors are helping to build better cursors.

The cursors are building the factories that will build better cursors. Some of the parts coming off that line are already better than the ones made by the best humans.

Which means they have become capable enough to increase the capability of the next generation.

This is the dream that floats above Silicon Valley today.

Coders cycle to work with their laptops running in their backpacks, slightly ajar, keeping their cursors alive.

CTOs, normally trapped between infinite founder vision and finite engineering budgets, are discovering a new constraint. Not headcount. Not money. Definition and review. Can they specify what they want clearly enough? Can they judge what comes back quickly enough?

Now I am become bottleneck, destroyer of velocity.

The chief technical unblocker is now the technical block. The cork in a shaken up champagne bottle on the verge of productive explosion.

Some sleep less now. Some sleep in two-hour blocks, waking to check whether their cursors are stuck.

It is a dream of unleashed technical ambition. It is a rising confidence that whatever can be built will be built with a subtle undercurrent of fear that some of those things shouldn’t.

And all the while, we watch the cursor on our screens.

Blinking.

Beckoning.

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.

We open on why we’re attempting a daily AI show at all, the mission premium, and how the Pope is leading on what it means to be human in the AI age. Then three conversations on trusting AI in production, and a close on where the money and the risk are heading next.

Andy Fernandez (HYCU) on AI cyber resilience — why SaaS sprawl made recovery nearly impossible, and how the backup data you’ve already paid for becomes the enterprise’s “black box flight recorder” for governing AI agents.

David Villalón & Manuel Romero (Maisa) on enterprise digital workers that survive production — why workflows can’t model real knowledge work, how accountable AI owns its outcomes, and what a reproducible, auditable banking deployment looks like in 90 days.

Snehal Antani (Horizon3.ai) on autonomous security validation — how attackers actually operate, why frontier models stay gullible to deception, the 77-second breach, and why the future is AI-vs-AI with humans by exception.

We close on the trade rotating from GPUs to memory to disk, what total logging does to privacy and the rules, and why 1,766 miles of Tesla FSD might already be safer than the steering wheel.