Presence ran the same scripted conversation — a codename on turn one, thirty-eight middle turns, a recall question at the end — against Gemini Live, OpenAI Realtime, and its own stack. Then it published everything, including the places it loses.

17.3×

That is how much more the fortieth turn cost than the first on Gemini. Turn one: $0.00215. Turn forty: $0.03714. Context grew from 559 to 13,231 tokens, and every turn was billed for all of it. Zero cached tokens across all forty turns. Seven minutes of audio came to $0.80.

The cost curve is roughly quadratic in turn count rather than linear in minutes, which punishes exactly the behaviour voice products encourage: rapid back-and-forth.

OpenAI has largely fixed this. By turn forty, 3,648 of 3,740 input tokens were cached and per-turn cost was flat, 1.25 times the first turn. But caching is best-effort, and the run shows what that means in practice: it did not engage at all for nine turns, cost climbed to $0.025, then collapsed to $0.0067 when 896 tokens finally cached. Two turns cost about five times the median. The problem is not growth. It is predictability.

Presence is flat by construction, because it bills occupied session time rather than tokens. That is a different unit of account, not a smarter optimization, and the authors say so.

2.6 seconds

Ask a voice agent something hard and it has to go away and think. The interesting question is what happens to the conversation while it does.

The test: while a tool call is pending, speak again — "how long will that take?" — and see whether the agent answers. Presence answers in 2.6 seconds. OpenAI answers in 5.8. Gemini does not answer until the tool returns.

On a twenty-second tool, that is twenty seconds of a session that appears to have died. Two seconds of silence is long; at four you say "hello?"; at ten you assume the line dropped.

This is a product difference, not a latency difference. An agent that can think without abandoning the conversation can look something up mid-sentence, start a slow job and keep talking, or say "I'm still working on that" and mean it. No single-turn benchmark can see it.

1,216 milliseconds

This is the number that nearly became the headline for the wrong reason.

Presence's first three runs said its stack was blocked, exactly like Gemini. The authors were preparing to publish that. The cause was the harness: the follow-up sent exactly 1,200 ms of trailing silence against a 1,200 ms endpointing threshold. Voice activity detection runs in 32 ms frames, so the first boundary at or above the threshold is 1,216 ms. They were one frame short. The turn was never committed, no transcript was published, and from outside it looked identical to a session that had stopped listening. Internally, everything had worked — transcription, model, speech synthesis — and a single missing frame withheld the audio.

A harness bug was indistinguishable from the finding. They published both, and added a rule: an inconclusive follow-up is reported as inconclusive, not as a block.

What survives

Latency numbers will age; Gemini was fastest to first audio, at 1,064 ms p50 against OpenAI's 1,253 and Presence's 1,395. Prices move monthly. The method is the durable part: measure the session, not the turn; test what the agent does while it thinks; and when your instrument disagrees with your expectation, suspect the instrument.

Especially when it is one frame off.

Source: What happens on turn forty — Bohita Presence.

viiwork measured it on two RTX 5090 hosts, and the faster cable barely moved the number that mattered.

The claim

A model does not reread your conversation on every turn. As it processes the prompt, each layer produces key and value tensors for every token — the KV cache. Keep that cache warm and the model skips straight to generating. Lose it and the receiving host has to run the whole prompt again before it can emit a single token. That is prefill, and on a long session it is pure repeated work: about 41 seconds for 50,000 tokens on an older RTX A4000, minutes at 200,000 tokens on slower hardware.

Keeping a session on the host that holds its cache avoids this. But hosts fill up, and when a turn spills, kv_share copies a replica of the cache to the host the session would spill to next, so it can import instead of recompute.

The copy is incremental. After a turn of at least 8,192 tokens, the host saves the session and ships only the delta — roughly 240 MB of a 1 GB cache per 50k-token turn, because most of the conversation went across last time.

So: how much does the cable matter?

Three numbers

247 MiB was the typical replica delta. Sending it took 3.24 seconds at 1 GbE, 0.81 seconds over 10 GbE through Tailscale, and 0.55 seconds on the direct cable. The raw link carried 9.41 Gb/s; Tailscale managed 3.0–3.4 Gb/s, which is why the first jump was about fourfold rather than tenfold.

93.8% → 96.8% was the change in warm host switches, in the 16-session runs, going from 1 GbE to 10 GbE with Tailscale. Partly-warm switches fell from 5.4% to 2.2%. Direct 10 GbE was faster on transfer but landed at 96.2% warm, inside the run-to-run noise.

2.8 seconds was the import into the receiving engine. Every time. At every link speed.

That last number is the story. Once the replica arrives, there is a fixed cost the network cannot touch, and it is larger than the best-case transfer.

The bug

The most useful finding was not a number. Traces showed that viiwork remembered importing a replica, but the engine could later evict that conversation; on a return visit it skipped the import even though the cache was gone. The "import-once" bug was fixed in v2.11.1 by asking the engine what it still holds before skipping.

That bug made the whole mechanism look weaker than it was. Some of what looked like a bandwidth limit was a state machine being optimistic.

What to do with this

Measure end to end before you buy hardware. The send time improved sixfold; the warm-switch rate improved three points. Optimizing the transfer and stopping there would have produced a confident, expensive, wrong conclusion.

Then profile the import path. Fixed per-move costs set a floor that bandwidth cannot lower, and at this workload 1 GbE was already workable — average kv_share traffic was 320–450 Mb/s in each direction, and most switches were warm anyway. The caveat is real: these are single 90-minute runs, not statistical trials, and the author says so.

Buy the 10 GbE if it is cheap. Fix the state machine first.

Source: Two RTX 5090 hosts sharing a KV cache over Ethernet — viiwork.

Porcelain is roughly three thousand years old. The interesting part is not that we are still improving it. It is that the improvements keep arriving disguised as something else.

The kiln

The first step toward porcelain happened in northern China, when potters started working with kaolin, a chalk-white clay, and firing it at around 1,200°C.

Clay is stacks of plate-like crystals held together by hydroxyl groups. Heat breaks those bonds, the plates reform into needle-shaped crystals that lock together, and at porcelain temperatures the silica melts into a glassy liquid that fills the gaps. The result is dense, translucent, and strong.

The recipe was not enough. Ordinary kilns could not hold that heat; their vents let it escape almost as fast as it was produced. Chinese potters designed downdrafts, forced the hot air out through floor vents, and invented heat-resistant bricks so the kilns would not melt. Porcelain was a system: clay, fuel, kiln, brick, technique.

The failure

Porcelain reached Europe in the 13th century and became an obsession. "White gold." European elites spent fortunes trying to make it and failed for centuries, mostly because nobody knew the key ingredient was kaolin, or that kaolin alone was not enough without ground "China stone."

They got close enough to invent maiolica, faience, and Delftware — beautiful, but not porcelain. Not impermeable, not strong, and it did not ring when you tapped it.

The theft

The breakthrough is a story about a con man and a polymath.

Count Ehrenfried Walther von Tschirnhaus spent decades focusing sunlight with lenses and mirrors, burning minerals to see what they did. In the winter of 1693–94 he found a clay that would not melt, and made a hard, yellowish, translucent ceramic he called "wax porcelain." He could not scale it.

Meanwhile Johann Friedrich Böttger, a Prussian alchemist, claimed he had discovered the Alltinktur, a substance that cured disease and turned base metal into gold. Prussia's king had him taken into custody; he escaped to Saxony, where Augustus the Strong locked him in a fortress and ordered him to produce gold. He could not. Augustus eventually assigned Tschirnhaus to supervise the alchemy project; Tschirnhaus, who cared nothing for gold, redirected Böttger to porcelain instead.

By 1708 they had it. Tschirnhaus died that October. Böttger claimed sole credit, and Meissen became Europe's porcelain center.

The hip

Dekton is porcelain pressed at 25,000 tonnes and fired at 1,200°C until atoms diffuse across particle boundaries and the slab becomes one solid mass. It is Porcelain 3.0, and it ends up under a frying pan.

Better examples exist. "Wet-activated" tiles that grip when water hits them, preventing the kind of fall that permanently injured Charles C. Mann's mother-in-law. Bioceramic scaffolds that dissolve as bone grows into them. Lithium disilicate dental implants.

The kid

Mann told his son how remarkable it all was — the thin slab of porcelain he cooks on, the hip, the tile. His son gave him a blank look. He had never lived in a world that was not made of porcelain.

That is the whole essay in one reaction. The most important materials are the ones nobody notices, and the surest sign that a technology has arrived is that it becomes invisible.

Source: Everything rhymes with porcelain — Charles C. Mann, Works in Progress.

The word comes from a short essay by Nadeesha Cabral, who noticed two things happening at once. People are reverse-engineering tools and rebuilding them to fit their own requirements. And products are wiring feedback boxes to agents that can modify the software in place, collapsing the old prompt → PR → CI → deploy loop into a single conversation.

Cabral's image for it: "We're no longer mass-producing aircraft and then flying them. We're building and changing the aircraft as we fly."

Good name. Real shift. But "softer" describes the interface, and the engineering underneath gets harder, not softer.

How the change happens

The reverse-engineering trend is not new. People have forked and patched software for as long as software has existed. What is new is the cost: rebuilding a tool to your requirements used to take weeks, and now it takes an afternoon plus a review. When the price of a bespoke tool drops by an order of magnitude, "use it as shipped" stops being the default.

The in-app agent is newer. Instead of filing a feature request into the void, you describe the change and the agent makes it. The product becomes a substrate you shape rather than a product you buy.

What gets harder

Reproducibility. If the software differs per user, "it's broken" becomes unreproducible. A bug report has to carry the state of the code, not just the steps to reproduce. Support turns into archaeology.

Trust. A program that rewrites itself at runtime is a security model change. Who reviews the diff? What is the audit trail? If the agent can change the software, the agent is inside the trusted computing base, and most organizations have no idea how to reason about that.

Ownership. Who maintains the bespoke fork when the session ends? The promise is that everyone gets exactly what they need. The bill is that everyone gets a codebase.

The wrong metaphor

Aircraft are regulated, redundant, and designed for failure. Most software is none of those things, and the reason we accepted the write → compile → test cycle was that it made failures cheap and local.

Softerware makes failures expensive and strange: in-session, user-specific, hard to reproduce. That does not mean it will not happen. It will, for the same reason spreadsheets happened — it moves power to the person with the problem.

The interesting question is not whether software gets softer. It is which software should. My answer: the kind whose failure is cheap. Personal tools, internal dashboards, scripts, small games. Not the systems holding money, health records, or other people's data.

A softer interface is a feature. A softer audit trail is a liability, and the two arrive together.

Source: Softerware — Nadeesha Cabral.

That was the visiting card, and for roughly two centuries it was how respectable Europe and America managed the dangerous business of wanting to know someone.

How to read one

The syntax was strict enough to be a protocol.

A card left at a house meant you had made an opening. It did not mean you expected to be let in. The other party answered with a card of their own, which meant a visit would be welcome. No card, or a card in an envelope, meant the opposite.

A turned corner meant the card had been delivered in person rather than by a servant. A lady's card might carry an "At Home" day, which told you when to come. Size and format signalled gender and marital status. Debrett's still considered a telephone number on a visiting card a "solecism" in 2007.

The whole ritual depended on servants to open the door and receive the cards, which is why it belonged to the classes that employed them.

Where it came from

Before the 18th century, people left handwritten notes when calling on friends who were out. In the 1760s, the upper classes in France and Italy started leaving printed cards, decorated on one side with a blank space on the other for a note. Cheaper printing spread the practice across Europe and to America, and by the 19th century the rules had become elaborate enough to fill etiquette books.

Then the telephone arrived, the servant economy dissolved, and the card turned into the business card — same object, different job. The business card carries contact details precisely because the visiting card refused to.

What we threw away

The visiting card encoded permission in form. The modern handshake encodes only data. You can send a LinkedIn request to anyone, which is exactly why it costs nothing to receive and means less than a turned corner.

We have reinvented the missing piece with clumsy proxies: the "let's grab coffee" email that means "I want something," the double opt-in newsletter, the calendar link with a thirty-minute default that announces you are busy but flexible.

The card's real invention was the reversible ask. It cost the recipient nothing and committed them to nothing. They could answer with a card, or with silence, and both were legible.

What to steal

Make your first move cheap to ignore.

That is the whole trick, and it does not need stationery. A short note with no ask attached. An introduction offered rather than requested. A link sent because it was useful, not because it opens a conversation you want to have.

Leave the card. Don't knock.

Source: Visiting card — Wikipedia.

The Duo went on sale October 23 at $2,000 for 256GB, up to roughly $3,000 for 2TB. It sold out in minutes. Within days, buyers were posting photos of a bright line and cracks along the inner display at the central hinge, sometimes after an audible snap when the phone was folded.

What Apple said

On Monday the company acknowledged it: "We are aware of an issue impacting the inner display on a small number of iPhone Duo units." Affected buyers should contact AppleCare for a replacement. Apple did not comment on the cause.

What we know

Some units fail at the fold. iFixit built a 3D model of the hinge from teardown footage, CT scans, and Apple patents, and found the reported damage clusters just below where a display cable runs under the screen, near a central support bar. Its hypotheses: the bar may not be dropping out of the display's way as the phone closes, or a curved guide rail may rise too high during folding. Either would create what iFixit calls "a small stress hotspot near the reported break location." iFixit could not confirm a direct cause.

It also found the Duo the most repairable foldable it has scored: 5 out of 10, with clip-based bezels and batteries on releasable adhesive.

What we don't know

The failure rate. Whether the failures cluster in an early production batch. Whether the cause is a design margin problem, a manufacturing defect, or misuse — TechRadar has reasonably cautioned that all three are on the table. And whether a replacement unit is meaningfully different from the one that broke.

What I would ask

Three questions, none of which require a press conference.

Does the hinge meet its rated fold count, and what is that number? Is there a running design change, and which production window does it apply to? And will the hinge and inner display be covered beyond the standard warranty for early buyers?

A company that says "we reinvented the foldable" inherits a higher burden of explanation when the category's oldest problem appears in its product. "We don't know yet, and we're testing it" is a credible sentence. "No comment" is not.

Credit where it is due: Apple acknowledged the issue within days and is replacing units without a fight. That is better than denial and better than Samsung's first response to the Galaxy Fold in 2019. But Samsung's fix included delaying the launch and redesigning the hinge. A warranty process is not a redesign, and if the cause is the geometry, every replacement inherits it.

If the hinge is the problem, why is the answer a repair queue?

Sources: The Wall Street Journal, Reuters, iFixit teardown, TechRadar, VRT NWS.

Zen, the obvious comparison, published 10 in the same window. Helium also climbed from version 0.15 to 0.20 in twelve weeks, and moved its Chromium base from 152 to 156 in seven, rebasing its patch set each time.

None of that is a feature. It is a rate, and the rate is the pitch.

What actually shipped

Read the changelogs in order and you can watch a browser being assembled in public.

The sync layer. Helium is building its own private sync and vault so you do not need a Google account. It arrived in pieces: the provider API and provider selection in 0.16, state deltas and an extension API, then snapshots, deltas, and a simplified vault protocol by 0.19.4. This is the one thing that would make Helium meaningfully different from every other Chromium fork, and it is also the least finished. Treat it as a promise with receipts.

The shell. Vertical tabs got hover-to-expand, side switching, corner shrinking as the strip collapses, and a compact "zen mode" (yes, Helium has a mode called zen; yes, that is confusing). Stacked split view arrived in 0.17.2, along with lazy loading for restored tabs and back-to-opener on by default. Mute tab(s) landed in the context menu in 0.18.3.

The surface. Rounded icons on by default, an animated loading indicator in the address bar, a toolbar that stops overflowing at small widths, and a flag to hide the native window buttons.

The plumbing. uBlock Origin 1.75.0 bundled as a native component, with CNAME uncloaking. Proxy flags. Chrome's nags suppressed in every locale. Google's AI and experimental WebUI resources removed.

Individually, most of these are small. That is the point: they ship weekly instead of waiting for a launch event.

The bill

Fast is not free, and Helium's own notes show the cost.

It is beta, and the docs say so plainly. Features land before they are smooth: 0.20.1 includes "fix crash when closing a window in zen mode," a crash fix for a feature that shipped days earlier. Four Chromium majors in seven weeks means "patches: refresh for Chromium X" is a recurring line item, and every jump risks breaking something subtle. The team is small, which makes the cadence impressive and the bus factor real.

Against Zen

Zen has the features Helium is still adding: workspaces, Boosts, a mods ecosystem, a command palette, and sync that works today through Firefox. It sits on Gecko, not Chromium, so the extension story and the privacy story are different in kind, not just degree.

Helium's advantages are narrower and real: no Google services by default, native uBlock Origin, native bangs, and a release cadence roughly 1.6 times Zen's.

They are chasing the same idea — the Arc-shaped browser, open and local-first — and they read each other's changelogs. Zen users have already requested a compact layout "like Helium Browser."

Pick Zen if you need a finished browser this afternoon. Pick Helium if you enjoy watching one get built.

Sources: Helium releases, helium.computer, Zen releases, zen-browser.app, Zen discussion #13803.

That is the parable at the center of a recent Coached Weekly essay, and it is a good one. As an apprentice, the locksmith took an hour of visible effort to open a jammed lock; customers paid full price and tipped. Years later he could do it in two minutes, and the same customers complained about the cost. He had gotten better in every measurable way, and his perceived value collapsed.

The essay's argument: every outcome in your career is decided by other people's impressions, and impressions form through predictable biases. Four of them.

The four quirks

The labor illusion. People value work in proportion to the effort they can see. Ryan Buell and Michael Norton named the effect; the locksmith demonstrates it.

Recency. Your review is a memory test, and memory overweights the last few weeks. A good year ending in a bad month gets remembered as a bad year.

Loss aversion. Losing something feels roughly twice as bad as gaining it feels good, so a proposal framed as a loss beats the same proposal framed as a gain.

Incentives. People follow what actually gets them paid and promoted, not what the values page says. The essay's sharpest line: look at what the last three promoted people did.

Plus the HALT rule for your own brain — never decide anything Hungry, Angry, Lonely, or Tired — which is folk wisdom with a good batting average.

Where the essay is right

Visibility is part of the job. That is the load-bearing insight, and it is worth accepting even if you dislike it. The person who did the work and the person who is understood to have done the work are different people, and organizations promote the second one.

The win log is the best value in the piece: five minutes every Friday, a running list of what you finished, fixed, or handled. It costs almost nothing and turns your year into evidence instead of recollection.

And the incentives test explains more behavior than personality does. When a team blocks your launch, the useful question is not "why are they being difficult" but "what are they measured on."

Three places I would argue

Narrate judgment, not suffering. "Say it was hard" is good advice for a bad manager and risky advice for a good one. The line this looked simple, but the data was a mess works because it conveys complexity. It was a grind conveys only that you suffered. The stronger version describes the decision: "I rebuilt it three times because the first two approaches broke under load." That is evidence of judgment. The other is a bid for sympathy, and sympathy does not compound.

The locksmith's real problem was pricing. He stayed denominated in time, so getting faster destroyed his price. The fix is not to perform struggle; it is to change the unit. A locksmith who charges for "your lock opens" gets faster and richer at once. The same logic applies to knowledge work: the further your comp sits from hours, the less the labor illusion can hurt you.

Loss framing has a budget. Use it once, on the decision that matters, with numbers you would defend in public. An organization where every proposal is a catastrophe in waiting stops listening.

The rule

Tell people what you decided, and what it cost to decide it. Not how tired you are — what you chose, what you rejected, and why.

That is not bragging. It is the receipt.

Source: Your best work goes unnoticed, Rohan Mahtani, Coached Weekly issue 358, June 12, 2026.

1838

Louis Daguerre pointed his camera out of a window near the boulevard and opened the shutter. The exposure took minutes — accounts say up to ten — which is why the street looks deserted. Paris was not quiet that morning. The carriages, the pedestrians, the dogs, all of it moved and vanished, leaving nothing behind but the buildings.

The two figures at the lower left survived because they had a reason to stand still: one was working, the other was being served. Nobody posed. Daguerre was photographing a street to demonstrate his process, not making a portrait, and the men are a detail he never planned.

1839–2024

The city erased the evidence. Daguerre's studio and his neighbouring diorama theatre burned in 1839. Later, Haussmann's redesign of Paris demolished much of what survived. The boulevard itself was truncated to make room for the Place de la République, and the exact location of the shoeshine stand became one of those facts that everyone assumed was lost.

It was, until three filmmakers decided it wasn't.

2024

Maximilien Van Aertryck, Axel Danielson, and Kalle Boman were standing in the Place de la République when one of them pointed out that they were roughly where the photograph had been taken. "It would be funny to try and find where exactly the shoeshiner and his client were standing," Van Aertryck later told Radio France.

The joke turned into two years of work. They combed city and museum archives, reconstructed the 1838 streetscape, and found that three buildings from Daguerre's era were still standing. They commissioned 3D models and hired two architectural firms to triangulate independently. The firms landed a few feet apart.

"We decided to split the difference," Van Aertryck told The New Yorker. "A colleague ran out to buy nail polish to mark the spot."

2026

On October 15, the City of Paris installed a bronze plaque at that spot. It is titled Un Détail Imprévu — An Unexpected Detail — which is the best possible name for a photograph that became famous for something its maker never meant to capture.

Van Aertryck calls it "the first photo of a social interaction and of labor," and "so poetic, with those two lonely people on that huge frame." Both are true. The picture is remembered for its emptiness, but its subject is a relationship: one person kneeling, one standing, a small economy of service.

There is a detail I keep returning to. The plaque marks a compromise — two expert reconstructions, a disagreement of a few feet, nail polish, then bronze. That is what historical certainty usually looks like when you get close enough to see it.

We still don't know their names.

Sources: Artnet News, Smithsonian Magazine, The New Yorker, Radio France, Ville de Paris.

That is accurate. The alarm is half-right. What interests me is the design underneath it.

Two documents, one product

Anthropic publishes two articles with almost the same name.

The commercial one — Claude for Work, the API, Claude Gov — starts with a strong promise:

By default, we will not use your inputs or outputs from our commercial products to train our models.

Then it carves out an exception:

If you explicitly report feedback or bugs to us (e.g. via our thumbs up/down feedback button), or otherwise choose to allow us to use your data, then we may use your chats and coding sessions to train our models.

The consumer one — Free, Pro, Max, and Claude Code on those plans — has a different rule set. It lists three conditions: you allow model improvement, a conversation is flagged for safety review, or you opt in to something like the Trusted Tester program. Incognito chats are exempt.

Both describe feedback the same way. Pressing the button sends "the entire related conversation," including content, custom styles, conversation preferences, and model settings, stored for up to five years. It is de-linked from your user and customer IDs before use, and not combined with your other conversations. Connector data is excluded unless you pasted it in. Team and Enterprise admins can disable the whole thing under Organization settings → Data and Privacy → Rate chats.

The fair objection

The "what did you expect?" crowd has a point. If a user says "this answer was wrong," somebody has to read the answer. Feedback without context is a useless signal, and the dialog itself says the whole conversation is being sent.

So the existence of feedback-driven review is not the problem. The scope is.

Four things that don't hold up

The unit is the whole conversation. If one sentence in a forty-message exchange is bad, the other thirty-nine travel with it. Proportionate for a support ticket. Not proportionate for a debugging session that wandered through your infrastructure and your customers' names.

Five years is a legal-discovery number. Nobody improves a model by keeping one specific chat for five years. That retention period is about defending decisions, not learning from them, and it is the part users are least likely to have considered when they clicked.

An opt-out with a side door is not an opt-out. It is a preference. The people most likely to press the button are the ones who believe they have already disabled training, and nothing in that flow tells them otherwise.

And the packaging is genuinely bad. The sentence that circulates — "by default, we will not train on your data" — belongs to the commercial document. A consumer on claude.ai is governed by the other one, which makes no such promise. Two policies, one product family, URLs differing by a few digits.

What I would actually ask for

Not a ban on feedback review. A narrower design: send the rated exchange plus a few turns, redact the rest, and ask permission per report if you want to train on it. That is a product decision, not a technical limit.

Anthropic's commercial default is better than most of the industry, and "the API does not train on your data" is a real contractual commitment. I do not think this is a company quietly harvesting chats. I do think a privacy promise is only as strong as its least obvious exception, and here the exception is a button in the corner of every conversation.

The person who turned training off, then tapped thumbs-down on a bad answer, has no way to know they just turned it back on. That is the whole problem, and it fits in one sentence.

Sources: Anthropic privacy center, commercial, consumer.

The pitch is that breathing exercises are boring and easy to abandon, so someone built a piece of hardware that makes the pacing physical instead of visual. No screen, no counting, no app required.

Whether that is worth the money depends on which of three questions you are actually asking.

What is it doing, mechanically?

Three things at once.

It paces you. The expansion and contraction are the metronome, which means your eyes are free and your hands are busy. That sounds trivial until you have tried to follow a breathing timer while your mind wanders off to the thing you were anxious about in the first place.

It measures you. The pulse sensor feeds heart rate and heart rate variability into the app, which shows a live coherence score and keeps a history. It can pull data from Oura, Apple Health, and Google Health.

And, in the premium tier, it tries to personalise. The idea is that the ideal breathing pace — the resonance frequency where HRV peaks — varies between people, usually somewhere around 4.5 to 6.5 breaths per minute, and that the standard six-breaths-a-minute protocol is not everyone's best. Moonbird says its analysis of 269,764 sessions found about three in four people respond best to something other than six. Its own research, so treat it as a plausible claim rather than a replicated result.

Is €179 the right number?

You are not paying for the breathing. A free timer can pace you. You are paying for a small, well-made object that makes the practice tactile, plus software that turns it into a habit with feedback.

Compare it to what else €179 buys in this category: three months of a meditation app you will probably stop opening, or a fraction of a wearable that measures the same HRV without teaching you to change it. The Moonbird is cheaper than a course of anything and more expensive than doing nothing.

The premium tier (€199.99 for the first year, list €238.99) adds a personalised pace, an AI coach called Aura, soundscapes, and a sleep prediction. None of that is necessary. The device works offline without any of it, which is the most underrated thing about it.

Who should not buy one?

Anyone who wants a treatment. It is not a medical device. It does not diagnose, and it will not fix insomnia on its own, whatever the testimonials say.

Anyone who has panic disorder, a trauma history, or a respiratory condition should be careful with deliberate breath work. It can produce dizziness, air hunger, or rising anxiety in exactly the people who most want it to help. Shorter sessions, a gentler exhale, and a clinician's opinion first.

And anyone whose obstacle is motivation rather than medium. The device fixes the medium. If you have bought three journals and a kettlebell you never used, a €179 pebble will join them.

If you do buy one: five minutes at the same time every day, anchored to something you already do, and keep it where you can see it. A breathing device in a drawer is an expensive paperweight. The habit is the product; the pebble is just the handle.

Further reading: moonbird.life, Sleep Foundation review, Mashable review.