The Gayzette

About a quarter of them recalled an experience from the period when they were clinically dead. The interesting finding is not that they recalled something. It is what the recollections had in common. Reported as strikingly consistent structural elements: a sense of detachment from the body, an awareness of the resuscitation happening, a boundary of some kind, and encounters with deceased relatives. That consistency held regardless of the survivor's nationality or religious background, which is the opposite of what you would expect if these were culturally scripted stories people tell afterward.

There is a second result with practical teeth. Survivors whose experiences were clinically acknowledged, rather than waved off as "just a dream," fared better psychologically than those whose accounts were dismissed. That is a finding about how clinicians should talk to patients, and it stands on its own regardless of what the experience is.

Now the honest part, and it matters. This comes to us through a single secondary report of the underlying Resuscitation Plus paper, so treat the specifics as early reporting rather than settled. The study itself states it cannot rule out residual brain activity as the explanation. A shared structure is not proof of a shared destination; it may just mean brains fail in similar ways.

Accuracy was no better than chance. That is the headline result, and the team published it as such, which is worth pausing on. A null result is a study that found nothing, and in this field nulls have a habit of quietly not getting written up. This one did.

Then the part they did not plan for. Two participants, unprompted, described the researchers' locations and what they were doing behind closed doors. That was not the target, not the hypothesis, and not something the protocol was designed to score. The team reported it anyway as an unplanned finding, published in the journal Explore, with the researchers named on the paper: Marina Weiler, Damon Abraham, David J.P. Acunzo and Niffe Hermansson.

The caveat is the one the researchers themselves imply by calling it unplanned. An anomaly noticed after the fact, in two people out of 21, with no pre-set criteria for what would have counted as a miss, is a lead and not a result. It is the thing you design the next study around, not the thing you conclude from.

Which is the whole point. This is what the field looks like when it behaves: say what you predicted, report that it failed, and report the odd thing too.

The interesting casting is on the professional side. Adele Zaikman joins the pro roster after winning "Dancing with the Stars: The Next Pro," a spinoff competition, and is paired with Conner Leavitt. Longtime pro Sharna Burgess returns after a multi-year hiatus.

Zaikman's route is the part worth noticing. The traditional path onto a show like this runs through competition circuits, conventions, and whoever happens to know a casting producer, which is to say through a network that is mostly invisible from outside. A televised competition for the job makes that pipeline public and legible: here is the standard, here is the audition, here is who cleared it. That is a small structural change in how a closed room gets entered, and it is the same change that shows up whenever any gatekept field publishes its criteria.

The caveat is that one spinoff does not reroute an industry, and a competition show is still a show, cast for television as much as for technique.

Still, for a business built on developing dancers, the question of how somebody gets seen is not an abstraction. It is the product.

After a sold-out Hollywood run, the show is going out on the road for the first time in 15 years. Confirmed fall dates are Boston on September 27, Philadelphia on September 29, Washington DC on September 30, and New York's Tribeca Performing Arts Center on October 2 and 3, hosted by Patrick Bristow.

What makes this a story rather than a listing is the economics it demonstrates. Unscripted adult puppet comedy, with no script to license and no post-production, sold out a residency and is now worth the cost of moving 80 puppets and a crew between four cities in a week. The format is repeatable by design, because improv means the show regenerates its own material every night. That is a very different cost structure from scripted comedy, and it is the reason a 15-year gap can close this fast.

The caveat is scale, not quality: four cities and a handful of dates is a test tour, not a national rollout, and whether it extends past October is not something the announcement says.

The distribution ladder used to have a missing rung. Theatrical release was gated by studios and distributors who decided what was worth a screen, and a creator-made show was not in the conversation regardless of its audience. What the Nebula deal shows is that the gate now opens on a different question: can you fill the seats on a Saturday. If yes, the chain will book you, because an empty auditorium earns nothing and a creator with a devoted audience is a known quantity in a way a new film is not.

Sixty-one locations is not a wide release, and that is worth saying plainly. This is an event screening, one night, for people who already watch the show. But the path from "one night" to "a real window" is a path that exists now, and it did not a few years ago.

The receipt is the specificity. Not a partnership announcement in principle, but a date, a count, and tickets already on sale.

Ten titles are longlisted in the international category, including "What Is Intelligence?" by Blaise Aguera y Arcas and "Creativity in the Age of AI." Rights Director Laura Lesum put the reasoning plainly: "at a time when knowledge is increasingly processed by AI systems, making its origins visible matters more than ever." Winners are announced October 8 at the Frankfurt Book Fair.

The word doing the work in that quote is "visible." Not banned, not detected, not policed. Visible. A prize cannot verify how a manuscript was made, and this one is not pretending it can. What it can do is make provenance part of what it rewards, which turns a fuzzy anxiety into a stated value that authors and publishers can point at.

The caveat: this is one award in one category, and a criterion is not an enforcement mechanism. Nobody is auditing drafts.

But it is a publishing institution deciding, on the record, that the origin of a book is part of the book's worth. That is the kind of quiet institutional signal that tends to precede a norm rather than follow one, and ghostwriting sits in an interesting spot next to it: a discipline that is, by definition, a human voice standing behind somebody else's page.

Auto-Tagging labels customers automatically based on what they do, like making a purchase or filling out a form. A tag is just a sticky note on a contact record, but an automatic one means the sorting happens at the moment of the behavior rather than in a spreadsheet later. Trigger Filters let you segment an audience by the data people entered in a form response, so the answers to your intake questions become the basis for who gets what. Conditions split an automation into branches, so different customers walk different follow-up paths from the same starting point. The same update added downloadable, branded invoice PDFs.

None of these is new as a concept; dedicated email platforms have had branching for years. What is new is that they are sitting inside the website builder, one layer away from the form and the checkout that generate the behavior in the first place. The value of that is not features, it is fewer joints. Every integration between two systems is a place where a contact record goes stale or a tag never arrives, and removing a joint removes a class of silent failure.

The caveat: a branching automation is only as good as the segments you define, and a badly designed condition just delivers the wrong email faster.

For a small operation, that is still the difference between treating everyone identically and treating people like you noticed what they did.

On September 10, OpenAI released the Agents API in public beta. It exposes the same agent, environment and session infrastructure that powers Codex as a standalone service. Three pieces matter for a solo builder. Automatic context compaction means the agent trims its own working memory as a job runs long, instead of hitting a wall. On-demand tool search means it fetches a tool's definition only when it needs it, rather than paying tokens to carry every tool description in every request. And native multi-agent delegation means subagents run in parallel, each with its own context, so one long task splits into several short ones.

Sandboxes, the isolated machines where the agent's code actually runs, can be hosted by OpenAI or by nine partner platforms including Cloudflare, Vercel, DigitalOcean and E2B. There is no platform fee. You pay for tokens and tool usage, full stop.

The caveat is the phrase "public beta," which in this industry means the interface will change and you should not wire anything load-bearing to it yet without expecting to rewire it.

Cognition released SWE-2 on September 10, post-trained from Kimi K3, a base model with 2.8 trillion parameters. Parameters are the tunable weights inside a model, roughly a measure of raw capacity, the way engine displacement is a rough measure of an engine. Post-training is the specialization pass on top of that base, where a general model gets taught to do one job well.

The scores: 50.0% on FrontierCode 1.1 Main, against 50.9% for Claude Fable 5.1. That is within a single point of a frontier model, at 64% less cost. On Terminal-Bench 2.1, which tests whether a model can actually operate a command line rather than just write code that looks right, it hits 92.8%. Cognition also reports it needs 58% fewer turns than its predecessor SWE-1.7 to land a real edit, which is the number a solo builder feels most directly, because every turn is a round trip you sit through.

The caveat, and it is a standing one for the whole category: these are the vendor's own benchmark disclosures on benchmarks the vendor selected. Cognition publishes them in more detail than most, which is a point in their favor, but a benchmark is not your codebase.

The trend underneath is the real receipt. The cost floor for agentic coding keeps dropping while capability holds flat or climbs.

On September 9, Suno launched v6, the experimental v6-wild, and the free v6-mini, all trained from scratch on catalog licensed from Warner Music Group, BMG and Believe, including its TuneCore arm, with those rightsholders paid per generation. Per generation is the load-bearing phrase. It means the payment is metered to use, not a one-time settlement, which is the structure a licensing chain needs if it is going to be a business rather than a truce. Every prior Suno model was retired the same day, which is a costly way to say the old ones were not defensible.

The gap is real. The training data excludes Universal Music Group and Sony Music Entertainment recordings: neither label is among Suno's three licensing partners, and both labels' infringement suits against Suno continue. So this is the first AI music generator with a real, paid rights chain behind part of its catalog, not all of it.

The practical version for anyone putting generated music under a monetized video: the provenance story is much better than it was last week, and still not complete. Read Suno's own output-license terms before you rely on it, because a clean training set and a clean output license are two different documents.

On September 9 the FDA granted accelerated approval to sevabertinib, brand name Hyrnuo, for locally advanced or metastatic non-squamous non-small cell lung cancer carrying those HER2 activating mutations. In the SOHO-01 study of 69 previously untreated patients, the objective response rate, meaning the share of patients whose tumors measurably shrank, was 75%. Of those who responded, 73% were still responding at six months and 38% at twelve months.

Two caveats worth holding together. Accelerated approval is a conditional pathway: the FDA clears a drug based on a measure likely to predict benefit, like tumor shrinkage, while confirmatory trials run. It can be withdrawn. And 69 patients is a small trial, which is typical for a narrow genetic subgroup and still a small trial.

What makes it a receipt anyway is the shape of it. Sequence the tumor, find a specific mutation, take a pill twice a day. That is a workflow that barely existed fifteen years ago for lung cancer, and the patients in this subgroup previously had few good options.